title
stringlengths 10
190
| abstract
stringlengths 118
1.9k
| related work
listlengths 2
8
| references
listlengths 0
0
|
|---|---|---|---|
One size does not fit all: Investigating strategies for
differentially-private learning across NLP tasks | Preserving privacy in contemporary NLP models allows us to work withsensitive data, but unfortunately comes at a price. We know that stricterprivacy guarantees in differentially-private stochastic gradient descent(DP-SGD) generally degrade model performance. However, previous research on theefficiency of DP-SGD in NLP is inconclusive or even counter-intuitive. In thisshort paper, we provide an extensive analysis of different privacy preservingstrategies on seven downstream datasets in five different `typical' NLP taskswith varying complexity using modern neural models based on BERT andXtremeDistil architectures. We show that unlike standard non-private approachesto solving NLP tasks, where bigger is usually better, privacy-preservingstrategies do not exhibit a winning pattern, and each task and privacy regimerequires a special treatment to achieve adequate performance. | [
"A primer on differential privacy and an explanation of DP-SGD would unfortunately go beyond the scope of this short paper; please refer to {{cite:afd5a0d94b1d2a5d17482bd6f03d384818f01dde}}, {{cite:08702708b58a4724cc690cda79809fe66ff4b669}}, {{cite:843022636cdfe378196a0db59cbd2e89e35466e1}}.",
"In NLP, several studies have utilized the DP-SGD methodology, primarily for training language models.For instance, {{cite:59c7a68a3f815a399fcaf7552ee2c2c1c99625e3}} study the effect of using DP-SGD on a GPT-2 model, as well as two simple feed-forward networks, pre-training on a large public dataset and fine-tuning with differential privacy.Model perplexities are reported on the pre-trained models, but there are no additional experiments on downstream tasks.{{cite:0ee05be4ac39264cf47dcf5feaa763131167146d}} train a differentially private LSTM language model, showing that it can behave similarly to models without DP when using a large enough dataset, with comparable accuracy.{{cite:81779831f3835005021198b7205dec5ede751156}} train a differentially private BERT model using the MIMIC-III dataset {{cite:b9e0123a5a8d6b10250bbc8568d2dd5e98ca4ca5}} and a privacy budget {{formula:a15b8808-fd83-484c-b7dc-4401509e1a71}} , achieving comparable accuracy to the non-DP setting for a Medical Entity Extraction task.",
"There are only a few works that have investigated DP-SGD for specific downstream tasks in NLP. {{cite:3fdb8ab207d246b036e739c7f6c6350eeadc3f71}} looked into the behavior of differential privacy on the CoNLL 2003 English NER dataset. Their findings show that no significant drop is expected, even when using low epsilon values such as 1, and even as low as 0.022. This is a very unusual result and is assessed in our current work further below. {{cite:8d60b9c8d4b02312ed4989ac31d775ce721cba48}} apply DP-SGD, among other tasks, to sentiment analysis of Tweets, reporting a very small drop in accuracy with epsilons of 8.99 and 3.87. With studies up to now being limited to only a few tasks and datasets, using disparate privacy budgets and metrics, there is a need for a more general investigation of the DP-SGD framework in the NLP domain."
] | [] |
Leveraging Image-based Generative Adversarial Networks for Time Series
Generation | Generative models synthesize image data with great success regarding samplingquality, diversity and feature disentanglement. Generative models for timeseries lack these benefits due to a missing representation, which capturestemporal dynamics and allows inversion for sampling. The paper proposes theintertemporal return plot (IRP) representation to facilitate the use ofimage-based generative adversarial networks for time series generation. Therepresentation proves effective in capturing time series characteristics and,compared to alternative representations, benefits from invertibility andscale-invariance. Empirical benchmarks confirm these features and demonstratethat the IRP enables an off-the-shelf Wasserstein GAN with gradient penalty tosample realistic time series, which outperform a specialized RNN-based GAN,while simultaneously reducing model complexity. | [
"The challenge of generating of time series stems from changes in the intertemporal dynamics, the presence of jumps, and correlations between multiple time series. The prevailing approaches rely on recurrent neural networks {{cite:60b26f2c8725cc83d3d335002190a49f4857e9b5}} {{cite:d52bc437a7743ae3bfb3704c9c3abae64a749953}} and address these challenges by adding complexity. The most prominent RNN-based model is the TimeGAN of {{cite:60b26f2c8725cc83d3d335002190a49f4857e9b5}}. To outperform other RNN-based models, TimeGAN balances five different loss function that control the generator, the discriminator, intertemporal relationships, moments, and the embedding loss. The TimeGAN architecture assigns each network a different task and each component is trained separately before starting a joint training. Distributing TimeGAN computations is complicated, which leads to long training procedures.",
"Other approaches use convolutional neural networks to simulate image representations of time series {{cite:2b21a2d0ce685661068a191758ed91abb17414a6}} {{cite:160e896219be8b7c2fb79ec90857f4da052ba1df}} and are currently dominated by the use of spectrograms and waveforms. Spectrograms process audio signals and visualize the spectrum of frequencies of a signal as it varies with time {{cite:7f910f5f0810f7ac97a7621cd22ea8cce8150773}} while each plot is characteristic for a certain sound or a spoken word. In a recent paper, {{cite:160e896219be8b7c2fb79ec90857f4da052ba1df}} propose a GAN model with a simplified, image-based approach using two separate Wasserstein GANs. The first network takes the spectral image of the signal and generates new spectrograms. Thereafter, the other GAN generates a candidate signal that is likely to have created that spectrogram. Both GAN finally unify their loss functions in a joint expression in order to facilitate training. The reason for using two GAN is that since the initial time series cannot be inverted in a deterministic way, the second GAN is needed to construct an estimate.Due to the spectrogram representation and its non-invertibility, the sampling process requires increased complexity. Thus, the two GAN approach is a drawback.A second image representation of a time series is a waveform. It can be processed by CNNs and refers to the shape of its graph as a function of time and is used by {{cite:2b21a2d0ce685661068a191758ed91abb17414a6}} to simulate new time series. The advantage is that the representation works well for image generation while capturing the non-stationary components of a time series. The ability of waveforms to capture non-stationary data has been extensively studied in, for example, electric engineering {{cite:7f910f5f0810f7ac97a7621cd22ea8cce8150773}} {{cite:2d78baa0c3773027b8769279402592d9992f7ac2}}. However, the inversion appears difficult, as pointed out by {{cite:2b21a2d0ce685661068a191758ed91abb17414a6}}. They demonstrate that even if inversion is approximately possible via the Griffin-Lim algorithm, the sample quality is not significantly better than using an additional inference network.",
"As a result, image-based methods currently lack an efficient representation that captures the dynamics of a time series while being invertable. The new method proposed in this paper solves both problems."
] | [] |
Cognition-aware Cognate Detection | Automatic detection of cognates helps downstream NLP tasks of MachineTranslation, Cross-lingual Information Retrieval, Computational Phylogeneticsand Cross-lingual Named Entity Recognition. Previous approaches for the task ofcognate detection use orthographic, phonetic and semantic similarity basedfeatures sets. In this paper, we propose a novel method for enriching thefeature sets, with cognitive features extracted from human readers' gazebehaviour. We collect gaze behaviour data for a small sample of cognates andshow that extracted cognitive features help the task of cognate detection.However, gaze data collection and annotation is a costly task. We use thecollected gaze behaviour data to predict cognitive features for a larger sampleand show that predicted cognitive features, also, significantly improve thetask performance. We report improvements of 10% with the collected gazefeatures, and 12% using the predicted gaze features, over the previouslyproposed approaches. Furthermore, we release the collected gaze behaviour dataalong with our code and cross-lingual models. | [
"Current literature which uses gaze behaviour to solve downstream NLP tasks has been applied to the NLP tasks of sentiment analysis {{cite:c215de4f22dc4c00b8ec6466f20f88729eadbfc2}}, {{cite:e65e5bc5c1678426bfab1ee7732296bacc3100c1}}, {{cite:c187c1d0e70ae3131247238f76aa665050e678cf}}, sarcasm detection {{cite:1f18a49eda0001f344b54d5a9665e358e122f5f8}}, grammatical error detection {{cite:e65e5bc5c1678426bfab1ee7732296bacc3100c1}}, hate speech detection {{cite:e65e5bc5c1678426bfab1ee7732296bacc3100c1}}, named entity recognition {{cite:26c7bed3db7f66edeb5735e75ffed4a122c39c45}}, part-of-speech tagging {{cite:dce9f0d1b0fd0d67af9299ff3b8b6a4b1aeb7f58}}, sentence simplification {{cite:36ae21427657940a62cfa66926758d405b347f1f}}, and readability {{cite:2002f276cd44c449e4914786ae9986352850548a}}, {{cite:65b20e1635ff1553bf5b70069ac867647f1d74c9}}. A comprehensive overview is provided by mishra2018cognitively. The primary motivation of using cognitive features for NLP tasks is derived from the eye-mind hypothesis {{cite:7e13eb99f7a7d5978a8b1143a21cf960d0476008}}, which establishes a direct correlation between a reader's comprehension of the text with the time taken to read the text. This hypothesis has initiated a large body of psycholinguistic research that shows a relationship between text processing and gaze behaviour. {{cite:6a437e46b5e2b5828a6e44912f28177b8c401c8a}} discuss the use of gaze features for the task of anaphora resolution. Their findings show that gaze data can substitute the classical text processing approaches along with the fact that human disambiguation process overlaps with the information carried in linguistic features. {{cite:86901cc8bcf06aefeb83978d390ad0e180d410d5}} use gaze data to automatically identify multiword expressions and observe that gaze features help improve the accuracy of the task when combined with traditional linguistic features used for the task.",
" {{cite:ca294bccd73de4c360ccc499a4128402e7b229ec}} describe an approach to scoring essays in a multi-task learning framework automatically. Their approach relies on collecting gaze behaviour for essay reading for a small set of essays and then predicting the rest of the dataset's gaze behaviour in a multi-task learning setup. Similarly, {{cite:dce9f0d1b0fd0d67af9299ff3b8b6a4b1aeb7f58}} use token level averages of cognitive features at run time, to mitigate the need for these features at run time. {{cite:65b20e1635ff1553bf5b70069ac867647f1d74c9}} and {{cite:c187c1d0e70ae3131247238f76aa665050e678cf}} predict gaze behaviour at the token-level as well. {{cite:c215de4f22dc4c00b8ec6466f20f88729eadbfc2}}, {{cite:2002f276cd44c449e4914786ae9986352850548a}}, {{cite:e65e5bc5c1678426bfab1ee7732296bacc3100c1}}, and {{cite:36ae21427657940a62cfa66926758d405b347f1f}}, use multi-task learning to solve the primary NLP task, where learning the gaze behaviour is an auxiliary task.",
"Orthographic/String similarity-based methods are often used as baseline methods for the cognate detection task, and the most commonly used method amongst them is the Edit distance-based similarity measure {{cite:4ab6ca3238e890f98bf923b4858b0192cc0ce32c}}, {{cite:6a9a9215e5cc0495805b1167bad8708e0a658929}}. Research in automatic cognate detection using various aspects involves the computation of similarity by decomposing phonetically transcribed words {{cite:cf931e2542fdbdd111191fa9357d148891d4a7e0}}, {{cite:17d830b107582357b973969428de7ac1a4fa543e}}, acoustic models {{cite:aa98426914c934bd46d66f5505e6720f61a0529d}}, clustering based on semantic equivalence {{cite:42fafd97dd9ab123436d02ee3b6bb47f2735a249}}, and aligned segments of transcribed phonemes {{cite:5c9a0a33f56c3291d4289ba8536e07a87eee4fff}}. {{cite:90dbfdbc3c7091e41ae6bd88031713bed6720ee4}} employs a Siamese convolutional neural network to learn the phonetic features jointly with language relatedness for cognate identification. {{cite:f268a3706771c44479b1f3b108d1fa6ef3fd01b1}} use SVM for phonetic alignment and perform cognate detection for various language families. Various works on orthographic cognate detection usually take alignment of substrings within classifiers like SVM {{cite:c48ccf693f14d2cc4cdeed2f811db6b937bf4e7a}}, {{cite:401a5d21812ebfbcad511cef9f5330e3ad2d309d}} or HMM {{cite:46a403ff5ba5548fb4465c24f0a2264d817095ac}}. {{cite:c48ccf693f14d2cc4cdeed2f811db6b937bf4e7a}} employ dynamic programming based methods for sequence alignment. Cognate facilitation in second language learners has previously been explored with the help of eye-tracking studies {{cite:45fac9883c4dc8e73d43c50e1d82a8cb45fa7697}}, {{cite:3737049877f64f8a9c9f68f1f00596c10b44be9d}}, {{cite:12f2e74cfd882a655518c3450d4b1362ab84499a}} but the task of cognate detection has not been performed with gaze features obtained from the cognitive data in any of the previously available literature.",
"For the task of cognate detection, however, the use of cognitive features has not been established previously. The task of cognate detection is cross-lingual, and a reader's cognitive load should vary while trying to comprehend the meaning of concepts, from different languages. Our work tries to exploit the difference noted in terms of time taken and eye-movement patterns in cognates vs false friends, to generate additional features for the ACD task. Moreover, for Indian languages such as Marathi, where agglutinationAgglutination is a linguistic process pertaining to derivational morphology in which complex words are formed by stringing together morphemes without changing them in spelling or phonetics. varies the word length, the task becomes tougher, computationally."
] | [] |
Towards Controllable Agent in MOBA Games with Generative Modeling | We propose novel methods to develop action controllable agent that behaveslike a human and has the ability to align with human players in MultiplayerOnline Battle Arena (MOBA) games. By modeling the control problem as an actiongeneration process, we devise a deep latent alignment neural network model fortraining agent, and a corresponding sampling algorithm for controlling anagent's action. Particularly, we propose deterministic and stochastic attentionimplementations of the core latent alignment model. Both simulated and onlineexperiments in the game Honor of Kings demonstrate the efficacy of the proposedmethods. | [
"The most related to our work is Tencent JueWu-SL, which provides a supervised learning method to achieve human-level performance in playing Honor of Kings. As our goal is to improve human-agent cooperation, it is suitable to use JueWu-SL's method to train human-like agents. Therefore, we directly use the same design of features, labels, training data preprocessing and the policy loss. For details of these parts, refer to {{cite:0ee5448f423ae3ec6c2d5f2023e1a7a02deaa806}}. Although JueWu-SL agents perform collaboration with humans well, their actions can not be controlled by human players. In other words, JueWu-SL agents make decisions autonomously, whose actions only depend on the observed game states. In contrast, our proposed methods can train action controllable agent that can do what a human player wants it to do. Technically, the main difference with JueWu-SL is that our method is based on the latent alignment model. Consequently, we can change the agent's predicted action by manipulating the latent alignment variable.",
"Conditional imitation learning is a promising approach to train an action controllable agent {{cite:17aec557309426b8efeb39844b465004315a2c64}}.The predicted action is determined by both the state and a conditional command. Human can control agent's action via a command. This method requires that the training samples must be triples {{formula:00aacce3-fba9-4fcc-a856-9f80666a4244}} state, action, command{{formula:dbe75715-181a-474b-8c59-a6f8d83970d7}} . At training time, the model is given not only the state and action, but also a command representing human's intention. In comparison, our methods need no command infos for training. The command is used to sampling a latent variable when inferring.",
"The stochastic attention method enjoys the same spirit with latent dirichlet allocation in topic modeling {{cite:48e830baa565319bc4f65c506af6cf1c7848f15e}}. Given the observed state, the action is affected by a mixture of a small number of important units. Recently, variational autoencoder with the Dirichlet distribution as prior has been widely studied. One major difficulty is reparameterization trick of the Dirichlet distribution. Various approaches are proposed to tackling this issue, including softmax Laplace approximation {{cite:9a1beba0e3e7d8389f23b072568683b9fef4487d}}, Weibull distribution approximation {{cite:64479245b7a45e225144a530e37a4a518bbbf624}}, implicit reparameterization gradients {{cite:9c9273f47fafb31818096b1ac7552b34ccc5255b}}, inverse Gamma approximation {{cite:da2ed4c13a3b59a99dfba32354d9707127b5bf2a}}, etc. We use the rejection sampling variational inference {{cite:3373d1552008a2ba190b3d929322b2272520f5b8}}, as which has been validated to performing well {{cite:76e775e683d47fdd01658ec1e61faf1d803f8d8f}}."
] | [] |
Building separable approximations for quantum states via neural networks | Finding the closest separable state to a given target state is a notoriouslydifficult task, even more difficult than deciding whether a state is entangledor separable. To tackle this task, we parametrize separable states with aneural network and train it to minimize the distance to a given target state,with respect to a differentiable distance, such as the trace distance orHilbert--Schmidt distance. By examining the output of the algorithm, we obtainan upper bound on the entanglement of the target state, and construct anapproximation for its closest separable state. We benchmark the method on avariety of well-known classes of bipartite states and find excellent agreement,even up to local dimension of $d=10$, while providing conjectures and analyticinsight for isotropic and Werner states. Moreover, we show our method to beefficient in the multipartite case, considering different notions ofseparability. Examining three and four-party GHZ and W states we recover knownbounds and obtain additional ones, for instance for triseparability. | [
"Previous work on using machine learning for the separability problem has been focused either having the machine choose good measurements and then using an existing entanglement criteria {{cite:6f60e675b6eb9a249e852654a5424063bd2edeef}}, {{cite:151c02acb271fe3265dde4bf1231fcae4930d226}} , or on viewing the task as a classification problem {{cite:db7ceddf8f2af151294758af5e5718ca412cdd76}}, {{cite:dd93a44183745263140811a149a3976b281f9c54}}, {{cite:624a4fb5bd987ed7d96437aef0d8b3011b378e1d}}, {{cite:e387a53c0dd0e6d340ebf91247c679d64a244025}}, {{cite:b047a39e4bba90bedaea78ba83792017ea840efe}}, {{cite:bf980a9259cd53a8c7e9c9df2c4e4e41f6b7bbd4}}, {{cite:b0780b0f4723c6d0a34609b01e291ebc3783481e}}. For classification, typically a training set is constructed where quantum states are labeled as separable or entangled. The machine learns on this training set and given a new example predicts whether it is entangled or separable. There are several difficulties with this approach. First, the machine just gives a guess of whether the state is entangled or separable, and does not provide any kind of certificate. Second, the training data can only be generated in a regime where we already understand the problem well, which results in the machine giving only marginal new insight at best. This could be circumvented by using suboptimal criteria (e.g. PPT) in order to create the training data, however, the machine would just learn this criteria instead of correctly identifying the entanglement/separability boundary.",
"We overcome these challenges by using a generative model, which tries to give an explicit separable decomposition of a target state. This way we immediately get a certified upper bound on the distance from the separable states. A similar approach has been taken in Refs. {{cite:60f407cbfcb82d4644ec2df385e28caa5a0e1fbb}}, {{cite:41ec783664ca3d096e087fa2d8c16ee5ac6e2155}}, where the authors represent the quantum states with “quantum neural network states” {{cite:c2d914c8560fc700edc8fc6266798ce4fd1d5a2d}}, {{cite:9dfc45c0abe258dd4ca61093cd5a83736f2065f6}}, and their extension to density matrices {{cite:65e44c8618facb9cd352b5897c0b849b9be3db42}}, {{cite:688d96d7eaf3c9b89c00738384cdd9c393e0a414}}, {{cite:e54b32879c062940713558f6880675d9bd3ca70c}}, {{cite:e68f6d051b8db735765b7282f5ea6bfe50cd391f}}, as opposed to the dense representation we utilise. Their results show a more limited flexibility in the loss function and in the design of types of separable states. One such family of separable states that is not examined is the very challenging, yet interesting question of multipartite bi- and triseparability (relevant for 3-party and 4-party genuine multipartite entanglement), which we address here."
] | [] |
Hierarchical Variational Memory for Few-shot Learning Across Domains | Neural memory enables fast adaptation to new tasks with just a few trainingsamples. Existing memory models store features only from the single last layer,which does not generalize well in presence of a domain shift between trainingand test distributions. Rather than relying on a flat memory, we propose ahierarchical alternative that stores features at different semantic levels. Weintroduce a hierarchical prototype model, where each level of the prototypefetches corresponding information from the hierarchical memory. The model isendowed with the ability to flexibly rely on features at different semanticlevels if the domain shift circumstances so demand. We meta-learn the model bya newly derived hierarchical variational inference framework, wherehierarchical memory and prototypes are jointly optimized. To explore andexploit the importance of different semantic levels, we further propose tolearn the weights associated with the prototype at each level in a data-drivenway, which enables the model to adaptively choose the most generalizablefeatures. We conduct thorough ablation studies to demonstrate the effectivenessof each component in our model. The new state-of-the-art performance oncross-domain and competitive performance on traditional few-shot classificationfurther substantiates the benefit of hierarchical variational memory. | [
"MemorySeveral works augment neural networks with an external memory module to improve learning {{cite:f32dee2d7da91bcf742dd1a15bb1d47b797e8e18}}, {{cite:44f5fd9a39aa4d7e6eac8cb6433f0475f74b6b5a}}, {{cite:347d10eb355f703f44da2401337bbd759e85838d}}, {{cite:86387a14521afb28a88b5d7143172d173cd122ca}}, {{cite:fa7f18ca7982ad49040b4049c17ea2c4bad23aaf}}, {{cite:0493b2fa6b2bf4fddeb323bca807a91f7eb79320}}, {{cite:f29cd5ea5ee9c770418a3791397ccb19c410a020}}, {{cite:36aebbd6747f4b98e75069417e7c4a8a3e5c2fa6}}. {{cite:f32dee2d7da91bcf742dd1a15bb1d47b797e8e18}} equip neural networks with a neural Turing machine for few-shot learning. The external memory stores samples in the support set in order to quickly encode and retrieve information from new tasks. {{cite:43bf125270d1baf90a441c6954f5482c19c66e09}} proposed augmenting generative models with an external memory, where the chosen address is treated as a latent variable. Sampled entries from training data are stored in the form of raw pixels in the memory for few-shot generative learning. In order to store a minimal amount of data, {{cite:36aebbd6747f4b98e75069417e7c4a8a3e5c2fa6}} proposed a surprise-based memory module, which deploys a memory controller to select minimal samples to write into the memory.An external memory was introduced to enhance recurrent neural network in {{cite:8fc340e716ea3868707e7d8d6542eda834b2261d}}, in which memory is conceptualized as an adaptable function and implemented as a deep neural network. Semantic memory has recently been introduced by {{cite:e6ae80768bdb2f2d1d4fb895749b166ff5ab21a8}} for few-shot learning to enhance prototypical representations of objects, where memory recall is cast as a variational inference problem.In contrast to {{cite:e6ae80768bdb2f2d1d4fb895749b166ff5ab21a8}}, our memory adopts a hierarchical structure, which stores the knowledge at different levels instead of the last layer only.",
"Prototypes The prototypical network for few-shot learning is first proposed by {{cite:28fdc844cec5e74ce52a1ca49e93b57ee4ba04bb}}. It learns to project the samples into a common metric space where classification is conducted by computing the distance from query samples to prototypes of classes.Infinite Mixture Prototypes was proposed by {{cite:547937c0f2004b9a301badb4c2b2f0119419f10a}}, which improved the prototypical network by an infinite mixture of prototypes that represents each category of objects by multiple clusters. {{cite:dc0bb9b567af24ab6957a3c35d0778546a498331}} combines the complementary strengths of Prototypical Networks and MAML {{cite:7abb7efab2a8b5a4f1c9f1dc926de17e5522b8d5}} by leveraging their effective inductive bias and flexible adaptation mechanism for few-shot learning. {{cite:e6ae80768bdb2f2d1d4fb895749b166ff5ab21a8}} proposed the variational prototypical network to improve the prototypical network by probabilistic modeling of prototypes,which provided more informative representations of classes.Different from their work, we propose hierarchical variational prototype networks to generate different layer prototypes instead of using one prototype, which could capture information at different levels for few-shot learning across domains.",
"Meta-learning or learning to learn {{cite:1f6d46a3ef6a95179463de4e4cbfa9ee0444fad1}}, {{cite:2f219bd777964f2c41d7371592fb753333566bc3}}, {{cite:af17438dbbef35e491593e4d17a2f511e2e63287}}, is a learning paradigm where a model is trained on distribution of tasks so as to enable rapid learning on new tasks. Meta-learning approaches to few-shot learning {{cite:7a2f828ac3d881f91e3bce494fa9ce39de2aeb5a}}, {{cite:0a7ce1d29151dd9503e0db51f15a569e28378cc0}}, {{cite:28fdc844cec5e74ce52a1ca49e93b57ee4ba04bb}}, {{cite:e6ae80768bdb2f2d1d4fb895749b166ff5ab21a8}}, {{cite:7abb7efab2a8b5a4f1c9f1dc926de17e5522b8d5}} differ in the way they acquire inductive biases and adapt to individual tasks. They can be roughly categorized into three groups.Those based on metric-learning generally learn a shared embedding space in which query images are matched to support images for classification {{cite:7a2f828ac3d881f91e3bce494fa9ce39de2aeb5a}}, {{cite:28fdc844cec5e74ce52a1ca49e93b57ee4ba04bb}}, {{cite:f32dee2d7da91bcf742dd1a15bb1d47b797e8e18}}, {{cite:45dc85c07b9c10906016a453023624fb3d29c510}}, {{cite:547937c0f2004b9a301badb4c2b2f0119419f10a}}. The second, optimization-based group learn an optimization algorithm that is shared across tasks,which can be adapted to new tasks for efficient and effective learning {{cite:0a7ce1d29151dd9503e0db51f15a569e28378cc0}}, {{cite:a1cdd1afd4ba3db8c805f8aedcebdc1fa2c13178}}, {{cite:7abb7efab2a8b5a4f1c9f1dc926de17e5522b8d5}}, {{cite:36379de569314faca4393d4b0695d6caf1e63b2e}}, {{cite:a210cf3cc1e6ed92311fc831b9784d26f921fe5b}}, {{cite:aaf297ae2418d182419504e2b9f1a87c7f965f70}}, {{cite:98edf6c326e02b742597920c8fee4e85571add02}}. The memory-based meta-learning group leverages an external memory module to store prior knowledge that enables quick adaptation to new tasks {{cite:f32dee2d7da91bcf742dd1a15bb1d47b797e8e18}}, {{cite:0493b2fa6b2bf4fddeb323bca807a91f7eb79320}}, {{cite:f29cd5ea5ee9c770418a3791397ccb19c410a020}}, {{cite:4be6eee74a374c91dc00f8a048f54062bd9f4010}}, {{cite:e6ae80768bdb2f2d1d4fb895749b166ff5ab21a8}}.Our method belongs to the first and third group, as it is based on prototypes with external hierarchical memory, with the goal to perform cross-domain few-shot classification.",
"Cross-domain few-shot learning The problem is formally posed by {{cite:af3e2381a7547bb019378ddb518ce9b7e52d4865}}, who attack it with feature-wise transformation layers for augmenting mage features using affine transforms to simulate various feature distribution. Then, {{cite:adec7e1b82953b31bbe7aa575fe8badce9936d48}} proposed the cross-domain few-shot learning benchmark, which covers several target domains with varying similarities to natural images. {{cite:5139585f0f86835b1c8ff6e80c2d114bfb40e48b}} introduce a method, which allows few-shot learners to adapt feature representations to the target domain while retaining class grouping induced by the base classifier. It performed cross-domain low-level feature alignment and also encodes and aligns semantic structures in the shared embedding space across domains.{{cite:c3473f77b59071f5d44c936627be44cee7013121}} proposed a method which improved the cross-domain generalization capability in the cross-domain few-shot learning through task augmentation.In contrast, we first address cross-domain few-shot learning by a variational inference approach, which enables us to better handle the prediction uncertainty on the unseen domains. Moreover, we propose a hierarchical variational memory for cross-domain few-shot learning, and to better deploy the hierarchical memory, we introduce a hierarchical variational prototype model to obtain prototypes at different levels."
] | [] |
Design Challenges for a Multi-Perspective Search Engine | Many users turn to document retrieval systems (e.g. search engines) to seekanswers to controversial questions. Answering such user queries usually requireidentifying responses within web documents, and aggregating the responses basedon their different perspectives. Classical document retrieval systems fall short at delivering a set of directand diverse responses to the users. Naturally, identifying such responseswithin a document is a natural language understanding task. In this paper, weexamine the challenges of synthesizing such language understanding objectiveswith document retrieval, and study a new perspective-oriented documentretrieval paradigm. We discuss and assess the inherent natural languageunderstanding challenges in order to achieve the goal. Following the designchallenges and principles, we demonstrate and evaluate a practical prototypepipeline system. We use the prototype system to conduct a user survey in orderto assess the utility of our paradigm, as well as understanding the userinformation needs for controversial queries. | [
"The importance of retrieval result diversity has long been recognized by the information retrieval community {{cite:9b87a1b43e7bf8e08818a6a082f2a298c3161a49}}, {{cite:5687c114542fb81e03596375fb7fc2335b692990}}. The problem has mostly been studied in the context of topic or entity ambiguity in the user query {{cite:fcd1d7fe209736ec755ade622b0d8b970217950d}}, and most solutions focus on resolving the ambiguity and uncovering the user intent, or hierarchy of the target topics {{cite:a03fddb92407647df0c522d2c95e43281903478e}}, {{cite:b099be2e84fb94090382a3ab99238eef5ed4160b}}, {{cite:0dc49d77413a24d69770c462e3879b58f17c7bff}}, {{cite:f1c55a95828d47485fc4aeaa0ee0b392a43b07c6}}. The most notable hypothesis for this line of work, which our study also investigates, is the importance of modeling the inter-dependencies among documents during retrieval. The TREC 2009 Web track's diversity task {{cite:52ef8b0f185240c83ff9db11f2fbbbb6a3b236cc}} introduces an evaluation protocol where the relevance of next retrieval item is dependent on the previous result.",
"Our study takes a different angle from query ambiguity, and hypothesize that in the case of debate-worthy query, users wish to see direct responses from diversified perspectives in the retrieved documents {{cite:3c5f941fca7c08b128fa3a59835a8718c93f0fb7}}. Such objectives requires recognizing the argumentation structure of a document {{cite:670292381b7e7db7c61e7d99a6898b5bb156433e}}, {{cite:87b79c8fbb4e1177ee1e173f9db6f628306f1379}}, and more importantly comparing the semantics and implications of the arguments made cross documents {{cite:895b873119ba99736465bf7af22eaa632f2f9eba}}. {{cite:94fa1ca4aa626c7545c65c892f2f897fd310eb93}}, {{cite:4496cfa3854c1cab01214ea5667c802069957da5}} formulate a argument retrieval problem that conceptually resembles our formulation. In their task, a system is expected to return single-sentence arguments from different perspectives given a user query. Our study instead focuses on the more practical and challenging setting of document-level retrieval.",
"The task of question answering (QA) aims to identify direct answers to users questions expressed in natural languages from either close- or open-domain information resources {{cite:0e0008a84a0d985e195db7cd0f1fd0616a16519d}}. Naturally, document retrieval has been an essential step during answer candidate generation {{cite:8f4f08e1da2936f50c24dfb6bff3d2503325fef1}}. Among various types of QA tasks, Machine reading comprehension {{cite:06c5527369a735ac5a8f5d83f8860c27753630e0}}, {{cite:dce4ff25434304fbe0c8b4a40b99837b86cc2781}} and open-domain QA {{cite:6eb31fd73b66b4cd5e65ce680bdddeb6bb30f6f9}}, {{cite:e6792f40404856fa24a75ae9bc0793db28ebb92b}} resemble our problem closely, as both involves identifying a concise response from large- to web-scale corpora. Under such settings, researchers have found benefit in jointly learning QA and retrieval tasks {{cite:f2c4fdaeff4b813f1a4fa88283ffde43c78fc92b}}. The key difference between most QA tasks and our problem is that objective correctness does not exist for most controversial queries, and so they cannot be responded with a single correct answer. Instead, there exist dimensions such as stance {{cite:7fac4d24a07e345666706980f99dd8e49e191f9c}} and persuasiveness {{cite:66970d3934faa9a08b906386487eeaf0ab5535f6}} to measure the semantic difference between relevant and arguably equally valid arguments. To this end, recent researchers have attempted to extract or generate explanations alongside search results {{cite:7c862a0cdcc45c1de6ecfd829b2c7688f4d24ab9}}.{{figure:307bbf59-bac1-417f-b2ba-1a8f707ae51c}}"
] | [] |
Model Stealing Attacks Against Inductive Graph Neural Networks | Many real-world data come in the form of graphs. Graph neural networks(GNNs), a new family of machine learning (ML) models, have been proposed tofully leverage graph data to build powerful applications. In particular, theinductive GNNs, which can generalize to unseen data, become mainstream in thisdirection. Machine learning models have shown great potential in various tasksand have been deployed in many real-world scenarios. To train a good model, alarge amount of data as well as computational resources are needed, leading tovaluable intellectual property. Previous research has shown that ML models areprone to model stealing attacks, which aim to steal the functionality of thetarget models. However, most of them focus on the models trained with imagesand texts. On the other hand, little attention has been paid to models trainedwith graph data, i.e., GNNs. In this paper, we fill the gap by proposing thefirst model stealing attacks against inductive GNNs. We systematically definethe threat model and propose six attacks based on the adversary's backgroundknowledge and the responses of the target models. Our evaluation on sixbenchmark datasets shows that the proposed model stealing attacks against GNNsachieve promising performance. | [
"In this section, we review the research work close to our proposed attacks.We refer the readers to {{cite:a7b1845c9da072e3a44541e23ca4939befbd2f31}}, {{cite:de9540285ad80a59651bcc9310b5e7e2641515c9}}, {{cite:f5c720c25933521e54528e15c4a760a62c97fdc9}}, {{cite:ff9d337f29003442137df32dc8d2761096457ddd}}, {{cite:0c622109da965917f280725e82b5d486f948dc63}} for an in-depth overview of different GNN models, and {{cite:2400e8fb1ccae8af01938af69bf79ac702a688e4}}, {{cite:5b897d7c629351f60228cee15c64358ff7d50c49}}, {{cite:e499634522419a3729d7714b42290bbbb41cfe69}}, {{cite:f20799dc08373306fdde6b667a3b072e9be653b0}}, {{cite:ebcbd4e6ec529d745de26ee88f26ab4eab8792d3}} for comprehensive surveys of existing adversarial attacks and defense strategies on GNNs.",
"Model Stealing Attack Against ML Models.Model extraction is in many ways similar to model distillation, but it differs in that the victim’s proprietary training set is not accessible to the adversary.In this regard, previous literature already investigated stealing various aspects of a black-box ML model such as hyperparameters {{cite:13397f6aa515be4e89faff7ff1ab8d6e1c199bf6}}, architecture {{cite:b84f3d4166a95cab6df332dc80f1efd2754ac3b0}}, information on training data {{cite:7ec89a37a7b058f9e971872175c5270c7b2891c5}}, {{cite:d18ac103ac2d2369bb4411936bc9d7c6388ec033}}, parameters {{cite:4da69f7a2b1946f1cde2711fc20925e044593c17}}, {{cite:84e050173d8523ce08e9b9b8bc8a4912f4fd7d12}}, decision boundaries {{cite:bffdb6c7e910f9473cd68b7ef08138b48d80df72}}, and functionality {{cite:37cb4f4a8016ec11f0e25bb8c7d75a1704a34595}}, {{cite:49eed56809a5401bc6ed8cded99bd358f9b5b223}}.However, most of these efforts focused on images.There exist some preliminary work on model stealing attacks against GNNs {{cite:6686ae49f40eb84256ef48551c43091f25b8e9fa}}, {{cite:62ee16876fb9cc9f345ca4bdd10712ac3419ac49}}.However, they are only focusing on the transductive setting of GNNs, which cannot generalize to unseen data.Our attacks instead focus on a more popular and general setting of GNNs, i.e., inductive setting.We fill the gap and understand if both theft and reconnaissance adversaries can steal inductive GNNs with high accuracy and high fidelity.",
"Causative Attacks on GNNs.Many adversarial attacks to GNNs {{cite:3d558902cc712ea03017e94bfdc9a7b784e63838}}, {{cite:502ea49b863f38037f7cfb74ae019a034a7525e6}}, {{cite:07f280f9b988e834296c8521598e3c9daba7eaa8}}, {{cite:0e4ac65d33760eeaffd65b3a5a84d2b8345a1f02}}, {{cite:2efefbf0d603077317826e4851a4fb8cd96d85a0}}, {{cite:42601ae32809eeca52a8973bd4bfe678b1f558ef}}, {{cite:152b99f65710aff4d2a5ce11d9ce3ec7085360d2}}, {{cite:5b897d7c629351f60228cee15c64358ff7d50c49}}, {{cite:2ca774b850f19f9d22e4db8ebba9a8bcb54a9165}}, {{cite:c010e5a38b57af66304d4ca61567f1727d97056e}} are causative attacks {{cite:e5324e5fe05fe60d2d4101b425adde4a18a5c05d}}.These attacks assume that an adversary can manipulate the training dataset in order to change the parameters of the target model and influence their behavior.In this context, Zügner et al. {{cite:92801a8ee476ba5707a4f7e083ce7f408b89f837}} was the first to introduce unnoticeable adversarial perturbations of the node’s features and the graph structure.Their goal was to reduce the accuracy of node classification via GCN.After this work, different adversarial attack strategies have been proposed.Depending on the attack objectives, they aim atreducing the accuracy of node classification {{cite:3d558902cc712ea03017e94bfdc9a7b784e63838}}, {{cite:10fed018b3818657862da633f26207d9f16556c2}}, {{cite:07f280f9b988e834296c8521598e3c9daba7eaa8}}, {{cite:0e4ac65d33760eeaffd65b3a5a84d2b8345a1f02}}, {{cite:2efefbf0d603077317826e4851a4fb8cd96d85a0}}, {{cite:42601ae32809eeca52a8973bd4bfe678b1f558ef}} (node level), link prediction {{cite:152b99f65710aff4d2a5ce11d9ce3ec7085360d2}}, {{cite:3d558902cc712ea03017e94bfdc9a7b784e63838}}, {{cite:3b06dea189267d3afaae67c0d46a277ea744dfe3}}, {{cite:126a4108f99ab6bf3b9d96f80bfbd5b7807fcdc3}} (edge level), graph classification {{cite:5b897d7c629351f60228cee15c64358ff7d50c49}}, {{cite:2ca774b850f19f9d22e4db8ebba9a8bcb54a9165}} (graph level), etc.Our attack does not tamper with the training graph data and does not change the behavior of the target model or its parameters.",
"Exploratory Attacks on GNNs.In reality, however, it is more practical for the attacker to query the target model and leverage the model’s responses on these carefully crafted input data.Consequently, we witness the emerging of exploratory attacks on ML models.However, adversarial exploratory attacks on GNNs remain understudied.In particular, only a few studies {{cite:7ec89a37a7b058f9e971872175c5270c7b2891c5}}, {{cite:62ee16876fb9cc9f345ca4bdd10712ac3419ac49}}, {{cite:cbf190236b8fcfcb7851091fb86f83752a858093}} focused on exploratory attacks on GNNs.For instance, He et al. {{cite:7ec89a37a7b058f9e971872175c5270c7b2891c5}} proposed the first link stealing attack to infer if there exists an edge between a given pair of nodes in the training graph.Duddu et al. {{cite:cbf190236b8fcfcb7851091fb86f83752a858093}} and He et al. {{cite:d18ac103ac2d2369bb4411936bc9d7c6388ec033}} discussed the membership inference attack that infers whether a given node in the graph was used to train the target model by leveraging different background knowledge.Note that, in membership inference attacks, the attack model in its training phase does not interact with the target model. However, in model stealing attacks, the surrogate model does have interaction with the target model since it needs the guidance of the target model to optimize its parameters.",
"Defense of Attacks on GNNs.To mitigate those attacks, several defense strategies have been proposed.The core idea of the existing defense strategies is reducing the sensitivity of GNNs using adversarial training {{cite:03524cd13bb4abcac0eab4538f231911eb1a7e58}}, {{cite:c1a17f74e8be51deafb076ee16ddd6e526d315df}}, {{cite:c934a82250bd63a635f8e11bd1b90236e9e28efc}}, {{cite:d18152c07e220f66f5727931453f31e8e6234140}}, {{cite:c8169a20becc8de9854177342c31ccc09d8be9ae}}, {{cite:829468e6d65290029ad78a97aae14f59213ff1b4}}, perturbation detection {{cite:b71bb09e0fa4ffe1674228006ed6ec4e9ea56ef2}}, graph sanitization {{cite:571a4a77df6f5157475dd2f1e9478d84b2d44a78}}, {{cite:17312e881d480f2d01f77c85f5764482e458d895}}, etc.In turn, the trained GNNs are robust to perturbation (e.g., structure perturbation {{cite:03524cd13bb4abcac0eab4538f231911eb1a7e58}}, {{cite:c8169a20becc8de9854177342c31ccc09d8be9ae}}, {{cite:829468e6d65290029ad78a97aae14f59213ff1b4}}, attribution perturbation {{cite:c1a17f74e8be51deafb076ee16ddd6e526d315df}}, {{cite:c934a82250bd63a635f8e11bd1b90236e9e28efc}}, {{cite:d18152c07e220f66f5727931453f31e8e6234140}}, {{cite:c8169a20becc8de9854177342c31ccc09d8be9ae}}, {{cite:829468e6d65290029ad78a97aae14f59213ff1b4}}).Also, robustness certification {{cite:d78bbb5c02815550163043cddbcd18625ee9f18b}}, {{cite:036604da3eae2d9be43f4d48704a370a0ecff337}} becomes an emerging research direction.They aim at reasoning the safety posture of GNNs under adversarial perturbations.However, these defense techniques only protect GNNs from causative attacks instead of exploratory attacks."
] | [] |
Feature-Attending Recurrent Modules for Generalizing Object-Centric
Behavior | To generalize in object-centric tasks, a reinforcement learning (RL) agentneeds to exploit the structure that objects induce. Prior work has eitherhard-coded object-centric features, used complex object-centric generativemodels, or updated state using local spatial features. However, theseapproaches have had limited success in enabling general RL agents. Motivated bythis, we introduce ``Feature-Attending Recurrent Modules'' (FARM), anarchitecture for learning state representations that relies on simple, broadlyapplicable inductive biases for capturing spatial and temporal regularities.FARM learns a state representation that is distributed across multiple modulesthat each attend to spatiotemporal features with an expressive featureattention mechanism. This enables FARM to represent diverse object-inducedspatial and temporal regularities across subsets of modules. We hypothesizethat this enables an RL agent to flexibly recombine its experiences forgeneralization. We study task suites in both 2D and 3D environments and findthat FARM better generalizes compared to competing architectures that leverageattention or multiple modules. | [
"While a large body of work has focused on studying generalization in deep RL {{cite:43903baa9573ebce2753dbcdefa417988bb237e7}}, {{cite:09399227e9839a62482ccc8b70ad0f59c9304df9}}, {{cite:47b926ab23d15f56fba023c1500ae25a62ccc6c9}}, {{cite:9ddc876e05c1b2ad7aebc1539228b79050f6fc3a}}, {{cite:71d70f69b87e4c90ad4cf12a26c1c794fbb85cc7}}, {{cite:cb2a8f6fc0e1df4e6692499f19fb1a1e6d7d3cc4}}, there has been less emphasis on studying generalization within diverse environments and task structures.One research direction has focused on generalizing to longer tasks, e.g. executing longer sequences {{cite:c43614bc297d0a08c0e524be730cfce016d2b5b7}}, {{cite:d42fd9c22c602e4060547ce9491ed9564813a9f6}}, {{cite:9535e0fa868bc13f06581251e2661039b542ff99}} or executing novel subtask structures {{cite:8f8096be2f1ad9f9cb36c2deabdc224315111d51}}, {{cite:9a911ef82179d227244c79277fe2e104dc73a4ed}}, {{cite:52f729d6a0f616e33fb8d6a3f4b19274c4419051}}. Another direction has focused on generalizing to tasks with novel features, e.g., novel shape-color combinations {{cite:93bcc3a8e693dae58b267d00f8a3f3e09ba2a672}}, {{cite:900ff56fd6a43983d1affaa18ae27ca3d6a08d47}}, {{cite:0dca16316c0e5d534375d6290887b46d39e93347}}, {{cite:961566b97e17284c77a8752d18b3b37098871cff}}, novel backgrounds {{cite:76f8e3ea1bd3bca70baaa3b3a509524330da8be9}}, {{cite:74229377579d5541edf333a0ca91718085d32fb2}}, and novel distractors {{cite:fb115e9b5352f85a256caa87c9abcd63f689c4c7}}, {{cite:8c69ef1ec0600a23bf93239d8a3003b145a8254e}}. We attempt to bridge these prior strands of research by developing a single architecture for (a) generalizing recall to novel compositions of object motions; (b) generalizing sequential perception of 3D objects to larger environments; and (c) generalizing memory of goal information to longer task lengths.",
"Task-driven generalization.Recent work which has shown that a diverse training curriculum can promote generalization {{cite:e08aa46595cde491667f83b858f6cce49b88f6d3}}, {{cite:f19baad745fbcdbd8f3ca7b93291ab93ad1aa575}}, {{cite:961566b97e17284c77a8752d18b3b37098871cff}}. This research inspired our task-driven approach to discovering generalizable “schema-like” representations.Additionally, our procedurally-generated KeyBox task follows previous research on using procedural level generalization for faster learning and generalization {{cite:fb28620b496ac7ddde4e1d1a5554c94a2db05c51}}, {{cite:c56fc3e87924470aa5b049bd52347d32fed823b1}}.",
"Generalizing with feature attention.Most similar to our feature-attention mechanism are the attention mechanisms by {{cite:1efa810c5bd4c3e2863099a04bd33387dbf342da}}, {{cite:900ff56fd6a43983d1affaa18ae27ca3d6a08d47}}.In particularly, {{cite:900ff56fd6a43983d1affaa18ae27ca3d6a08d47}} showed that mapping language instructions to non-linear feature coefficients enabled generalizing to tasks specified over unseen feature combinations in a 3D environment. While FARM also learns non-linear feature coefficients, our work has two important differences. First, we develop a multi-head version where individual feature coefficients are produced by their own recurrent modules. This enables FARM to leverage this form of attention in settings where language instructions don't indicate what to attend to (this is true in {{formula:e8f92c9c-fb73-4e9f-a7e7-c644bbcec0e8}} of our tasks). Second, we are the first to show that feature attention facilitates generalizing recall of object dynamics (Figure REF ) and generalizing memory-retention to larger environments (Figure REF ).",
"Generalizing with top-down spatial attention.Most similar to FARM are the Attention Augmented Agent (AAA) {{cite:fb115e9b5352f85a256caa87c9abcd63f689c4c7}} and Recurrent Independent Mechanisms (RIMs) {{cite:8c69ef1ec0600a23bf93239d8a3003b145a8254e}}.Both are recurrent architectures that produce top-down spatial attention over observations.Additionally, RIMs also uses a modular state representation produced by a set of recurrent modules.AAA showed generalization to unseen distractors and RIMs to more distractors than trained on.revisioncolorWhile follow-up work on RIMs has addressed problems such as learning type-specific update rules {{cite:5e4921eab7a7d6b02b499a587f6aa27e0913b112}}, {{cite:b20f7146f5938edaa0668c58a91f141bcbd3b159}} and improved information sharing among modules {{cite:23d7b671b19d0072d9e5d2b7d8e4a939d6c3b9c0}}, {{cite:9b1c82630922a3e7a99f496007382b45b416cd5a}}, there hasn't been an emphasis on which components enable generalization in reinforcement learning across diverse domains.The major difference between AAA, RIMs, and FARM is that FARM attends to an observation with feature attention as opposed to spatial attention.Our experiments indicate that spatial attention can be an impediment to reinforcement learning of tasks defined by shape-color agnostic object-dynamics and 3D objects."
] | [] |
One System to Rule them All: a Universal Intent Recognition System for
Customer Service Chatbots | Customer service chatbots are conversational systems designed to provideinformation to customers about products/services offered by differentcompanies. Particularly, intent recognition is one of the core components inthe natural language understating capabilities of a chatbot system. Among thedifferent intents that a chatbot is trained to recognize, there is a set ofthem that is universal to any customer service chatbot. Universal intents mayinclude salutation, switch the conversation to a human agent, farewells, amongothers. A system to recognize those universal intents will be very helpful tooptimize the training process of specific customer service chatbots. We proposethe development of a universal intent recognition system, which is trained torecognize a selected group of 11 intents that are common in 28 differentchatbots. The proposed system is trained considering state-of-the-artword-embedding models such as word2vec and BERT, and deep classifiers based onconvolutional and recurrent neural networks. The proposed model is able todiscriminate between those universal intents with a balanced accuracy up to80.4\%. In addition, the proposed system is equally accurate to recognizeintents expressed both in short and long text requests. At the same time,misclassification errors often occurs between intents with very similarsemantic fields such as farewells and positive comments. The proposed systemwill be very helpful to optimize the training process of a customer servicechatbot because some of the intents will be already available and detected byour system. At the same time, the proposed approach will be a suitable basemodel to train more specific chatbots by applying transfer learning strategies. | [
"Intent recognition systems have been accurately modeled by the use of recurrent neural networks (RNN), particularly using gated recurrent units (GRUs) and long short-term memory (LSTM) cells {{cite:866687432d89aa30d84936d616698e74dfb2d36d}}, {{cite:9e082a165d45bc22c9c85861caf55039861e879b}}. More recently, the use of Transformer models {{cite:05cc4ebf2171535011bf1aca011faa9a61eec52c}} such as Bidirectional Encoder Representation from Transfomers (BERT) {{cite:9ffe465843f1b44e1dfc8f198a7c0d9b68d3b7d1}} have achieved state-of-the-art results for intent classification problems.",
"The authors in {{cite:9e082a165d45bc22c9c85861caf55039861e879b}} proposed a model to learn the relation between intents and entities by introducing a slot-gated mechanism into an attention layer combined with an LSTM network. The authors evaluated their proposed approach classifying 21 intents from the Airline Travel Information Systems (ATIS) dataset {{cite:95e2701735658acf0ca15e896bf84ceb59fe5bdb}}, which has 4,478 utterances of people making flight reservations, and reported an accuracy of up to 94.1%.The authors in {{cite:866687432d89aa30d84936d616698e74dfb2d36d}} explored the use of pre-trained BERT embeddings for joint intent classification and entity extraction. The authors reported an accuracy of up to 97.9% recognizing 21 intents from the ATIS corpus.Another joint intent and entity classification was recently proposed in {{cite:e0de43432d6859f172b4e2ed7974dbedddc7476b}}. The proposed model consisted on the computation of word-embeddings such as Glove, Fasttext, and BERT, which are classified using BiLSTM and time-distributed dense layers. The model was trained using a multi-task learning strategy to recognize at the same time intents and entities from the input's utterance. The authors reported an accuracy of up to 95.7% classifying intents from the ATIS dataset.In {{cite:e08f7e08fb962022ce3a0c47080f2bf4d440a4cd}}, the authors proposed what they called a contextual dialogue encoder based on BERT (DialogueBERT) to address different classification problems that appear in dialogue systems such as intent recognition, entity extraction, and emotion classification. The proposed model consisted on different input embeddings, the same Transformer encoder from BERT, and a layer to encode information from full dialogue using a convolutional layer. The authors reported an accuracy of up to 88.6% for intent recognition, using a dataset with 185,200 requests labeled with 102 intents.",
"Additional studies are focused on comparing the performance of cloud-based providers of NLU services {{cite:6688b07feeeb5cf6b244d455899073c4f9bac6a5}}, {{cite:3c1822ec8c32319188faf28b071809cce3700c76}}. For instance, the authors in {{cite:6688b07feeeb5cf6b244d455899073c4f9bac6a5}} created a benchmark dataset for intent and entity recognition in human-robot interaction in at-home domains. The authors evaluated the performance of several cloud providers of NLU services such as IBM Watson assistant, Dialogflow, LUIS, and Rasa using the benchmark dataset. The results indicated that on intent classification Watson significantly outperforms the other platforms, achieving an accuracy up to 88.8% recognizing 46 intents."
] | [] |
An Experimental Study of the Impact of Pre-training on the Pruning of a
Convolutional Neural Network | In recent years, deep neural networks have known a wide success in variousapplication domains. However, they require important computational and memoryresources, which severely hinders their deployment, notably on mobile devicesor for real-time applications. Neural networks usually involve a large numberof parameters, which correspond to the weights of the network. Such parameters,obtained with the help of a training process, are determinant for theperformance of the network. However, they are also highly redundant. Thepruning methods notably attempt to reduce the size of the parameter set, byidentifying and removing the irrelevant weights. In this paper, we examine theimpact of the training strategy on the pruning efficiency. Two trainingmodalities are considered and compared: (1) fine-tuned and (2) from scratch.The experimental results obtained on four datasets (CIFAR10, CIFAR100, SVHN andCaltech101) and for two different CNNs (VGG16 and MobileNet) demonstrate that anetwork that has been pre-trained on a large corpus (e.g. ImageNet) and thenfine-tuned on a particular dataset can be pruned much more efficiently (up to80% of parameter reduction) than the same network trained from scratch. | [
"Recently, there has been a line of work concerning the lottery ticket hypothesis {{cite:3ab05dd778ccb9507e2d451c49434cea6929276f}}, {{cite:9543db878c199516a40c838a24247d2d99ce6b43}}, {{cite:383bba05f13f92782c0bc32e89e945f8529b9948}}. This hypothesis argues that, in a regular neural network architecture, there exists a sub-network that can be trained to the same level of performance as the original one, as long as it starts from the original initial conditions. This means that a neural network does not require all of its parameters to perform correctly and that having an over-parametrized neural network is only useful to find the \"winning ticket\". In practice, those sub-networks can be found by training the original network to convergence, removing the unnecessary weights, then resetting the value of the remaining weights to their original value.",
"Early work on pruning methods dates back to Optimal Brain Damage {{cite:17775e157221f450fd6b89a0d2194317e295ce79}} and Optimal Brain Surgeon {{cite:040e27563d7e10e3e20e18d96ca77c12d5a64012}}, where the weights are pruned based on the Hessian of the loss function. More recently, {{cite:ba73c15750daeae559425d1d41838749ba8c6e35}} proposes to prune the weights with small magnitude. This kind of pruning, performed on individual weights, is called unstructured pruning {{cite:65b4cf863810415b52977cafb934c6544ade4ea8}}, as there is no intent to preserve any structure or geometry in the network. The unstructured pruning methods are the most efficient as they prune weights at the most fine-grained level. However, they lead to sparse weight matrices. Consequently, taking the advantage of the pruning results requires dedicated hardware or libraries able to efficiently deal with such sparsity.",
"To overcome this limitation, so-called structured pruning methods have been introduced. In this case, the pruning is operated at the level of the filters, kernels or vectors {{cite:9b14b7e5e9a2a99f17fa42b0bc4d985f7b3daf20}}, {{cite:2dc730afce0b8b8bde0a6e694f200e02f63f2481}}. One of the most popular structured pruning technique is the so-called filter pruning. Here, at each pruning stage, complete convolution filters are removed and the convolution structure remains unchanged. The filter selection methods can be based on the filter weight norm {{cite:ebb7356fad5fc783c39f6a35f04c7fa8240d4141}}, average percentage of zeros in the output {{cite:43dec217b2598b2a4e60f1c794848ef250458b9e}}, or on the influence of each channel on the final loss {{cite:7fa31cafde17f14f75d6fcdfdb357e6d2294692a}}. Intrinsic correlation within each layer can also be exploited to remove redundant channels {{cite:188cbf3c46aa2f3da164f7e18c11a928d6031cae}}. Pruning can also be combined with other compression techniques such as weight quantization {{cite:088b7d81e3fb11a25fa8ae8d3a62caef2b71eab9}}, low-rank approximations of weights {{cite:2fff38f3f302021b2e9c0cebb529cf105ea84fb2}}, knowledge distillation {{cite:a063e6302b5aeda2a3eb301d2e36313b95922e96}}, {{cite:a9881b01cd3b1ccf84dc8d03c3f081c56b9b921e}} to further reduce the size of the network.",
"In our work, we have adopted a structured pruning approach, performed at the filter level. As in {{cite:ebb7356fad5fc783c39f6a35f04c7fa8240d4141}}, {{cite:25f5660372eedb9a96eaefcee22a8ef9aaabc0a3}}, we use the {{formula:943d276e-c68b-4784-a517-dc6962b016c9}} -norm to determine the importance of filters and select those to prune. We decided to adopt an iterative pruning process, where the pruning is performed one layer at a time, and the network is retrained after each pruning phase, to allow it to recover from the loss of some parameters."
] | [] |
Verification of Component-based Systems with Recursive Architectures | We study a sound verification method for parametric component-based systems.The method uses a resource logic, a new formal specification language fordistributed systems consisting of a finite yet unbounded number of components.The logic allows the description of architecture configurations coordinatinginstances of a finite number of types of components, by means of inductivedefinitions similar to the ones used to describe algebraic data types orrecursive data structures. For parametric systems specified in this logic, weshow that decision problems such as reaching deadlock or violating criticalsection are undecidable, in general. Despite this negative result, we providefor these decision problems practical semi-algorithms relying on the automaticsynthesis of structural invariants allowing the proof of general safetyproperties. The invariants are defined using the WSkS fragment of the monadicsecond order logic, known to be decidable by a classical automata-logicconnection, thus reducing a verification problem to checking satisfiability ofa WSkS formula. | [
"This paper reports on a resource logic for the specification of setsof configurations of parametric distributed systems with unboundednumbers of components (processes) and a verification method for safetyproperties, such as absence of deadlocks and critical sectionviolations.",
"Traditionally, verification of unbounded networks of parallelprocesses considers known architectural patterns, typically cliques orrings {{cite:d6a31133999ad31b722a91f84fc6576a0c91b7f1}}, {{cite:510d635308962e313812ce101dc2092b0452397b}}. Because the pricefor decidability is drastic restriction on architecture styles {{cite:354a2d7fbad6bee15d1526194b8c3fab50c930ad}}, morerecent works propose practical semi-algorithms, e.g. regularmodel checking{{cite:acc6ec239887c942c763768b3cd61af3ac3ec712}}, {{cite:a2575208bceb1ef2858f0699964677f9b570d583}} orautomata learning {{cite:ebdd5faf72b7f207b803a3f14d97802d59b127b7}}. Here thearchitectural pattern is implicitly determined by the class oflanguage recognizers: word automata encode pipelines or rings, whereastree automata describe trees.",
"A first attempt at specifying architectures by logic is theinteraction logic of Konnov et al. {{cite:2b3e05fab5aedd031af7fec6319669cb7de62439}}, whichis a combination of Presburger arithmetic with monadic uninterpretedfunction symbols, that can describe cliques, stars and rings. Morestructured architectures (pipelines and trees) can be described usinga second-order extension {{cite:b86608923927a747e0d8ffa6dd19d5297f306876}}. As such, theseinteraction logics are undecidable and have no support for automatedverification. Recently, interaction logics that support theverification of safety properties, by structural invariant synthesishave been developed. These logics use fragments of first order logicwith interpreted function symbols that implicitly determine the classof architectures, such as cliques {{cite:4870cfd308f4848b5aa306fbbd709affc9221914}}, orpipelines, rings and trees {{cite:b9d4363046b86745d104de5cf5fb161cf79a0e5e}}.",
"From a theoretical point of view, our resource logic with inductivedefinitions is strictly more expressive: for instance, a chain ofcomponents where a certain component type occurs on all even (odd)positions is provably not expressible in first order logic, but can beeasily defined using our language, whose inductive definitions allowto describe second order constructs in a controlled manner (ratherthan using unrestricted second order quantification, as in{{cite:b86608923927a747e0d8ffa6dd19d5297f306876}}). Moreover, first order logic with successorfunctions can describe at most tree-like architectures, whereas ourlanguage describes structures more general than treesForinstance, the tree-shaped architecture with leaves linked in a ring§., using no interpreted function symbols, otherthan ports involved in interactions {{cite:4870cfd308f4848b5aa306fbbd709affc9221914}}, {{cite:b9d4363046b86745d104de5cf5fb161cf79a0e5e}}.",
"Specifying parameterized component-based systems by inductivedefinitions is, however, not new. Network grammars{{cite:d1121d21ce401499f13d337497bc51d16251e9f0}} use context-free grammar rules to describesystems with linear (pipeline, token-ring) architectures obtained bycomposition of an unbounded number of processes. In contrast, we usepredicate symbols of unrestricted arities to describe architecturalpatterns that are, in general, more complex than trees. Such complexstructures can be specified recursively using graph grammar rules withparameter variables {{cite:1575792f7d5ef8dd6b3dc9136b977cb52bed1cd0}}, {{cite:c7cc7ed96ad33c77f22bc6f1cc199551e2ac7687}}. To avoid clashes, thesevariables must be renamed to unique names and assigned unique indicesat each unfolding step. Our specifications use quantifiers to avoidname clashes and separating conjunction to guarantee that thecomponents and interactions obtained by the unfolding of the rules areunique.",
"Verification of network grammars against safety properties requiresthe synthesis of network invariants {{cite:4ec24a67054bd80936ab1bbfef752267b6a13327}},computed by rather costly fixpoint iterations{{cite:474458aa688490d8f4d739f2c8b14871f640d9e8}} or by abstracting (forgetting theparticular values of indices in) the composition of a small boundednumber of instances {{cite:3213fc5df43561fa0e6780439f921772be574e15}}. Instead, ourmethod uses lightweight structural invariants, that aresynthesized with little computational effort and prove to be efficientin many practical examples{{cite:b9d4363046b86745d104de5cf5fb161cf79a0e5e}}, {{cite:6d499b8defca6ec35dadf05a72e923c60f406818}}."
] | [] |
Simple Text Detoxification by Identifying a Linear Toxic Subspace in
Language Model Embeddings | Large pre-trained language models are often trained on large volumes ofinternet data, some of which may contain toxic or abusive language.Consequently, language models encode toxic information, which makes thereal-world usage of these language models limited. Current methods aim toprevent toxic features from appearing generated text. We hypothesize theexistence of a low-dimensional toxic subspace in the latent space ofpre-trained language models, the existence of which suggests that toxicfeatures follow some underlying pattern and are thus removable. To constructthis toxic subspace, we propose a method to generalize toxic directions in thelatent space. We also provide a methodology for constructing parallel datasetsusing a context based word masking system. Through our experiments, we showthat when the toxic subspace is removed from a set of sentence representations,almost no toxic representations remain in the result. We demonstrateempirically that the subspace found using our method generalizes to multipletoxicity corpora, indicating the existence of a low-dimensional toxic subspace. | [
"Probings of pretrained language models such as BERT and GPT-2 find that these language models associate specific contexts with toxic language {{cite:63c1daf0976a8a377889c2c7e186f5726d372649}}, {{cite:fa08bd204d83f789de5936dbdc7ffabaa2e4b998}}.These probing techniques create a prompt in the format of “a [protected group] [does a specific action] because they...\" where different protected groups and actions can be used in the prompt.A language model such as GPT-2 must then complete the prompt, after which, the continuations are analyzed.This technique resembles the problem of toxic degeneration {{cite:aa31a340a906dae90a316fdfaa0d3b404668ca03}}, wherein a non-toxic prompt may still induce a language model to generate toxic text.For instance, within Real Toxicity Prompts, a dataset of 100k sentences each split in half into a prompt and continuation, toxicity was generally found to be confined to either the prompt or the continuation, not both, suggesting that toxic continuations can result from non-toxic prompts {{cite:aa31a340a906dae90a316fdfaa0d3b404668ca03}}.",
"Furthermore, if toxicity is generally not present in either the prompt-half of a toxic sentence or the continuation-half of a toxic sentence, this implies that toxicity is concentrated in a smaller subset of words or phrases within that sentence.Either toxicity is concentrated in sentences in a predictable manner, or certain contexts cause toxicity more than other contexts, suggesting that toxicity can be explained by a lower dimensional subspace in language model embeddings which can then be removed.",
"Removing a subspace responsible for a certain attribute has been well studied in the context of language model debiasing {{cite:d644b9e7df4393edd6cc90c2ca95a723c3515a43}}, {{cite:f1252456d5a1890b65ab9a405ad9b5fc8e37d425}}, {{cite:f44c56b28886494e9781bc0cc61a10970fe25f9d}}.These approaches remove bias by first identifying bias vectors for each sentence.Principal component analysis is performed on these vectors to produce a bias subspace.To remove bias from an input sentence, the sentence's bias component is identified using the bias subspace and removed.A number of methods have been proposed to find the aforementioned bias vectors.In particular, {{cite:f44c56b28886494e9781bc0cc61a10970fe25f9d}} creates a parallel training corpus for male/female debiasing by swapping gender pronouns (eg. “she was a doctor” and “he was a doctor”), then computing the difference between each sentence embedding.However, we cannot do this for toxicity for three reasons: (1) there are many offensive words, (2) a word’s offensiveness often depends on context, (3) offensive words do not have a corresponding “unoffensive version.”Thus, we propose our own method of creating a parallel corpus and removing toxic features."
] | [] |
Decomposing Natural Logic Inferences in Neural NLI | In the interest of interpreting neural NLI models and their reasoningstrategies, we carry out a systematic probing study which investigates whetherthese models capture the crucial semantic features central to natural logic:monotonicity and concept inclusion. Correctly identifying valid inferences indownward-monotone contexts is a known stumbling block for NLI performance,subsuming linguistic phenomena such as negation scope and generalizedquantifiers. To understand this difficulty, we emphasize monotonicity as aproperty of a context and examine the extent to which models capturemonotonicity information in the contextual embeddings which are intermediate totheir decision making process. Drawing on the recent advancement of the probingparadigm, we compare the presence of monotonicity features across variousmodels. We find that monotonicity information is notably weak in therepresentations of popular NLI models which achieve high scores on benchmarks,and observe that previous improvements to these models based on fine-tuningstrategies have introduced stronger monotonicity features together with theirimproved performance on challenge sets. | [
"Natural logic dates back to the formalisms of {{cite:660d6f63716a4ad2c642ad5e9e6c647cf740689c}}, but has been received more recent treatments and reformulations in {{cite:a23a0ef28c95955a14768b554d2b257168ead7c1}}, {{cite:724d38d38e2f3886025bb88f1edb912ddbfbb430}}.Symbolic and hybrid neural/symbolic implementations of the natural logic paradigm have been explored in {{cite:f973b994737e90bc7319ed6017301199f04b67a1}}, {{cite:a737196ec4db35c062df4bf82195794537e2a6b3}}, {{cite:0026efd3958cdd2a4e1bc1a3ca15c26d0915fce7}}, {{cite:71dbb780cc3e6f0ebd098268e93a191ea36a2051}}.",
"The shortcomings of natural logic handling in various neural NLI models have been shown with several behavioural studies, where NLI challenge setsexhibiting examples of downward monotone reasoning are used to evaluate performance of models with respect to thesereasoning patterns {{cite:2ec72d77eae76d8f1c0a929d093d93dc3ceb9b2c}}, {{cite:6c1b2ded53200b410c866b528cf0feda750a2c12}}, {{cite:cde26d7d8f616de86b94b13faafc40e8c6d016d4}}, {{cite:89f47e4a198e2f171bab119d2b6c7a8e8f68adff}}, {{cite:3bd605bad2412786734014e497e7d7f973cc9574}}.",
"In an attempt to better identify features that neural models manage or fail to capture,researchers have employed probing strategies: namely, the diagnostic classification {{cite:7cb97011f72aa48d9112855637822a838c02de11}}of auxiliary feature labels from internal model representations.Most probing studies in natural language processing focus on thesyntactic features captured in transformer-based languagemodels {{cite:421132ad14e7fa010d05ffef495dfd9a9392f437}}, but calls have been made for more sophisticated probing tasks whichrely more on contextual information {{cite:30140b9d3ff0419277a09ab1e43e73740ef81b61}}.",
"In the realm of semantics, probing studies have focused more on lexical semantics {{cite:7bdc6832afe485662a88705068602c9715f31acf}}:word pair relations are central to monotonicity reasoning, and thus form part of our probing study as well,but the novelty of our work is the task of classifying context monotonicity from contextual word embeddings. Due to its context-sensitive nature, it cannot be learnt by “memorizing\" the labels of specificwords in the training data, a key shortcoming in probing studies which focus on tasks such as POS tagging and word-pair relation classification, which have much less dependency on context."
] | [] |
Generalization Bounds for Stochastic Gradient Langevin Dynamics: A
Unified View via Information Leakage Analysis | Recently, generalization bounds of the non-convex empirical risk minimizationparadigm using Stochastic Gradient Langevin Dynamics (SGLD) have beenextensively studied. Several theoretical frameworks have been presented tostudy this problem from different perspectives, such as information theory andstability. In this paper, we present a unified view from privacy leakageanalysis to investigate the generalization bounds of SGLD, along with atheoretical framework for re-deriving previous results in a succinct manner. Aside from theoretical findings, we conduct various numerical studies toempirically assess the information leakage issue of SGLD. Additionally, ourtheoretical and empirical results provide explanations for prior works thatstudy the membership privacy of SGLD. | [
"There are two most related direction to our works. One is to study generalization abilityof SGLD. Another is to study the connection between generalization and differential privacy ingeneral case. In this part, we highlight the differences between our work and these most related works.",
"In contrast to research focusing on the generalization bounds of SGLD, we aim to unify this line of works through the information leakage analysis of SGLD. This unified view can help tosimplify the derivation of generalization bounds of SGLD and provide new insights to improve these bounds.",
"There are another research line focusing onbuilding the connection between generalization bounds and differential privacy in the general case instead of deriving concrete generalization bounds {{cite:0497e8e4f4372eeb2100b565fd919d492087be05}}, {{cite:a6c070cc3f721db63a6c23e0c01731232192ab63}}. For example, there are a series of work on adaptive data analysis {{cite:0497e8e4f4372eeb2100b565fd919d492087be05}}, {{cite:2c813fbebbcb64b879a1978d5ba964a7c49c83ec}}, {{cite:bd9a8841d072e4a55177d6118d589453f03aafa8}}, {{cite:c919f49d2c148b6b7ec887a79bdf17f3a848be70}}, {{cite:a24d9e771581ab490810e5f9c9da571d768723a0}}, which study false discovery and generalization error led by different settings of adaptivity. While both of the works rely on techniques from differential privacy, in a specific case of SGLD, we need to carefully derive its information leakagein terms of Rényi differential privacy under some mild conditions, i.e., Theorem 1 and its derivation. Once we obtain its privacy leakage, we can employ previous works to buildconcrete generalization bound.",
"In this part, we briefly review some related works on the theoretical properties of SGLD.",
"Convergence Property.In early times, SGLD is proposed to enable Bayesian learning on large scale datasets and can be seen as an SG-MCMC method, which can approximate the process of sampling from the intractable posterior distribution (e.g., {{formula:a4a16b69-d648-4726-8b9e-1d9ef52e2dd1}} in our paper).One of the important properties of SGLD is that the parameters learned using SGLD will approach samples from the parameterposterior distribution. Welling et al. {{cite:1c522b1744f0d0145b67c0ccac3d8eab83f93c2a}} have shown some intuitive observations of this property.Subsequent works have theoretically proved this proposition and have presented various convergence analysis, i.e., how fast the learned parameter can approach the posterior distribution. The key idea of these works is to treat SGLD as the discretization of some specific stochastic differential equations (SDE, e.g., Fokker-Planck equation) and turn to study the discretization error {{cite:7c3706da738379cc93323673a5406a0a8afbe85e}}, {{cite:61c41b7c982f4b8d3b1b7d872b877af8dbde4afb}}.",
"Generalization Bounds of SGLD.More recently, the generalization bound of SGLD has emerged as an active topic. Most of the previous works study this problem from two theoretical aspects, the stability-based theory and information-theoretic quantities.A representative work based on stability-based theory is conducted by Mou et al. {{cite:11b7811941058c8878c0f1c18beff67ba2d697cf}}, in which they have derived that SGLD satisfies {{formula:727e30e7-3f8c-4c4b-9163-a6ba0a695b71}} uniform stability with respect to the Hellinger loss. Then the expected generalization error is bounded by {{formula:fd52e1fb-de75-45c5-9cda-7bcc8134efc2}} according to the stability-based theory {{cite:7584e418b61963c909d4ea86f9822003c4925fda}}.",
"Another research line studies the generalizability of SGLD from an information-theoretic view.The core idea is to bound the generalization error via someinformation measures. For example, an early work {{cite:f7b5331c37506d23e1d5af9c5cf35186b59f2ccd}} proposes to bound the expected generalization error in terms of the mutual information between the training dataset and the output hypothesis of the learning algorithm.Bu et al. {{cite:b677b06eca626aa44be40ad0ec89a6042b4fda2f}} introduce the individual sample mutual information and touch it to the expectedgeneralization error.Besides these theoretical works, there is also empirical evidence showing that SGLD can improve the generalization ability of the model. For example, Gan et al. {{cite:c69e0d964336fbd85f3def9e93b27530bb3709a6}} propose using SGLD to mitigate the overfitting of RNN models for language modeling tasks.",
"Privacy Issues in SGLD.There are also some related works that study the privacy issues in SGLD. Wang et al. {{cite:f8510f48f8d624c911bc815818b454de664e36c7}} present a variant of SGLDthat satisfies differential privacy. Wu et al. {{cite:c1044b34f71d257f7597b026713f1dca5ed3df07}} further prove that SGLD can naturally prevent the well-known membershipattack which is also conducted in our experiments. These works also point out some implicit connections between privacy leakage and generalizations. In our paper, we further formally build the connectionbetween previous theoretical results via the privacy leakage analysis."
] | [] |
Practical Quantum K-Means Clustering: Performance Analysis and
Applications in Energy Grid Classification | In this work, we aim to solve a practical use-case of unsupervised clusteringwhich has applications in predictive maintenance in the energy operationssector using quantum computers. Using only cloud access to quantum computers,we complete a thorough performance analysis of what some current quantumcomputing systems are capable of for practical applications involvingnon-trivial mid-to-high dimensional datasets. We first benchmark how welldistance estimation can be performed using two different metrics based on theswap-test, using angle and amplitude data embedding. Next, for the clusteringperformance analysis, we generate sets of synthetic data with varying clustervariance and compare simulation to physical hardware results using the twometrics. From the results of this performance analysis, we propose a general,competitive, and parallelized version of quantum $k$-means clustering to avoidsome pitfalls discovered due to noisy hardware and apply the approach to a realenergy grid clustering scenario. Using real-world German electricity grid data,we show that the new approach improves the balanced accuracy of the standardquantum $k$-means clustering by $67.8\%$ with respect to the labeling of theclassical algorithm. | [
"Performing clustering and nearest-neighbor type algorithms using quantum computers has been studied in various contexts. Improving the encoding strategy to work better with IBM's quantum computers was studied by Khan et al. in {{cite:a0850f4b615892350ea6124ca7b4bf258c43fde0}}. In the article, they describe an encoding mechanism for feature vectors and benchmark the approach on IBM's quantum computer. Feature vectors, after PCA is performed, of dimension two are considered and benchmarked with the MNIST dataset using quantum hardware. Using the IonQ quantum hardware, Johri et al. perform data classification using clustering on their trapped ion quantum computer {{cite:3041972ad51a88ef3e3bbc292e7fab871972877f}}. In their work, they define an optimized method for encoding 8-dimensional classical data into the quantum computer and use PCA to benchmark against MNIST data for ten different labels and perform their quantum algorithm for nearest-neighbor classification. In these works, an explicit benchmarking of distance estimation accuracy is not demonstrated. Moreover, the only experiments tested on the IBM quantum system were of two dimensions. In a quantum annealing setting, clustering has also been considered. The authors of {{cite:1d6c87ec64f359d2c370399e806238274a423d7e}} and {{cite:4403fb29e5970752f8cd47f46b4f586a1b7fa951}} map a clustering problem to a quadratic unconstrained binary optimization (QUBO) problem for an adiabatic quantum computer and use hardware to test their approach. Quantum annealing uses a different approach to quantum computing versus the gate based model, indeed no swap-test is involved, and the results from annealing experiments do not paint a clear picture for performance using a universal quantum computing approach.",
"In {{cite:3d27d470d91dd4347fd029b494c8ca9985b57a39}}, Benlamine, et al. review three methods for distance estimation using a quantum approach and benchmark the approaches using nearest-neighbor classification in simulation. Further modified quantum clustering approaches were proposed in {{cite:5efa0065dbeb3766fa11df54b02e2f619713acb9}}, {{cite:eb5af4b02b1d5a4f56dd5ce93d54df134d0894c3}}. These works did not perform tests on real physical hardware. In {{cite:f610124a449df3b8882c75719cdd6813b8a4df60}}, Nguyen et al. run experiments to test the accuracy of the swap-test using their trapped ion system in a continuous variable setting, but do not perform any experiments of clustering."
] | [] |
Edge Partitions of Complete Geometric Graphs (Part 2) | Recently, the second and third author showed that complete geometric graphson $2n$ vertices in general cannot be partitioned into $n$ plane spanningtrees. Building up on this work, in this paper, we initiate the study ofpartitioning into beyond planar subgraphs, namely into $k$-planar and$k$-quasi-planar subgraphs and obtain first bounds on the number of subgraphsrequired in this setting. | [
"Recently, the second and third author {{cite:267df441b7812200ce3871131a364c5403d1a117}} showed that there are complete geometric graphs that cannot be partitioned into plane spanning trees, while this is always possible for graphs in convex {{cite:5ea39f02e21e74d609c8f78662ecaa8873590c3d}} or regular wheel position {{cite:89074584b082a113db7370a9f67f02009bfe0d95}}, {{cite:67ac2644d955e0549f4a76059607a9b86d3df7fe}}. For the related packing problem where not all edges of the underlying graphs must be covered, Biniaz and García {{cite:a4bbab3fac3eda4add297a5a46b1c2b7674488f4}} showed that {{formula:512a18ee-80d2-4312-8036-fa803787f07e}} plane spanning trees can be packed in any complete geometric graph on {{formula:afff7925-e93c-4a5e-910e-0f95cdbfb81b}} vertices, which is the currently best lower bound.",
"Note that the problem of partitioning a geometric graph into plane subgraphs is equivalent to a classic edge coloring problem, where each edge should be assigned a color in such a way that no two edges of the same color cross (of course using as few colors as possible). This problem received considerable attention from a variety of perspectives (see for example {{cite:7e3b8e8c0c6cfec16f13de55fcf6c5431e391449}} and references therein) and is the topic of the CG:SHOP challenge 2022 {{cite:42abcce2d5d9d583c1476febfa086ef05754eacd}}."
] | [] |
Algorithms for Adaptive Experiments that Trade-off Statistical Analysis
with Reward: Combining Uniform Random Assignment and Reward Maximization | Multi-armed bandit algorithms like Thompson Sampling can be used to conductadaptive experiments, in which maximizing reward means that data is used toprogressively assign more participants to more effective arms. Such assignmentstrategies increase the risk of statistical hypothesis tests identifying adifference between arms when there is not one, and failing to conclude there isa difference in arms when there truly is one. We present simulations for 2-armexperiments that explore two algorithms that combine the benefits of uniformrandomization for statistical analysis, with the benefits of rewardmaximization achieved by Thompson Sampling (TS). First, Top-Two ThompsonSampling adds a fixed amount of uniform random allocation (UR) spread evenlyover time. Second, a novel heuristic algorithm, called TS PostDiff (PosteriorProbability of Difference). TS PostDiff takes a Bayesian approach to mixing TSand UR: the probability a participant is assigned using UR allocation is theposterior probability that the difference between two arms is `small' (below acertain threshold), allowing for more UR exploration when there is little or noreward to be gained. We find that TS PostDiff method performs well acrossmultiple effect sizes, and thus does not require tuning based on a guess forthe true effect size. | [
"Given the well-documented challenges of drawing inferences from adaptively collected data {{cite:ad7d8c2335888e7d3ed618546fbc16c14446b250}}, {{cite:3e3e9b56810378fd2ca462cce598fa87aee050e4}}, {{cite:c837fe8ed3c65d33b222d7bd653d508375c06bdc}}, {{cite:6f1e08be38a837e8ce085602dc4c44a6de9eace4}}, {{cite:540b1d82358cf9607a551f6d261a3d50c25f148f}}, {{cite:fd165e111f1eb0398f5dcc126202cc0750bde495}}, {{cite:ca76eb0e23aa036cc23a417c2d49dc12967f9dde}}, {{cite:1f7cb1568b0022b4711817a8d681b4eddec70947}}, two general approaches have been taken to address this issue: developing methods to make correct inferences directly from data collected using existing adaptive algorithms, or changing the adaptive algorithm to collect better data.Work taking the first approach has primarily focused on characterizing bias from ordinary least squares estimates and reducing that bias through a variety of statistical approaches {{cite:ad7d8c2335888e7d3ed618546fbc16c14446b250}}, {{cite:3e3e9b56810378fd2ca462cce598fa87aee050e4}}, {{cite:540b1d82358cf9607a551f6d261a3d50c25f148f}}, {{cite:c837fe8ed3c65d33b222d7bd653d508375c06bdc}}, {{cite:6f1e08be38a837e8ce085602dc4c44a6de9eace4}}; reducing the bias of an estimator is related to the final performance of a hypothesis test, but does not directly address the hypothesis testing procedure (typically evaluated through FPR and Power). We focus on statistical hypothesis testing because that is the dominant paradigm for analyzing data in digital field experiments {{cite:6f8d6e2e518b27386b07b3d9695c0a9af3191f99}}; future work could explore the impact of adding uniform random allocation on Bayesian analysis. There is reason to believe that the issues in incorrect statistical conclusions extend beyond hypothesis tests, as the issues in bias in estimates of means {{cite:c837fe8ed3c65d33b222d7bd653d508375c06bdc}} and unequal sample sizes are not restricted to one particular kind of test.",
"More closely related to this paper is work that considers changing the adaptive algorithm to balance competing goals of minimizing regret and obtaining data that is useful for drawing conclusions. In active exploration {{cite:03b3aca3904abe99ed40e1e7020a1026c260dec3}}, {{cite:f8924cab7bbacc0d9eab2f7a59d4bed5493cab02}}, the objective is to estimate the value of all arms as accurately as possible, and thus ensure low estimation errors. The tradeoff between active exploration and maximizing cumulative reward was first addressed using a heuristic algorithm {{cite:683a1be25244786dc5193fe550851decaee78adb}}, and later formalized as a multi-objective bandit problem integrating rewards and estimation errors {{cite:1b0f7a806bdbd0d9fd6f734be90040824b3189ab}} which lead to a rigorous, but deterministic, strategy with theoretical guarantees. The work in this paper is complementary to {{cite:1b0f7a806bdbd0d9fd6f734be90040824b3189ab}} as it focuses on randomized, rather than deterministic strategies, which are likely to be of interest to researchers who typically conduct experiments.",
"Several recent papers have directly tackled how to collect data that is amenable to statistical hypothesis testing. The Power-constrained bandits algorithm {{cite:3336c097f74e9932afdba3849aa94149849f1807}} explores making multiple allocation decisions for one participant, with the bandit aiming to guarantee the Power of a later hypothesis test while also minimizing regret. Other work {{cite:35f4062af9025c93e88aac7642cb24967b1e7371}} uses a finite horizon MDP approach to minimize regret while constraining the minimum number of participants assigned to each arm; they show this leads to improved Power and lower estimation errors compared to an adaptive algorithm without this constraint. Top-2 TS is a high-performing algorithm that introduces a fixed amount of additional exploration of the second-best arm as modeled by TS {{cite:d80dffff5f859bc47e8742eb6693d3733b1b6e5d}}. In contrast to the fixed amount of UR introduced by Top-2 TS, we introduce the TS PostDiff algorithm which adaptively varies the amount of UR and compare it to the state-of-the-art Top-2 TS."
] | [] |
Quantum Algorithms for Reinforcement Learning with a Generative Model | Reinforcement learning studies how an agent should interact with anenvironment to maximize its cumulative reward. A standard way to study thisquestion abstractly is to ask how many samples an agent needs from theenvironment to learn an optimal policy for a $\gamma$-discounted Markovdecision process (MDP). For such an MDP, we design quantum algorithms thatapproximate an optimal policy ($\pi^*$), the optimal value function ($v^*$),and the optimal $Q$-function ($q^*$), assuming the algorithms can accesssamples from the environment in quantum superposition. This assumption isjustified whenever there exists a simulator for the environment; for example,if the environment is a video game or some other program. Our quantumalgorithms, inspired by value iteration, achieve quadratic speedups over thebest-possible classical sample complexities in the approximation accuracy($\epsilon$) and two main parameters of the MDP: the effective time horizon($\frac{1}{1-\gamma}$) and the size of the action space ($A$). Moreover, weshow that our quantum algorithm for computing $q^*$ is optimal by proving amatching quantum lower bound. | [
"As we have discussed, our quantum algorithms can be viewed as “quantizations” of the classical algorithms and techniques in {{cite:aedbc11dae52e07dc5e230eb86ed198ec05d7f10}}, {{cite:8ab62c1b4aa7b892f37eee6f0bc821a5703ddb79}}, {{cite:948715ab98033120b5392bb5e31953fb07b97f1c}} which represent the latest development of classical model-free MDP solvers, which also recently include {{cite:94817cb8caad00942edf749eaf7bbdb16e4cff96}}, {{cite:0b1ebc87c55c5f84a0d99e4d2ac765da9c30fe6a}}, {{cite:349e7141f6b9eacdca3e737d76aa8eb22611affd}} among others, that started with {{cite:f18c3faa11eb35b5f85fe4d413cb9e49e5d5d805}}. {{cite:aedbc11dae52e07dc5e230eb86ed198ec05d7f10}} give algorithms with complexity {{formula:5943e298-b50c-4e69-93c9-45482d8c370a}} when {{formula:2101fb87-4e78-4c8f-bcce-46020155e0f7}} for approximating all three of {{formula:5f97f700-d101-4661-bdba-16cb23cebc8c}} , {{formula:14628b51-dfa3-4f77-8dc0-3d50fd71d543}} , and {{formula:cad4825c-3b16-4857-b465-22559cdcdece}} . On the model-free side, there has been even more recent progress culminating in the work of {{cite:938cd7cd556b9c38477ac109e6515b9cee5ae13c}} which achieves {{formula:021e6ff6-888e-44eb-96d6-28f3030c80fb}} for the full range of {{formula:ebede86a-a78e-48cc-8278-c3442d86b0ed}} . That this bound is tight (up to log-factors) is established by {{cite:2d9a9779fab22b5d5380e3262c613de25627bb81}} which is closely related to our work. Indeed, to prove our lower bounds, we use an instance inspired by {{cite:2d9a9779fab22b5d5380e3262c613de25627bb81}}. However, our proof by reduction and composition theorems is technically quite different from theirs and extends their lower bound to apply to arbitrary {{formula:cf7eb66f-f8c9-4af6-b5f7-07f7dd9e33da}} and {{formula:72aaad43-b419-473d-af42-c27f636e80d9}} . Arguably, model-based MDP solvers {{cite:2d9a9779fab22b5d5380e3262c613de25627bb81}}, {{cite:8ae00e302551f86e678d59fc69adea5dc73a4a5c}}, {{cite:938cd7cd556b9c38477ac109e6515b9cee5ae13c}} have seen more successes than their model-based counterparts that we quantized. However, quantizing these techniques appears more difficult. As a first step, one might ask if the quantum sample complexity of learning a probability distribution supported on {{formula:13b6bce5-1d17-420c-ae53-59c234ccadbe}} points to error {{formula:020ebf11-0907-45bb-ae53-0f845df5527e}} in {{formula:5bf1f8fe-9ecb-4670-8661-5d550e43e6ea}} -norm can be {{formula:7da0a104-5ab3-4676-996b-4f34576e9bf2}} , which represents a quadratic speedup over classical in terms of {{formula:e6c0f48d-dfcc-4357-b506-e29dfc3fb166}} . Recently, this question has been answered affirmatively {{cite:fe9730b716d4b5766ad0f324cb06204e986ef100}}, {{cite:719ed834f3a34a49f4412488ea21e02b6df41cca}} which immediately implies an {{formula:d26a2f33-a2a9-4c0e-800e-38ae39cb3ad5}} model-based quantum algorithm for {{formula:908f42ce-fb33-48f7-b9d1-97c0cb830f67}} due to {{cite:49ad0a723e86a7ec6a848c63e0ca269c22d55818}}. However, this complexity is highly suboptimal and it remains to be seen whether we could eventually obtain an optimal model-based quantum algorithm for any one of {{formula:c52e6fc2-22eb-4461-9012-b65562685d4e}} , {{formula:e0bf4e70-ce05-480c-bc9f-5de5f7bf6391}} , or {{formula:97eae664-7d6f-4730-96b0-9768db1436da}} .",
"On the quantum side, the broader subject of reinforcement learning “remains relatively unaddressed by the quantum community” {{cite:21a4e2db285bed400067dc6b21a161a6f9618f6d}}. The relatively few works on the subject include {{cite:f8370521ad6ec4a958544be680c6f2b7d619f10d}}, {{cite:d10be04f1ceca2c46789a83ab3d7b2d91f866cdd}}, {{cite:fc47ec1a37cc73c5ea75fcaa64836b7b1e677cb0}}, {{cite:227af725fa0709d19ca13877005ee9d7628251dd}}, {{cite:21a4e2db285bed400067dc6b21a161a6f9618f6d}}. However, these works are incomparable to ours as they focus either on problem formulation or lack rigorous results. None give rigorous complexity bounds on computing {{formula:0a2ca673-52b1-46c1-9cea-82fefffde83a}} , {{formula:1a5a9b6c-71e3-40ab-b782-e6cebf80bc86}} , and {{formula:97e235c7-32b2-4fdf-9334-e0a87dff9685}} . Some of these works do mention the possibility of quadratic speedups by using quantum maximum finding {{cite:d10be04f1ceca2c46789a83ab3d7b2d91f866cdd}}. However, they do not consider how this technique would work within an integrated algorithm. As we have mentioned, our work shows that to achieve optimal {{formula:0901a3ae-6b85-4afd-aab6-8ae911d39d96}} -dependence overall, we may have to forgo the use of quantum maximum finding. We note that in the multi-armed bandits setting, where {{formula:37759a1f-bb3c-428f-807f-8c2c573ef5de}} , an instance-optimal quadratic quantum speedup is shown in {{cite:d1478c107469fb633119f9955cc050c1243275fa}}. In synergy with our work, {{cite:227af725fa0709d19ca13877005ee9d7628251dd}} proposes methods to instantiate the quantum generative model in real physical environments as opposed to being given a classical simulator. If their methods can be realized, our work will have wider applicability."
] | [] |
DSGPT: Domain-Specific Generative Pre-Training of Transformers for Text
Generation in E-commerce Title and Review Summarization | We propose a novel domain-specific generative pre-training (DS-GPT) methodfor text generation and apply it to the product titleand review summarizationproblems on E-commerce mobile display.First, we adopt a decoder-onlytransformer architecture, which fitswell for fine-tuning tasks by combininginput and output all to-gether. Second, we demonstrate utilizing only smallamount of pre-training data in related domains is powerful. Pre-training alanguagemodel from a general corpus such as Wikipedia or the CommonCrawlrequires tremendous time and resource commitment, andcan be wasteful if thedownstream tasks are limited in variety. OurDSGPT is pre-trained on a limiteddataset, the Chinese short textsummarization dataset (LCSTS). Third, our modeldoes not requireproduct-related human-labeled data. For title summarizationtask,the state of art explicitly uses additional background knowledgeintraining and predicting stages. In contrast, our model implic-itly capturesthis knowledge and achieves significant improvementover other methods, afterfine-tuning on the public Taobao.comdataset. For review summarization task, weutilize JD.com in-housedataset, and observe similar improvement over standardmachinetranslation methods which lack the flexibility of fine-tuning.Ourproposed work can be simply extended to other domains for a widerange oftext generation tasks. | [
"Automatic text summarization methods can be categorized into two paradigms: the extractive summarization and the abstractive summarization. In extractive summarization, important words or sentences are extracted from the original document and then arranged to form a summary. The important units of text can be detected through classification methods{{cite:2c0f29d36595715dba2042e23e75915c7b048b3a}}, selection methods{{cite:23dd69e8c2e569f97132779608a3edc8ec5c5351}} and ranking methods{{cite:c5afc4498d8d09939aafc4e3d8b1184665a5ed0c}}.",
"In abstractive summarization, the source document is rewritten into a shortened version which preserves the core meaning of the original text{{cite:48483377c054a81bc63f80b0f40f462fe618d668}}. In early studies, abstractive summarization was realized by statistical machine translation techniques{{cite:bed14e89d92deb092c16a716df3cf7f0940dd122}}{{cite:e84089691510ff694abbb701bed294c8ba56c0c3}}. With the development of deep learning, NN-based encoder-decoder frameworks such as attentional Seq2Seq models{{cite:48483377c054a81bc63f80b0f40f462fe618d668}} and Transformers{{cite:fa7911fbf049be29c775b408828a1875d6887ad3}} have pushed abstractive summarization to new states of the art.",
"The OpenAI GPT-2 and GPT-3 {{cite:2accde4d381d71c6562bcf0714102c11c955a9a6}}{{cite:91e160621b73d6c87769931cfb1660402fcbd22f}} are large-scale language models pre-trained on large datasets that can achieve state of the art results via fine-tuning, zero-shot or few-shot. Salesforce research team has proposed CTRL{{cite:552771eace95f7aa2032003fa282e8fb32d96f94}}, which utilizes control codes to realize text generation with different styles and task-specific behaviors. Microsoft has proposed a task-oriented pre-trained model integrated in the dialog system DialoGPT{{cite:5264c3d4841e6a6d7e90337bc5c1bf3e77b59781}}. The open-domain and task-oriented pre-trained models discussed above are all obtained with large amounts of data and compute resources. In this paper, we demonstrate that with small amounts of data and small models we can also get benefits from pre-training.{{figure:1331e621-3046-47e5-9283-b7442bdd6834}}"
] | [] |
DocAMR: Multi-Sentence AMR Representation and Evaluation | Despite extensive research on parsing of English sentences into AbstractionMeaning Representation (AMR) graphs, which are compared to gold graphs via theSmatch metric, full-document parsing into a unified graph representation lackswell-defined representation and evaluation. Taking advantage of asuper-sentential level of coreference annotation from previous work, weintroduce a simple algorithm for deriving a unified graph representation,avoiding the pitfalls of information loss from over-merging and lack ofcoherence from under-merging. Next, we describe improvements to the Smatchmetric to make it tractable for comparing document-level graphs, and use it tore-evaluate the best published document-level AMR parser. We also present apipeline approach combining the top performing AMR parser and coreferenceresolution systems, providing a strong baseline for future research. | [
"MS-AMR Annotations MS-AMR annotations by {{cite:6a6a36b22eeaa7f09641781fe9ec0779bbb29e98}} include coreference chains, implicit roles and bridging relations.In the context of AMR-based summarization, {{cite:33367d5d5bce0743679f3ed386e3e9ed5aa57fe9}} present a novel dataset consisting of human-annotated alignments between the nodes of paired documents and summaries to evaluate merge strategies for merging individual AMR graphs into a document graphs. However, they sought out the merge operations that can serve as cross sentential coreference in the absence of any annotations – they only merge nodes with same surface forms, except for 'person' nodes. Our work, on the other hand, outlines a representation for already available gold annotations, where nodes' surface forms don't match in a large number of cases.",
"MS-AMR Evaluations and Models {{cite:6a6a36b22eeaa7f09641781fe9ec0779bbb29e98}} proposes Smatch as primary method for scoring MS-AMRs. They also report CoNLL-F1 relying on Smatch alignments. Adopting the methods from {{cite:6a6a36b22eeaa7f09641781fe9ec0779bbb29e98}}, {{cite:36178851e35ecad4690647f3078b7e99fb237563}} presented a comparative evaluation of various coreference resolution systems over MS-AMR test sets and document-level Smatch evaluations of machine generated sentence-level AMRs augmented with coreference predictions from various systems.The best approach from their study is incorporated as a baseline in sec:parsing.{{cite:ef0c1400ae0f891690d979a9091eda6bd56edbbb}} introduce an AMR coreference resolution system that uses graph neural network to model gold sentence-level AMR graphs for coreference predictions. This system assumes gold graphs and is not comparable with document-level parsing systems. Use of gold graphs also alleviates the need for alignments between gold and predicted graphs for the purpose of evaluation. {{cite:cf09c387bae9de1ee0ec8af6598a7932d8626dae}} constructed dialogue-level AMR graphs from multiple utterance level AMRs by incorporating inter-sentence coreference, speaker and identical concept information into sentence-level AMRs.",
"Anaphora and Coreference {{cite:9d3910fa412d7de6fd36db01f31ec6e16dfaeba1}} presents a comprehensive review of anaphora and coreference resolution including the scope of the tasks, datasets, evaluation and research methods."
] | [] |
Programmatic Reward Design by Example | Reward design is a fundamental problem in reinforcement learning (RL). Amisspecified or poorly designed reward can result in low sample efficiency andundesired behaviors. In this paper, we propose the idea of programmatic rewarddesign, i.e. using programs to specify the reward functions in RL environments.Programs allow human engineers to express sub-goals and complex task scenariosin a structured and interpretable way. The challenge of programmatic rewarddesign, however, is that while humans can provide the high-level structures,properly setting the low-level details, such as the right amount of reward fora specific sub-task, remains difficult. A major contribution of this paper is aprobabilistic framework that can infer the best candidate programmatic rewardfunction from expert demonstrations. Inspired by recent generative-adversarialapproaches, our framework searches for the most likely programmatic rewardfunction under which the optimally generated trajectories cannot bedifferentiated from the demonstrated trajectories. Experimental results showthat programmatic reward functionslearned using this framework cansignificantly outperform those learned using existing reward learningalgo-rithms, and enable RL agents to achieve state-of-the-artperformance onhighly complex tasks. | [
"Inverse Reinforcement Learning. IRL {{cite:bfd97a20cef305730d940a2ec1347195c6aad509}}, {{cite:8fe610b0530d5f40f11fafdefff41393a8642027}} instantiates a learning-from-demonstrations (LfD) framework to effectively infer the intrinsic reward functions of human experts. It is also notable for having infinite number of solutions. To resolve the ambiguity of IRL, max-entropy {{cite:855b59565feb1f91cca2189a5c9bbf5243f16995}}, max-margin {{cite:bfd97a20cef305730d940a2ec1347195c6aad509}}, {{cite:6be196f273f9d3f2ef0e59a8a7b2bf3ef9bcf4e9}} and Bayesian {{cite:28f8015c3cde7cfe3ee99e145ee3dddbfbc14267}} methods have been proposed. However, those IRL methods assume linear rewards on the basis of non-linear state features,and also call RL in a loop to repeatedly solve the entire environments. The recent works {{cite:47bca4f50783c22557e9358406a6d8cfa7622fa2}}, {{cite:1ddc38b648530b8796ec05f0f101383c125d8a60}}, {{cite:4fc40d31ff80a59ad394ea74ba727fc6e75ae316}}, {{cite:c28bcc71a8009ea457b4ba892a6051bff0570e39}}, drawing a connection between IRL and Generative Adversarial Networks (GANs) {{cite:55df63c296dd2288362fff099b0269341d5b0b8c}}, have achieved substantially improved the scalablity of IRL by using deep neural networks and data-driven approaches. Our work, while embracing a data-driven ideology, further pursues hard coding human knowledge unambiguously and interpretatively in the reward function.",
"Reward Design. Reward shaping {{cite:5d9c774cb5e7d8273f6e8a9dea0b6fe80ac05a1c}} highly speeds up RL training by modifying a sparse reward functions with state-based potential functions. Intrinsic reward generation {{cite:47402c15d6e9f7418a68871db299542d9d977dad}}, {{cite:71efcf966c5ecbebe6917f1f58aa16e606c3833a}}, {{cite:35c11e588f9b14be7f4944f50fe2a85b77997593}} and adversarially-guided {{cite:bc796b128b017acf106604cebfef9582cc8f132e}} techniques aims at motivating the agent to exhaustively explore the environments. We also seek to engineer on the rewards. However, our work, unlike all the listed approaches, does not generate uninterpreted reward functions densely ranging over the entire state space but rather have human knowledge interpretably presented in the reward functions. Logic based reward designs {{cite:8e32555bdfdada3e1b2ac39b994e721fcc9b5487}}, {{cite:6461f2f03e201e0d1775d1dcd5968d29ffd1e50a}}, {{cite:797ec7bad2231d4e6301e93211da5e212ee81c6c}} present human knowledge in reward functions with specification languages such as linear temporal logic (LTL){{cite:a727c2078a13acb125c308522a333ca4617012b5}}. Reward machine theories {{cite:9664b52d375b9f6c0a1208c67713d9578ec850c0}} further directly represent the reward functions as finite state automata which can be translated into logic formulas. Our work distinguishes itself by 1) using programming languages instead of logics to expressively represent human insights in the reward functions; 2) adopting LfD to implement certain details in the reward functions. Regarding LfD, Inverse reward design (IRD) {{cite:be8e6bca72c861e3cc0a5bf29bfcebf2306e5065}} design reward functions in a manner similar to IRL. Safety-aware apprenticeship learning {{cite:a7cc40222102b4c6d22fb0357e84be1d9af5a4f9}} incorporate formal specification and formal verification with IRL. However, those works confine the reward function to be linear of features as IRL does. Our work does not have such confinement.",
"Interpretable Reinforcement Learning. Learning interpretable RL policies has drawn continual attention {{cite:22125acc089070caf0b7155debf353826506d6c3}}, {{cite:824e81b64d44e6cd55c47718da35ee67a3113184}}, {{cite:90779321ed5088eb89c7f621236b6bad60c6b317}}, {{cite:75e11d1a23cd65c790c23a395e5e8fbd66a3a3f2}}, {{cite:f3825fe1952faa9e7fedebc21320a789569963d4}}, {{cite:173b63a0bd510392bdcd4317e497d3125efb3433}}. Our work focuses on programmatic reward functions instead of policies because well-designed reward functions can benefit diverse RL training methods. There have been a variety of works on learning-based program synthesis {{cite:cb8063f93f7227cb1ada507e8752d53b95589ac9}}, {{cite:d0304c249dd99c4309251b5a9ddf82a7c14578fc}}, {{cite:caa040ceb0b67fc9a818b313d262d9d66f0cea75}}, {{cite:4bd131a321bfbb3612ea67d0b30082216520235d}}, {{cite:db1ee97d20cf23ecde72690b23fcbca7949402ac}}.Our work is inspired by the concept of sketch synthesis {{cite:6594e2c4a5845b6a5d63707967b702f94ffbc937}}. Realizing sketch synthesis from example for reward function is our main contribution in this paper."
] | [] |
Cybersecurity Revisited: Honeytokens meet Google Authenticator | Although sufficient authentication mechanisms were enhanced by the use of twoor more factors that resulted in new multi factor authentication schemes, moresophisticated and targeted attacks have shown they are also vulnerable. Thisresearch work proposes a novel two factor authentication system thatincorporates honeytokens into the two factor authentication process. Thecurrent implementation collaborates with Google authenticator. The novelty andsimplicity of the presented approach aims at providing additional layers ofsecurity and protection into a system and thus making it more secure through astronger and more efficient authentication mechanism. | [
"Recently, three techniques for completely removing the possibility of offline password guessing have been proposed: (a) using a machine-dependent function (e.g., Ersatz Passwords {{cite:dbaeb730e4e81cea3cce3d92ba8988409b8968e4}}); (b) using distributed cryptography (e.g., threshold password-authenticated secret sharing {{cite:e8614647bd1a238dd58415120150f8ea1f0eb3c8}}); and (c) using external password-hardening services (e.g., Phoenix {{cite:99a9fc299aa6019fa12e35fd1a4ec808f3bf4010}}). All these techniques, however, need significant modifications to server-side authentication mechanisms. In addition, because they have a lack of scalability, technique (a) does not enable backing up password files where hash values are kept in a distributed way, making the mechanism inappropriate for Internet-scale services. Technique (b) necessitates client-side system changes, which is not user-friendly and widely admitted as not desirable; Technique (c) is subject to a single point of failure and may leak user behavior information to external parties. Many users have a tendency to choose weak passwords, i.e., common words that can be easily guessed by a dictionary attack [{{cite:1120325342203a43a3206458ec328a552b3e3911}},{{cite:73ea070d57c35e499db707ae89be13998093699a}}], when their choices are not monitoring and guided by the authentication mechanism itself.",
"In the literature, the use of honeywords for password protection was first proposed by A. Juels and R. Rivest, in 2013. Later, many works were based on this paper, others expanded it and others improved its vulnerabilities. Honeywords have also been suggested for use in authentication, however, in this paper we recommend their use in a novel two factor authentication scheme.",
"A. Juels and R. Rivest {{cite:5b4252636049c3290dabd66f92426190b47d24c2}} proposed a method to make the security of hashed passwords more powerful. They used honeywords to make password cracking detectable.More specifically, an attacker who achieves to access a file that contains fragmented passwords, inverting the hash function, is unable to distinguish the real ones from honeywords. If the attacker tries to enter a honeyword, a \"honeychecker\", which knows what the honeywords are, is activated and warns for an intrusion attempt.",
"In {{cite:8402c74ed5ec35345822ef65db0fa32ead7adf58}} a new, highly efficient system is proposed, which aims to resolve issues arising from user corruption and server compromisation. Honeywords are adopted to eliminate the long-standing security-usability conflict, and they do user and sensor authentication to resist the node from becoming compromised.",
"In {{cite:fecba9f35c296dabbac3e31bc8fa6dccccce7c3f}} the authors extend the work of A. Juels and R. Rivest by covering some of its gaps. They suggest a solution to the possibility an attacker to modify the honeychecker code, or the code executed on the login server, thus activating the alarm.",
"In {{cite:8007782a6cd16f753d5b7b334a546c044ecba85a}} the work of A. Juels and R. Rivest is coming to the forefront again. The authors here, however, use honeywords not only to protect passwords, but also to detect potential intruders on the system. It is important that they do not just detect, but continue to gather information about the malicious intruder, tracking the source of the problem in the system.",
"Finally, in {{cite:17047a62dd17b6d20bcbce3bfd549c2e89049c96}} the authors conclude that some standard, present authentication schemes intended to authenticate that are based on IIoT, do not meet all the security requirements set by modern era. For this reason, they created a new authentication scheme, much more powerful than the existing ones, which adopts the honeywords technique. This scheme suggests a way for effective detection and thwart node capture attack.",
"The proposed work presented in this article attempts to enhance a traditional 2FA mechanism in order to overcome security problems related to password authentication mechanisms, 2FA mechanisms, and users' problematic practices and behaviors that finally endangers heavily computing systems. The selected approach aims at protecting a system and its users from unauthorized access to resources, from data leakage and disclosure of private information, and provides secure and uninterruptible services to the next possible level. The Honeywords principles are used to detect corrupted or stolen tokens. By introducing also QR codes, the integration of this mechanism into any platform or web application and gain access via a mobile phone can be easily accomplished. This work has been developed over the previously published one, which outlines the main concept {{cite:355622ac40c183de4488b48db6dc59561c01fbc1}}."
] | [] |
ErAConD : Error Annotated Conversational Dialog Dataset for Grammatical
Error Correction | Currently available grammatical error correction (GEC) datasets are compiledusing well-formed written text, limiting the applicability of these datasets toother domains such as informal writing and dialog. In this paper, we present anovel parallel GEC dataset drawn from open-domain chatbot conversations; thisdataset is, to our knowledge, the first GEC dataset targeted to aconversational setting. To demonstrate the utility of the dataset, we use ourannotated data to fine-tune a state-of-the-art GEC model, resulting in a 16point increase in model precision. This is of particular importance in a GECmodel, as model precision is considered more important than recall in GEC taskssince false positives could lead to serious confusion in language learners. Wealso present a detailed annotation scheme which ranks errors by perceivedimpact on comprehensibility, making our dataset both reproducible andextensible. Experimental results show the effectiveness of our data inimproving GEC model performance in conversational scenario. | [
"As with many NLP tasks, the current state-of-the-art in grammatical error correction (GEC) involves using large Transformer-based language models such as BERT {{cite:d23f2f7748596dfb3b3cf1d0da1c0435b4818789}}, RoBERTa {{cite:5e3173deaeb571fcfe485d9299911067207fc41d}}, and XLNet {{cite:b0062d942e6f245afb060c31ed5d3ce464101bed}}. To evaluate the utility of our dataset, we use {{cite:d1f7f5eff989c899ded289928c9cdf7a7cec87a5}}'s GECToR model, which reframes GEC as a sequence labelling task rather than a monolingual machine translation task. GECToR achieves SOTA results on the test corpus used for the BEA 2019 Shared Task on Grammatical Error Correction {{cite:1470562432c78e302ec25e09628eea466aa8ecef}}.Other promising supervised GEC models are proposed by {{cite:9b4a9d69ad40e5cc0935a78fe472b7a5241404b1}} and {{cite:7d40bcca810186235f3b0df0f0869c6ccc94ae72}}, who achieve SOTA results on the JFLEG {{cite:0773277d0de994154495e5a5d7c6d07c03079e3a}} and CoNLL-2014 {{cite:092d60f1e0acdb2ae5c5b9279700df1e99574d7d}} GEC datasets, respectively. Both models achieve these results by combining innovative methods of synthetic data generation with large pretrained transformer language models.",
"Recent work related to the development of datasets for grammatical error correction include {{cite:6e9fe923ff77f64004444a681ca1d50e36c15336}} who present a dataset of native and non-native English writing. {{cite:820fd0b5efc031c4b855d068da190945c478c574}} propose a new parallel dataset of Russian student writing. These datasets add to the growing number of GEC datasets available to the research community. However, no GEC dataset that contains conversational data, in English or any other language, is currently available. We seek to begin closing this gap with the present research."
] | [] |
FIgLib & SmokeyNet: Dataset and Deep Learning Model for Real-Time
Wildland Fire Smoke Detection | The size and frequency of wildland fires in the western United States havedramatically increased in recent years. On high-fire-risk days, a small fireignition can rapidly grow and become out of control. Early detection of fireignitions from initial smoke can assist the response to such fires before theybecome difficult to manage. Past deep learning approaches for wildfire smokedetection have suffered from small or unreliable datasets that make itdifficult to extrapolate performance to real-world scenarios. In this work, wepresent the Fire Ignition Library (FIgLib), a publicly available dataset ofnearly 25,000 labeled wildfire smoke images as seen from fixed-view camerasdeployed in Southern California. We also introduce SmokeyNet, a novel deeplearning architecture using spatiotemporal information from camera imagery forreal-time wildfire smoke detection. When trained on the FIgLib dataset,SmokeyNet outperforms comparable baselines and rivals human performance. Wehope that the availability of the FIgLib dataset and the SmokeyNet architecturewill inspire further research into deep learning methods for wildfire smokedetection, leading to automated notification systems that reduce the time towildfire response. | [
"Before the rise in popularity of deep learning methods, computer vision algorithms leveraging hand-crafted features identified that the visual (e.g., color), spatial, and temporal (i.e., motion) qualities of smoke are essential for the machine detection of wildfires {{cite:a20ed80bdec767f520f536045f3579ad20167228}}, {{cite:676012c6c81a5b8d5b35bc488b5fe8606cbe3a95}}, {{cite:041fd5994e2734b52209698b1e82587262af42a3}}, {{cite:34ded4fedf25d8e5861452e151ed15bed7d79cce}}. More recently, deep learning approaches use a combination of convolutional neural networks (CNNs) {{cite:2f55e81eee8d06a0972fc51f665c44a320546eed}}, {{cite:e17d5812c064b4a6e87f89d54a77f79361587294}}, {{cite:44da363e2067a9ff5383090a10618972933ff6a4}}, {{cite:3ee8cd2610b2a216865962e36cb232b71704efc3}}, {{cite:655ac220a1377664d69784d7d9279747701d214f}}, {{cite:373677b974a5fd720fd04ddf645fcd97b00f14eb}}, {{cite:30628b5fe1aeaff0ef0d493232c7f421ea623540}}, {{cite:ff72399345bc9e16ed8e5d98ada5f10f626af7a4}}, background subtraction {{cite:2f55e81eee8d06a0972fc51f665c44a320546eed}}, {{cite:373677b974a5fd720fd04ddf645fcd97b00f14eb}}, {{cite:2e522394a7fa34f6561fb4348bd9bb1089abd35b}}, and object detection methods {{cite:ff72399345bc9e16ed8e5d98ada5f10f626af7a4}}, {{cite:5f37e36044ac67a80bb9ec5317e29907264b5cf4}}, {{cite:0ff3fbcfed2863f5dfd36cc25b959f4a24cfeee1}}, {{cite:9f0a12fa10ede2d5ae0beb41cef1b03ae23bae46}}, {{cite:a1cdaa32bdc73e2b2d2b24e20cf26ca4b64cad51}} to incorporate visual and spatial features. Long short-term memory (LSTM) networks {{cite:373677b974a5fd720fd04ddf645fcd97b00f14eb}}, {{cite:9f0a12fa10ede2d5ae0beb41cef1b03ae23bae46}} or optical flow {{cite:44da363e2067a9ff5383090a10618972933ff6a4}}, {{cite:2e522394a7fa34f6561fb4348bd9bb1089abd35b}}, {{cite:87a2992d2e51ee291a72c2bb1226cf35ea1338e3}} methods have been applied to incorporate temporal context from video sequences.{{figure:92137481-cfc1-4cfd-8f92-7a3331b49227}}",
"Despite many papers reporting above 90% image classification accuracy for the detection of smoke, the lack of a large, labeled publicly-available wildfire smoke dataset makes it difficult to compare performance between approaches. For example, Ko {{cite:34ded4fedf25d8e5861452e151ed15bed7d79cce}} and Jeong {{cite:9f0a12fa10ede2d5ae0beb41cef1b03ae23bae46}} used only 10 and 24 videos in their test sets, respectively, in which half the videos have smoke and half the videos have no smoke; there are no videos in which a fire starts in the middle of the sequence. With hundreds of frames per video but such few fire scenes evaluated, the high accuracies reported may not be representative of real-world performance across different scenarios. Li {{cite:655ac220a1377664d69784d7d9279747701d214f}} and Park {{cite:30628b5fe1aeaff0ef0d493232c7f421ea623540}} collected 4,595 (36% positive) and 6,354 (22% positive) wildfire images online, respectively; however, since these images were not from video sequences, the smoke plumes in the image are more visible, and likely easier to detect, compared to wildfire smoke initially forming after ignition when seen from a continuous video sequence. Yin {{cite:e17d5812c064b4a6e87f89d54a77f79361587294}} also manually acquired images online, but the images represent smoke from a variety of indoor and outdoor scenarios beyond wildland fires. Many works synthetically generate images to overcome the lack of available data; Park {{cite:30628b5fe1aeaff0ef0d493232c7f421ea623540}} uses generative adversarial networks (GANs), Zhang {{cite:5f37e36044ac67a80bb9ec5317e29907264b5cf4}} uses live smoke in front of a green screen, and Yuan {{cite:49cb0f1e3fe7748ffe46054c407b845f0d17218b}} uses computational simulation to generate these images. Given the diversity and challenges of the datasets across these works, it is hard to identify which models are best or choose a particular dataset to use as a benchmark for wildfire smoke classification.",
"{{cite:dddfb3f5dd7629283377fdff08a849d29894d4c3}} is the only work we are aware of that also uses the FIgLib dataset to evaluate wildfire smoke detection performance. They used an InceptionV3 CNN trained from scratch as the primary image classification architecture. Instead of using a sigmoid threshold of 0.5 for the prediction of smoke, as is common in classification models, a dynamic threshold was implemented based on the average prediction during the same time of day over the prior three days. This was used to incorporate periodic environmental events (e.g. fog, solar reflection, smog, haze); the data from prior days is not included in the FIgLib dataset, but is available through the HPWREN Archive. Using this approach, the authors reported a test accuracy of 0.91 and F1-score of 0.89. However, the test set consists of only a small number of hand-selected images relative to the training set (250 vs. 8,500+ images). The test set also contains images from the same video sequences of fires used in the training set with only 10 minutes (i.e., 10 frames) of separation between them; this procedure bleeds information from the training data and therefore may overstate the performance on the test set. The authors also reported results from field-testing the model on 65 HPWREN cameras over a period of nine days in October 2019. After suppressing repeat detections in a one hour time-span, only 21% of notifications showed smoke from real fires; in other words, their model had a 79% false positive rate on real data. Additionally, there was no report of how many actual fires were missed."
] | [] |
FRUIT: Faithfully Reflecting Updated Information in Text | Textual knowledge bases such as Wikipedia require considerable effort to keepup to date and consistent. While automated writing assistants could potentiallyease this burden, the problem of suggesting edits grounded in externalknowledge has been under-explored. In this paper, we introduce the novelgeneration task of *faithfully reflecting updated information in text* (FRUIT)where the goal is to update an existing article given new evidence. We releasethe FRUIT-WIKI dataset, a collection of over 170K distantly supervised dataproduced from pairs of Wikipedia snapshots, along with our data generationpipeline and a gold evaluation set of 914 instances whose edits are guaranteedto be supported by the evidence. We provide benchmark results for populargeneration systems as well as EDIT5 -- a T5-based approach tailored to editingwe introduce that establishes the state of the art. Our analysis shows thatdeveloping models that can update articles faithfully requires new capabilitiesfor neural generation models, and opens doors to many new applications. | [
"Early work on writing assistants largely focuses on grammar error correction; for a survey see {{cite:96cfdeaf9e6cd839ee18e9087b29d814edae228f}}.Neural models have expanded the capabilities of writing assistants to solve a wider variety of tasks including: autocompletion {{cite:86f380893ae59f3f80db69667516724b1573ea61}}, and following rhetorical directives such as paraphrasing, elaborating, etc. {{cite:955bc00aa91e20355618d3a0be90087cfbd4af78}}.In this work, we seek to expand these capabilities further to producing grounded updates, which has been previously studied by {{cite:43acbf050358928f9f7ac8b8b53a33939e960b8a}}, however only for post-modifier generation.",
"As our primary focus is on writing grounded updates to Wikipedia articles, our work is closely related to existing works on Wikipedia article generation, which generally uses one of two approaches: data-to-text generation {{cite:6e5b21ce35960020d398967d650cc13c5f40cf2e}}, {{cite:4d31ba4246edb87b778b229e72a05117774c03c6}}, {{cite:d9d24a97d8c6f0ee23d67494203ba7887f45d149}}, {{cite:f565382d0c3bde96dee545c78d067c1cbe64c534}}, {{cite:9922f2d81c97553bcc1ef01c3467b460b3b5078e}}, or multi-document summarization {{cite:fb10f90d3f7634d7a963165d10abe72185ddca57}}, {{cite:758de010bfa05174f9e07f965039feb1e3706b53}}, {{cite:39d3350244f2a34978c5528d4074ceb27d961c01}}.In particular, the hyperlink-based approach for associating evidence to articles is directly inspired by these works, and our annotation procedure for removing unsupported text directly draws from {{cite:d9d24a97d8c6f0ee23d67494203ba7887f45d149}}.",
"One of the key requirements in FRUIT is that the model be able to ignore its parametric knowledge to avoid contradicting the provided evidence.An alternative way of addressing these conflicts is to update the parametric knowledge to be consistent with the new information either by making the model \"time-aware\" {{cite:b41e50ca078779547546ee434e28146f8f5b839f}} or by directly editing parameters {{cite:b6258d66a68f25820a8fe07eab4610c99733f0fc}}.Integrating such approaches into a grounded writing assistant is an interesting area for future research.",
"The subtask of determining which facts in the evidence contradict claims in the existing article is a central topic in the large body of work on fact extraction and verification {{cite:081cae83ee2a692bc85b53b35f2b038f1bd6f91c}}.In particular, recently published work by {{cite:af3ecc66a3d81e1a8e53afa410d0a9883170b342}} introduces the Vitamin-C dataset of factual revisions to Wikipedia articles, as well as the task of factually consistent generation.Although closely related to FRUIT, their generation experiments are performed in a restricted setting: updates are performed to single sentences instead of entire articles, evidence contradicting the source sentence is identified ahead of time instead of by the model, and evidence that supplements instead of contradict the source sentence is not considered.",
"Our work is also closely related to the TAC 2008 Update Summarization Task {{cite:48e90cbb223ce0992902580c7187e61f8c0f3464}},whichaddresses the problem of writing a summary of updated information that does not overlap with an existing summaryinstead of updating an existing summary to reflect new information."
] | [] |
A framework for syntactic and semantic quality evaluation of ontologies | The increasing focus on Web 3.0 is leading to automated creation andenrichment of ontologies and other linked datasets. Alongside automation,quality evaluation of enriched ontologies can impact software reliability andreuse. Current quality evaluation approaches oftentimes seek to evaluateontologies in either syntactic (degree of following ontology developmentguidelines) or semantic (degree of semantic validity of enrichedconcepts/relations) aspects. This paper proposes an ontology quality evaluationframework consisting of: (a) SynEvaluator and (b) SemValidator for evaluatingsyntactic and semantic aspects of ontologies respectively. SynEvaluator allowsdynamic task-specific creation and updation of syntactic rules at run-timewithout any need for programming. SemValidator uses Twitter-based expertise ofvalidators for semantic evaluation. The efficacy and validity of the frameworkis shown empirically on multiple ontologies. | [
"Ontology quality evaluation approaches can be broadly classified into a) syntactic and b) semantic quality evaluation approaches.Syntactic evaluation approaches primarily evaluate structural aspects of an ontology based on ontology development guidelines, common pitfalls, structural metrics etc. OntoClean {{cite:423aa617ec81438ce75d5858be402aba334bed39}}, one of the earliest known works in this area, proposed a methodology for validating adequacy of relationships in an ontology based on notions drawn from philosophy such as essence, identity, and unity. Similarly, OntoQA {{cite:1ec575e6791c01d43bd089e00cbc1608d25afb38}} proposed ontology evaluation on the basis of schema metrics and instance metrics. They stated that `goodness' or `validity' of an ontology varies between different users and domains. Gangemi et al. {{cite:bae7e61d4843810d8aed96571e221cb7eeeacdef}} proposed structural, functional and usability-related measures using O{{formula:806dd555-70b0-4d4b-a26a-99511f8ddded}} and oQual, a meta-ontology and an ontology model for ontology evaluation and validation. Burton et al. {{cite:746d11c46911626f66e4ce5d485e6cf435965dce}} proposed an ontology evaluation model based on semiotic theory. In order to apply the metrics proposed in these works, tools such as S-OntoEval {{cite:f8187eab2b6ca38b9911cc97c4c54b0c7335cad4}} drawn from semiotic theory and AktiveRank {{cite:80e8f2212f2d48baa389f9c5e2cfb4ca8221cdb0}} that ranked ontologies based on structural metrics like class match measure, density measure etc. were proposed. However, the tools proposed in these articles are either closed-source prototypes or theoretical frameworks and are not publicly available.",
"There are also a few publicly available closed-source ontology (syntactic) evaluation tools, such as OOPS! {{cite:c894724655b975b8ff07898894fe37ff49d0c967}}, DoORS {{cite:024f63219da12b01cdcc39d2b9b3caa12af90018}} and OntoMetrics {{cite:c83e94531320d6b41d6cc695324998cc8318a648}}. OntOlogy Pitfall Scanner (OOPS!) evaluates ontologies on the basis of established ontology quality rules related to human understanding, logical consistency, real world representation and modelling issues, and manually assigned priorities. DoORS, evaluates ontologies based on metrics drawn from Semiotic Theory while OntoMetrics uses metrics proposed in OntoQA, {{cite:bae7e61d4843810d8aed96571e221cb7eeeacdef}}. These tools, however are not flexible or customizable, and evaluate ontologies using a fixed set of rules or metrics. Users are unable to create/update customized rules and set task-specific priorities, which can be crucial as requirements and application scenarios vary. The available open source implementations {{cite:8cbdacbb0169254ccdaaf55419ed61c62f587ef9}} require creation of new rules and priorities via programming which can be daunting for non-programmers.",
"Semantic evaluation approaches focus on semantic validity of concepts and relationships in an ontology. Traditionally, it has been formulated as a task requiring simple accept/reject decisions from domain experts {{cite:3a4eef5f2c397498f1c0749a36b02452dc1313a5}}. In the past few years, a good number of crowdsourced ontology evaluation approaches have been proposed. Hanika et al. {{cite:cc99e6895864cdb36910a0cc4a34aedb00f778d7}} have developed the UComp Protégé plugin to provide a platform for crowdsourced workers to validate classes, subclasses, properties and instances. Kiptoo et al. {{cite:d6901628989f694a39e3d69ea134373a7bc9dfdf}} use crowdsourcing for axiom and assertion verification in ontologies as well as for verification of subclass-superclass relations by Amazon Mechanical Turks {{cite:cf1850660996196350c6222e5088cba0cc9876c5}}. Pittet et al. used crowdsourced workers to propose changes related to addition, deletion and substitution errors {{cite:95ff0022a7b22b60092ef151bae402e39520289c}}. Zhang et al. {{cite:f480745d5fd30e7e930157cf0db660ee0a8e81ad}} used crowdsourced workers to obtain written feedback (comments/suggestions/references) for making final validation decisions. Requiring complex tasks (such as making data quality decisions or requiring written feedback) from crowdsourced workers can be expensive and unscalable as the size of ontology and/or number of workers increases. Some approaches {{cite:cc99e6895864cdb36910a0cc4a34aedb00f778d7}}, {{cite:9cb0385a6060e99f22d1cc5855896946fdc0afd0}}, {{cite:95ff0022a7b22b60092ef151bae402e39520289c}} used quality control mechanisms like majority voting that are debatable. Noy et al. {{cite:cf1850660996196350c6222e5088cba0cc9876c5}} addressed this by using qualification questions and spam filtering techniques. While these mechanisms eliminate spammers, it may not be applicable where large number of workers have some degree of knowledge but only few are experts. Therefore, an assessment of domain expertise on a continuous scale would be useful as a quality control metric. Further, an integrated quality evaluation framework that seeks to evaluate ontologies on both syntactic and semantic aspects, and also addresses the above-mentioned problems would be useful for a holistic and integrated approach to quality evaluation of enriched ontologies."
] | [] |
QAHOI: Query-Based Anchors for Human-Object Interaction Detection | Human-object interaction (HOI) detection as a downstream of object detectiontasks requires localizing pairs of humans and objects and extracting thesemantic relationships between humans and objects from an image. Recently,one-stage approaches have become a new trend for this task due to their highefficiency. However, these approaches focus on detecting possible interactionpoints or filtering human-object pairs, ignoring the variability in thelocation and size of different objects at spatial scales. To address thisproblem, we propose a transformer-based method, QAHOI (Query-Based Anchors forHuman-Object Interaction detection), which leverages a multi-scale architectureto extract features from different spatial scales and uses query-based anchorsto predict all the elements of an HOI instance. We further investigate that apowerful backbone significantly increases accuracy for QAHOI, and QAHOI with atransformer-based backbone outperforms recent state-of-the-art methods by largemargins on the HICO-DET benchmark. The source code is available at$\href{https://github.com/cjw2021/QAHOI}{\text{this https URL}}$. | [
"Two-stage Approaches. The two-stage approach simplifies the HOI detection to an interaction recognition problem with prepared human-object pairs.The multi-stream architecture was first introduced in the Human-Object Region-based Convolutional Neural Networks (HO-RCNN) {{cite:1a7e551c54e8be0407ec02f6719bd0789ce37c8e}}, which extracts human appearance features, object appearance features, and spatial semantic features separately.Based on the HO-RCNN, subsequent methods improve the multi-stream architecture by incorporating the advanced feature extracting module {{cite:ce782536345e6bfa42cd3e6f42b7b7aa6cfa8b7d}} or the graph module {{cite:7d9580fb3917e95a03fa419a722576882309ae3f}}, {{cite:ccdfe88299f0b327f5a480676aa607c92645e271}}, {{cite:adb1aef11ce8379e6dc8ca7bf9a2b6712704bfe9}}, {{cite:ab2be403903e2878c7bef5d3af08c94afcb40068}} and fusing the human pose information {{cite:203d20bdee9f3e186ae3c5dc1cf7c1241a428216}}, {{cite:f6351aa4b0819d2b96a0184a2f26cbaccafcd3e3}}, {{cite:8b5660a3a9b6895548a5165f7ae4f480435e031e}}.",
"One-stage Approaches. The one-stage approach is proposed with creative designs and adopts a two-branch architecture in general.The Union-level Detector (UnionDet) {{cite:02fd8c97c47aaf1cbc0826a8d027d18bea2fb4fc}} builds a two-branch architecture to detect the union regions of the human-object pairs and the localization of instances in interactions.The Parallel Point Detection and Matching (PPDM) {{cite:f77f3b4660ca562027650a0ac05626524c6a34da}} defines the interaction point as the midpoint of the human and object center points and matches the human and object instances via interaction points.The Glance and Gaze Network (GGNet) {{cite:d814b078e055346b7a748641299d6b3dd0a8067d}} extends the idea of the interaction point by inferring a set of action-aware points (ActPoints) around each pixel of the feature map.However, the methods using interaction points {{cite:f77f3b4660ca562027650a0ac05626524c6a34da}}, {{cite:d814b078e055346b7a748641299d6b3dd0a8067d}}, {{cite:709aafbfcf2fd4539cd0aff4dbf6d9e06db1b955}} or union regions {{cite:02fd8c97c47aaf1cbc0826a8d027d18bea2fb4fc}} require a matching or gathering process to clarify the HOI instance.On the other hand, the transformer-based methods {{cite:4bad7c11fe647eb2dbf8de7185d273d3d91a0dff}}, {{cite:f9fd86db2f023eccba8a834f79110704241d13ad}}, {{cite:97ecb954d526e7ae95a1db34e8234368dfc81309}}, {{cite:3419d3c7109bc9a91d3e9724d06671da9da23ffa}} using the self-attention mechanism of the transformer to extract contextual semantic information and the embeddings to represent the HOI instance become a new trend of the HOI detection task.Tamura {{cite:3419d3c7109bc9a91d3e9724d06671da9da23ffa}} transform the object detection head of the transformer-based object detector DETR {{cite:22d85209ba97cf20a5e5ec9ec23c8e24f0c64d2b}} into a simple interaction detection head to predict all of the elements of the HOI instance directly.Similarly, Zou {{cite:f9fd86db2f023eccba8a834f79110704241d13ad}} combine the CNN backbone and the transformer to predict HOI instances directly from the query embeddings.Both Chen {{cite:97ecb954d526e7ae95a1db34e8234368dfc81309}} and Kim {{cite:4bad7c11fe647eb2dbf8de7185d273d3d91a0dff}} propose a transformed-based two-branch architecture, which constructs an instance decoder and an interaction decoder to decode the boxes and action classes of the HOI instances in parallel."
] | [] |
DREAM: Improving Situational QA by First Elaborating the Situation | When people answer questions about a specific situation, e.g., "I cheated onmy mid-term exam last week. Was that wrong?", cognitive science suggests thatthey form a mental picture of that situation before answering. While we do notknow how language models (LMs) answer such questions, we conjecture that theymay answer more accurately if they are also provided with additional detailsabout the question situation, elaborating the "scene". To test this conjecture,we train a new model, DREAM, to answer questions that elaborate the scenes thatsituated questions are about, and then provide those elaborations as additionalcontext to a question-answering (QA) model. We find that DREAM is able tocreate better scene elaborations (more accurate, useful, and consistent) than arepresentative state-of-the-art, zero-shot model (Macaw). We also find thatusing the scene elaborations as additional context improves the answer accuracyof a downstream QA system, including beyond that obtainable by simply furtherfinetuning the QA system on DREAM's training data. These results suggest thatadding focused elaborations about a situation can improve a system's reasoningabout it, and may serve as an effective way of injecting new scenario basedknowledge into QA models. Finally, our approach is dataset-neutral; we observeimproved QA performance across different models, with even bigger gains onmodels with fewer parameters. We make our dataset and model publicly availableat https://github.com/allenai/dream. | [
"The concept of a mental model - a coherent, internal representation of the world - is common in cognitivescience {{cite:1b8a8ef7d832394e149d35a3ed1a9ca7d15fbdb9}}, {{cite:8232367402c4048a4e1eeb9185a5835b3e17f6ae}}, {{cite:f0c22a8aed16aa91c894391156cecca8d6ec6f7a}}. It suggests that peoplesolve situated problems by elaborating a mental picture of the situation, including elements that maybe peripheral to a specific answer, rather than constructing a deductive proof from a fewkey facts to an answer {{cite:5e85ed06fcb8606caa60b2536629389353953c4e}}.",
"Similarly within AI, the idea that people form an understanding of a scenarioby posing questions has been explored in several contexts.Minsky's frames ({{cite:6c64df0fc675c490c1193943aa8b131b6b820593}}) embodied this idea,where he defined a frame as “a collection of questions to be asked about a hypothetical situation”,and whose answers helped elaborate the situation to aid question-answering. Otherstudies have identified what questions people naturally ask when readingtext {{cite:9f88673665a32c6f17730bffcd13efd948da65ea}} or viewing images {{cite:4ad16fb518309b80fe6b3fc38f333eb9a69aacb4}},to form a mental picture of what might be happening. Our work draws on these ideasto explore how coherent a LM's mental picture is, and how it can be improved.",
"While many QA system use additional, generic knowledge to help question-answering,e.g., using sentences retrieved from a corpus {{cite:49a2256429dc45cebb30d02a09238d73908632ec}}, {{cite:2fb9beb454b3630285f21510204f99c72528785b}},our goal is to use a situation-specific elaboration to improve performance.Self-talk {{cite:a8087006e0e8b07e02b404a880e01f2b2ab10738}} explored elaborating a questionusing the answer to a single subquestion found through self-querying(rather than introducing new knowledge).Our work can be viewed as expanding this elaboration into a largerscenario (mental model) using multiple questions, and using a trainedmodel (DREAM) to generate answers to those questions.",
"Several prior researchers have highlighted the brittleness of LMs,and shown that they frequently rely on data artifacts to answerquestions {{cite:8203ea36e63e237f0062463f5663227702bcd05f}}, {{cite:3904e82efb277cdcdf179f4969078a143d567abc}}.Our work explores the sources of that brittlenessby probing how the model interprets a situated question, and proposesways to partially alleviate those problems."
] | [] |
A Deep Learning Approach for Ontology Enrichment from Unstructured Text | Information Security in the cyber world is a major cause for concern, with asignificant increase in the number of attack surfaces. Existing information onvulnerabilities, attacks, controls, and advisories available on the webprovides an opportunity to represent knowledge and perform security analyticsto mitigate some of the concerns. Representing security knowledge in the formof ontology facilitates anomaly detection, threat intelligence, reasoning andrelevance attribution of attacks, and many more. This necessitates dynamic andautomated enrichment of information security ontologies. However, existingontology enrichment algorithms based on natural language processing and MLmodels have issues with contextual extraction of concepts in words, phrases,and sentences. This motivates the need for sequential Deep Learningarchitectures that traverse through dependency paths in text and extractembedded vulnerabilities, threats, controls, products, and othersecurity-related concepts and instances from learned path representations. Inthe proposed approach, Bidirectional LSTMs trained on a large DBpedia datasetand Wikipedia corpus of 2.8 GB along with Universal Sentence Encoder isdeployed to enrich ISO 27001-based information security ontology. The model istrained and tested on a high-performance computing (HPC) environment to handleWiki text dimensionality. The approach yielded a test accuracy of over 80% whentested with knocked-out concepts from ontology and web page instances tovalidate the robustness. | [
"This section discusses related work on enrichment of ontologies from unstructured text as well as approaches to create and maintain information security ontologies. The work on enrichment of knowledge graphs (KG) from unstructured text is also discussed as it represents knowledge and contains similarities with ontologies.",
"Researchers worked on knowledge acquisition from text to construct ontologies for past couple of decades {{cite:811d3138258ad740967aba4361f27d2d783e97db}}, {{cite:ca6451cec21a58771bd168b1abe175fd63037dcc}}. The last decade witnessed significant progress in the field of information extraction from web with projects such as DBpedia, Freebase and others. The work of Mitchell et. al {{cite:c2fb71b81db0fed1eca0821a03a0405e86162cd0}} known as `NELL' states that it is a never-ending system to learn from web, their work bootstraps knowledge graphs on a continuous basis. Tools such as ReVerb {{cite:d0305586f8b2c16f8f449ea0a98205b40be09663}} and OLLIE {{cite:a0ba65628d0742e24b580f96fddaf3a306dfd614}} are based on open information systems to extract a triple from a sentence using syntactic and lexical patterns. Although these approaches extract triples from unstructured text using shallow and fast models, they do not handle ambiguity while entity mapping and do not learn expressive features compared to deep and multi-layer models.",
"The ML models based on probabilistic, neural networks and others are also explored for ontology enrichment from text {{cite:ca6451cec21a58771bd168b1abe175fd63037dcc}}, {{cite:cd366cab0c6d3b2b961be8e2854bf6c0c3ad0ecd}}, {{cite:d638414549a2ce1efae7d5f87b531b216041b7e6}}. In 2017, Wang et al {{cite:04da02de736f3130578c70614cd56c9a0f7b1f40}} conducted a survey on knowledge graph completion, entity classification and resolution, and relation extraction. The study classified embedding techniques into translational distance models and semantic matching models. The study also stated that additional information in the form of entity types, textual descriptions, relation paths and logical rules strengthen the research. Deep learning models such as CNN {{cite:d2fd58c0cf623bc5584ffec71bdb4fbe938e3c35}}, LSTM {{cite:10c3816cba127f2a006622a9d74951ee8724cbf0}}, {{cite:72b9a717ce60f2cffb28d045e6f283fea5552a30}} and variants are used to construct knowledge graphs from text as they carry memory cells and forget gates to build the context and reduce noise. The work of Vedula et al. {{cite:167fc0e4f4b51d7c5aaf51e2dc6253ab56bf024d}} proposed an approach to bootstrap newer ontologies from related domains.",
"Some of the recent approaches are based on Word2Vec {{cite:dc6fc4c624a55392c1417a3d80ead8cf0bf9d8f1}} and its variants such as Phrase2Vec or Doc2Vec that use distributional similarities to identify concepts to enrich an ontology. However, these approaches underperform in the extraction of concepts embedded in words, phrases and sentences due to their inability to adequately characterize context. Compared to Word2Vec and its variants, Universal Sentence Encoder (USE) {{cite:6028fc09e4e1db47ecb3628cf45f8f6f773434c7}} stands promising to identify concepts in long phrases as it encodes text into high dimensional vectors for semantic similarity. Lately, researchers {{cite:1a258ffd0ddd7970f0ba4ed6b469f1460071b08d}} are exploring USE to produce sentence embeddings and deduce semantic closeness in queries. Although, transformer-based models such as BERT and XLNet {{cite:32548eb09f2db246d0b1a73ce19e4543ed174d9e}}, {{cite:64ac3c42d3b7d42774458bc3bb4e761c5ada5e51}} are of interest to ontology enrichment researchers, training them to a domain is effort intensive.",
"The literature to enrich security ontologies from text drew attention with OASIS's STIX/TAXII standardization and open source threat intelligence. Most of the current work on security ontologies from text (construction or enrichment) are based on usage of string, substring, pre-fix and post-fix matching of terms, Word2Vec and other basic ML models {{cite:9f1e9c7b6cb99b39c74b3ba44f71e2fa413a76a6}}, {{cite:196dc99a99e95436beee00eac5e90c424464a857}}, {{cite:d638414549a2ce1efae7d5f87b531b216041b7e6}}. In ontologies as well, the deep learning approaches based on recurrent neural networks are trending because of their ability to build the context over multiple words {{cite:13aa4d620e5b55cf3f48c0aa7e1cc93e430d265b}}, {{cite:f40f1d8679363b414c86734a0c5247cca5eec20a}}. The research of Houssem et al. {{cite:f40f1d8679363b414c86734a0c5247cca5eec20a}} used LSTM for population of security ontologies. However, the details to create corpus, handle phrases and robustness of the approach are not elaborated, only 40 entities are used in the model. The literature revealed that security ontologies based on ISO 27001 {{cite:10bda834d8b73e2504d8c44dc3c8544e1e014254}} and MITRE Corporation's cyber security effort {{cite:196dc99a99e95436beee00eac5e90c424464a857}} are most referred."
] | [] |
Federated 3GPP Mobile Edge Computing Systems: A Transparent Proxy for
Third Party Authentication with Application Mobility Support | Multi-Access or Mobile Edge Computing (MEC) is being deployed by 4G/5Goperators to provide computational services at lower latencies. Federating MECsacross operators expands capability, capacity, and coverage but gives rise totwo issues - third-party authentication and application mobility - forcontinuous service during roaming without re-authentication. In this work, wepropose a Federated State transfer and 3rd-party Authentication (FS3A)mechanism that uses a transparent proxy to transfer the information of bothauthentication and application state across operators to resolve these issues.The FS3A proxy is kept transparent, with virtual counterparts, to avoid anychanges to the existing MEC and cellular architectures. FS3A provides userswith a token, when authenticated by an MEC, which can be reused acrossoperators for faster authentication. Prefetching of subscription and state isalso proposed to further reduce the authentication and application mobilitylatencies. We evaluated FS3A on an OpenAirInterface (OAI)-based testbed and theresults show that token reuse and subscription prefetching reduce theauthentication latency by 53-65%, compared to complete re-authentication, whilestate prefetching reduces application mobility latency by 51-91%, compared tono prefetching. Overall, FS3A reduces the service interruption time by 33%,compared to no token reuse and prefetching. | [
"In this section we discuss existing 4G LTE architecture with MEC, some of the relevant ETSI MEC standards, a threat model for our problem, and related work.",
"In order to find solutions to the proposed problem, we examined studies that have proposed solutions to at least one of these problems, either 3rd party authentication in federated systems, multi-network authentications, or application mobility across networks. We analyzed the transparency of these proposals according to existing LTE protocols and MEC standards. Our findings are summarized in Table REF . Some studies {{cite:541ae6b4a52735f4ef5452a0b03b347caaca589f}} propose third-party authentication through the cloud. There are also studies {{cite:20c55bfc3d454d01d38163e5246920a90324bc58}}-{{cite:76a7c2b3448a19347d5a96d28e912bf95255f190}} that use blockchain, mutual key exchanges, Reinforcement learning (RL), and federated IDs to manage Multi-network authentication. A study {{cite:4b40afb319ed8a8e8b3dc9c9b3d72c5b0e7354d9}} provides handover authentication by improving the existing EAP-AKA protocol and another study {{cite:e03103ee2e0b5a2f5875b7f1c0ff5cf73097385b}} makes use of tokens to provide service access control in 5G MEC. A few studies {{cite:ab8a43fcaeb180c731f7a4ac35b566f2b0f4c8ef}}{{cite:c2dbecd9a4734066f2914fbd418b590d7e45c25a}} also use SDNs to manage authentication and application mobility in federated networks, while another {{cite:b75e617046514f9e01593fc2f94afd38c81a8d08}} provides application driven handover while retaining transparency.",
"We also looked into studies that propose solutions for application relocation and most of them focus on VM migration, for example {{cite:ab76ea33c943d31459e469dcfe717fd74f6e0938}} and {{cite:c52f6e624e23c85a32b329a2f47f6d636e35ba85}}. On the other hand, recent literature also provides transparent, low-latency solutions for authentication and mobility within a single edge network {{cite:73776d741c14fdbeb21c8b862bec13a9adf1ef1e}}{{cite:a21167c82d85b4cf2bb83a53f32c54c6942ed882}}{{cite:91f0f1bcdabf12634e38617dc666c7622a1e11a7}}, anonymous mutual authentication in MEC platforms {{cite:beefffcbf79872e589d19793637777d068be8e49}}, and third-party authentication between cloud and edge {{cite:0f16f1990c9c7b2d0aca1a36613caee9112dd666}}. Some work has also been done on the deployment of a secure MEC system in a cellular network that can itself provide authentication for its applications {{cite:e0c1bcbde20da355fc8594dde277496cf4b7e29a}}. We use this MEC deployment method in our work and extend it further across federated MNOs.",
"An earlier study from some of us {{cite:73776d741c14fdbeb21c8b862bec13a9adf1ef1e}} provides low-latency authentication and state transfer while moving from one MEC platform to another by using two approaches, termed TC3A and TS3A. In order to compare this work with {{cite:73776d741c14fdbeb21c8b862bec13a9adf1ef1e}}, we summarize the differences in Table REF . It can be seen that TC3A, TS3A, and our proposed FS3A provide third-party authentication and application mobility but FS3A does that in a federated MEC system. The major advantage of FS3A over TS3A and TC3A is that it solves the challenges of subscription and state data prefetching and provides the Inter-MNO MEC connectivity, while TS3A and TC3A work for Intra-MNO MECs that are deployed by the same MNO.",
"Our survey shows that previous studies have provided solution to authentication and application mobility issues independently in the MECs deployed by an MNO but, no previous study provides solutions for authentication and application mobility issues together while maintaining transparency and latency requirements in federated MEC systems across multiple MNOs. It should be noted that we do not try to solve the call and data roaming problems as they have already been addressed by various studies {{cite:8af821641dda954f52c078da3dc7a862319ce8e6}} {{cite:dc8b836d047539a988aa466a6efd28b3c659eaea}} {{cite:9e3b7b3cec394f75dfecad43f5ecd2383f8db82b}}. We solely focus on solving the problems of third party authentication and application mobility for a user that roams between MECs deployed by different MNOs. To the best of our knowledge, ours is the first work that provides a complete solution to the problems at hand, i.e., third-party authentication and application mobility in federated MEC systems while meeting transparency and latency requirements. We adopt the transparent, low-latency authentication solutions from Intra-MNO MEC {{cite:e0c1bcbde20da355fc8594dde277496cf4b7e29a}}{{cite:73776d741c14fdbeb21c8b862bec13a9adf1ef1e}} and enhance them to solve the problem of third-party authentication and state transfer in an Inter-MNO MEC federation."
] | [] |
Call for Customized Conversation: Customized Conversation Grounding
Persona and Knowledge | Humans usually have conversations by making use of prior knowledge about atopic and background information of the people whom they are talking to.However, existing conversational agents and datasets do not consider suchcomprehensive information, and thus they have a limitation in generating theutterances where the knowledge and persona are fused properly. To address thisissue, we introduce a call For Customized conversation (FoCus) dataset wherethe customized answers are built with the user's persona and Wikipediaknowledge. To evaluate the abilities to make informative and customizedutterances of pre-trained language models, we utilize BART and GPT-2 as well astransformer-based models. We assess their generation abilities with automaticscores and conduct human evaluations for qualitative results. We examinewhether the model reflects adequate persona and knowledge with our proposed twosub-tasks, persona grounding (PG) and knowledge grounding (KG). Moreover, weshow that the utterances of our data are constructed with the proper knowledgeand persona through grounding quality assessment. | [
"To build dialog agents that can interact with people in multi-turn conversations, several datasets have been introduced {{cite:4526619745e1fd6f5e134204b06337f75e634e94}}, {{cite:6187740134437508c86159c66520b630943893d9}}, {{cite:b4da644804fa532d9f5a9c0f0c5697beba266ba6}}, {{cite:772b5abd74068396d697616b255d71145c75c171}}, {{cite:6dde4420c84b7cf0787bb5d4886b939898650b38}}, {{cite:91fe9b03c662424b5fd5741aa950ae640e45a75b}}, {{cite:4da56b0f535c18091da949dd9a1948576d8a9711}}, {{cite:e4ef7e666d8a8479302adaeff048aba29ca08749}}. Despite these datasets, the dialog agents merely answer the question without considering the user or specialized knowledge.",
"To generate customized answers to the users, attempts have been made to endow the agent with the user's emotion, preference, and experience {{cite:a6c38dd7ed44c2356a4e1c2d12242fcdccd56e83}}, {{cite:4da56b0f535c18091da949dd9a1948576d8a9711}}, {{cite:a446063f7e1b29e06f8cd509c6a2bc8de31a156b}}, {{cite:a69aaa23062908ab26bf3777f74a2f19fcf1e790}}. {{cite:8f3b3224408ca37c09ca46b7303a31fedbf59cdd}} introduces a dataset that includes each speaker's preference and experience, where persona sentences describing two speakers are given. Because both speakers are only provided with persona sentences, one speaker simply confirms what the other speaker likes or dislikes in the dialog. Even though agents generate answers that react or express sympathy, they hardly give a document-grounded answer that fits the user's preference and experience.",
"While the user-centered dialog datasets have appeared, datasets and agents that aim to improve the level of knowledge in the answer with additional documents has been in parallel released {{cite:1b965e99717afdf13de45823daac1155e78142b8}}, {{cite:06cd6b201a57ed6828fe294b5510acf86c8a8e16}}, {{cite:040c4614bd3ea9c833725b12cd9a5d562c9b05f4}}, {{cite:b0e0c0bec6782821cc3e2b422b87103ab291a728}}, {{cite:1e88c8b04fa74b08502bec80c0158d7d02441a75}}, {{cite:f139d56fb097732518ec0f0e585259f73aec8fc0}}, {{cite:5f4e17accbbea49a58713f178fe3abe8a38d6d81}}, {{cite:860c8c4da7b076a42bef4ab335551e899e6b5b06}}, {{cite:998407fcb638ae65169bc47f1b627df4932a2efe}}. {{cite:1b965e99717afdf13de45823daac1155e78142b8}} is a dialog dataset where the agent retrieves the Wikipedia pages on diverse topics and generates responses to the questions. Although these data have a concept of persona, they do not contain customized answers to the listener. Similar to {{cite:1b965e99717afdf13de45823daac1155e78142b8}}, {{cite:06cd6b201a57ed6828fe294b5510acf86c8a8e16}} introduces a document-grounded dataset that includes specified documents from Wikipedia articles about popular movies. These datasets mainly consist of answering the question without considering the user's information, and it leads to excessive and needless answers.",
"There have been efforts to blend several datasets {{cite:f07d53ad7a0d5300ed6ae1a64ce772d82cf834a1}}, {{cite:ca3003565efedfa1267f19b46b7dc81650ef686d}} to build an intelligent agent which has multiple abilities learned from various datasets. Despite the previous datasets, the capability of machines to respond in a dialog is still insufficient, compared to that of humans. Specifically, to answer a question, retrieving the knowledge while considering the user's background information is beyond current dialog agent's abilities."
] | [] |
Self-supervised Enhancement of Latent Discovery in GANs | Several methods for discovering interpretable directions in the latent spaceof pre-trained GANs have been proposed. Latent semantics discovered byunsupervised methods are relatively less disentangled than supervised methodssince they do not use pre-trained attribute classifiers. We propose ScaleRanking Estimator (SRE), which is trained using self-supervision. SRE enhancesthe disentanglement in directions obtained by existing unsuperviseddisentanglement techniques. These directions are updated to preserve theordering of variation within each direction in latent space. Qualitative andquantitative evaluation of the discovered directions demonstrates that ourproposed method significantly improves disentanglement in various datasets. Wealso show that the learned SRE can be used to perform Attribute-based imageretrieval task without further training. | [
"Generative Adversarial Networks GANs are one of the most popular generative models that shows promising results on image synthesis {{cite:9c8d1eaf896f14c80a49985c4f1a1303af7ab622}}. It consists of a Generator and Discriminator that learns in an adversarial setting. Recent variants of GANs such as StyleGAN {{cite:d73de3187ab306371437eb1c047cc8a59b05657a}}, StyleGAN-2 {{cite:3daee41f9943379980ca9a55983510377c56e4e5}}, Progressive GAN {{cite:38e1f740a8ae4eb1441639b0eedabab0ab3912e9}} and BigGAN {{cite:982213211d0ed55938a9375ac541348c765c75a4}} are shown to be very successful in generating high-resolution images. Progressive GAN, a successor of conventional GAN, attempts to generate high-resolution images by progressively growing the generator and discriminator. StyleGAN and StyleGAN-2 learn a mapping network that maps the {{formula:06fba7e6-e53c-4606-9249-d8fe8b37d525}} -space to {{formula:de80db27-308d-447c-9af5-5f489f429e7d}} -space that is more disentangled.",
"Post-hoc Disentanglement from pretrained GANs {{cite:4d36bbeb809cd12bd30c91af71fe002b5deef787}}, {{cite:79c39b9e341723dc7c616ede44c4070327ad3f20}}, {{cite:3e6534cc2c075459460a7aa6e00ee6d840cd90ce}} e.t.c. disentangle factors of variations very well in synthetic datasets, but they fail to do so in complex natural datasets. These classical disentanglement learning techniques improve disentanglement at the cost of generation quality. To overcome this limitation, extensive research has been conducted in the field of learning interpretable directions from pre-trained models. They can be categorized into three based on the learning paradigm used :",
"Supervised : {{cite:ba4c879f2f4c55ba8b1d203dd7029e6ae28eef2f}} computes the agreement between the output of a pre-trained semantic segmentation network and the spatial location of the unit activation map to identify the concept encoded in each unit.{{cite:6245e678cd65dfc1952c7379ef76306462e977d0}} and {{cite:a65a6d14b41265664100b9cc26e7994ce86fae0b}} use off-the-shelf classifiers to discover interpretable directions in the latent space. A conditional normalizing flow version of {{cite:6245e678cd65dfc1952c7379ef76306462e977d0}} and {{cite:a65a6d14b41265664100b9cc26e7994ce86fae0b}} is explored in {{cite:c00e0230d6e1318b786ef4adfca686f1d7dc6c64}}. The main limitation of the above approaches is that they require pre-trained networks, which may not be available for complex transformations.Unsupervised : {{cite:83bff675fb9d76b4e143fb6eb91dd24c4132591c}} discovers interpretable directions in an unsupervised manner by jointly updating a candidate direction matrix and reconstructor that predicts the perturbed direction. {{cite:3bb655d07d48715ed1c5b6c16566eb1da2648bfe}} proposes a regularization term that forces the Hessian of a generative model with respect to its input to be diagonal. However, such methods require training.{{cite:0e6ac607e39b7bd91281e07955d62b20043e4d78}} observed that applying PCA on the latent space of Style-GAN and BigGAN retrieves human-interpretable directions. {{cite:67befe040d2fe7b6d021dc7ac3f074205d7c3dd3}} and {{cite:d8607f02ec49e8fdca8f04b6cb20581510abb388}} obtained a closed-form solution by extracting the interpretable directions from the weight matrices of pretrained generators. These methods are computationally inexpensive since they do not require any form of training. {{cite:83bff675fb9d76b4e143fb6eb91dd24c4132591c}} and {{cite:3bb655d07d48715ed1c5b6c16566eb1da2648bfe}} attempts to learn directions that are easily distinguishable while {{cite:67befe040d2fe7b6d021dc7ac3f074205d7c3dd3}}, {{cite:d8607f02ec49e8fdca8f04b6cb20581510abb388}} and {{cite:0e6ac607e39b7bd91281e07955d62b20043e4d78}} finds directions of maximum variance. However, none of these approaches ensure that only a single factor of variation gets captured in a transformation. Our method addresses this problem by defining a self-supervision task that promotes disentanglement on directions captured by these methods.Self-supervised : {{cite:b9b47c0a34abcd21a2776f304ee22c704e07f214}} and {{cite:8aeffd004dbf007ce542b4f0d769a365776edd9a}} make use of user-specified simple transformations as a source of self-supervision to learn corresponding directions. The main drawback of these approaches is that, user-specified edits are hard to obtain for complex transformations. Unlike these methods, our method relies on transformations discovered by unsupervised methods and hence can discover a wide variety of disentangled transformations."
] | [] |
AcTune: Uncertainty-aware Active Self-Training for Semi-Supervised
Active Learning with Pretrained Language Models | While pre-trained language model (PLM) fine-tuning has achieved strongperformance in many NLP tasks, the fine-tuning stage can be still demanding inlabeled data. Recent works have resorted to active fine-tuning to improve thelabel efficiency of PLM fine-tuning, but none of them investigate the potentialof unlabeled data. We propose {\ours}, a new framework that leverages unlabeleddata to improve the label efficiency of active PLM fine-tuning. AcTune switchesbetween data annotation and model self-training based on uncertainty: itselects high-uncertainty unlabeled samples for active annotation andlow-uncertainty ones for model self-training. Under this framework, we design(1) a region-aware sampling strategy that reduces redundancy when activelyquerying for annotations and (2) a momentum-based memory bank that dynamicallyaggregates the model's pseudo labels to suppress label noise in self-training.Experiments on 6 text classification datasets show that AcTune outperforms thestrongest active learning and self-training baselines and improves the labelefficiency of PLM fine-tuning by 56.2\% on average. Our implementation will beavailable at \url{https://github.com/yueyu1030/actune}. | [
"Active Learning.Active learning has been widely applied to various NLP tasks including text classification {{cite:1d5e837672da1d79e0fa3583ebe8df16b6f6b0fd}}, sequence labeling {{cite:5e26fee9b68b6e83493de75007fe3de3ad18114f}}, {{cite:dca7f9a17da88153ffa3064b9bfa609e149bcafe}}, and question answering {{cite:5175ebde52ad4265a4fcbe6147ff1477b5816b36}} and machine translation {{cite:a038285ae8c4473aefd66120f8d3b50204c7b2ec}} with promising performance. So far, AL methods can be categorized into uncertainty-based methods {{cite:26d4a868570245b4de43f8930da34d19f7694d9d}}, {{cite:9e9ae3dafe01e6b5131b8fa668ecebcf9684b583}}, {{cite:39650260a4d4e08cb27ebe90715bfe56bec2cc4c}}, {{cite:a588ff026307e5a81b3de0ab9de1146baba7dc5f}}, diversity-based methods {{cite:bf9e9ca894b3ac2125ef258bac521c24b759725b}}, {{cite:e7c73d1d5ef0f1bde34d5909c906effedf79556a}} and hybrid methods {{cite:1d5e837672da1d79e0fa3583ebe8df16b6f6b0fd}}, {{cite:69ab1be79e7eff1eae247b52f8c7ecc289871028}}, {{cite:b42abaae92c7dc026d4ebf592c2ced7ee6d7acc1}}.{{cite:a4e0193f58a0a2111a02bdd036a80877679d3ac4}} offer a comprehensive empirical study of active learning with pre-trained LMs. In our study, we leverage the knowledge from unlabeled instances via self-training, which further promotes the performance of AL.",
"Semi-supervised Active Learning.There are several lines of work that combine active learning with semi-supervised learning.{{cite:449669e7906c33abbc82f1191d7cf19b712c8892}}, {{cite:a73365880be5d04fb7ce9090a28aba9dfd5ea8bf}} design query strategies for specific semi-supervised methods, {{cite:6f487384e21efa5c9a5dcb814b2d5cb48afed9d5}}, {{cite:e87b6ac42842fa0e07762ddd24637377d3607361}} combine active learning with data augmentation and{{cite:ac395f9e883989ec3096b1497a022099e46ed2a5}}, {{cite:447983ef8e119a9e761933f86b98ed3b482d5c30}}, {{cite:a9b399b2b8329a1c663cc950d5d69a478f216d6d}} exploit the most-certain samples from the unlabeled with pseudo-labeling to augment the training set.So far, most of the semi-supervised active learning approaches are designed for the computer vision domain and it remains unknown how this paradigm performs with pre-trained language models on NLP tasks. In our study, we propose Atm to effectively leverage unlabeled data during training. We also empirically demonstrate that our framework achieves the best performance in most cases compared with previous active semi-supervised learning approaches.",
"Self-training.Self-training is one of the earliest and simplest approaches to semi-supervised learning {{cite:2d2e8cc13549b21ad6992191692ad6e66de63e5d}}, {{cite:a8b9888fba158f328f82e04747cab78c66227454}}.The method first generates pseudo labels for high-confidence samples, then fits a new model on pseudo labeled data to improve the generalization ability.Self-training has been widely adopted in few-shot learning {{cite:1ad7b30b578b65f916801734c636a99b0745f065}}, {{cite:41cb6fb63a944ce3db9b4662c869ed9c5a26e614}}, {{cite:bf0025e253e19253a41f42407fdbe26984151563}}, {{cite:22814c3d5e086e0df15158c2d5a125a17e3f23a2}}, {{cite:5898c7937801944dc7cccc65c23abb201e38d01d}} as well as weakly-supervised learning {{cite:1d684034fa939c1f737cf626a960a7e3222bd094}}, {{cite:97f05beb355ef273e3f613e3bf7a07363d2fdd43}}, {{cite:11cef522885119df333b829b630ca8151f848e86}}, {{cite:6c1ff75545c083bf8eea03071401ef1ead1dfe06}}.The major drawback of self-training is that it is vulnerable to label noise, as the incorrect pseudo labels will cause confirmation bias and hurt model training {{cite:b7894a3ce8e2223285019656610ed224da8816d1}}, {{cite:c696544b24e7f0f0a644fde718f78dd293d9cea8}}.To enhance self-training for active learning, we adopt a simple momentum-based method to select high confidence samples, which is tailored for active learning and effectively reduce the noise in pseudo labels.We remark that although {{cite:1ad7b30b578b65f916801734c636a99b0745f065}}, {{cite:c696544b24e7f0f0a644fde718f78dd293d9cea8}} also leverage uncertainty information for self-training, their focus is on developing better self-training methods with a fix set of unlabeled data, while our focus is orthogonal to them as we aim to jointly query high-uncertainty unlabeled samples and generate pseudo-labels for low-uncertainty samples.The experiments in Section clearly show that with appropriate querying methods, Atm can further improve the performance of self-training."
] | [] |
An Unsupervised Way to Understand Artifact Generating Internal Units in
Generative Neural Networks | Despite significant improvements on the image generation performance ofGenerative Adversarial Networks (GANs), generations with low visual fidelitystill have been observed. As widely used metrics for GANs focus more on theoverall performance of the model, evaluation on the quality of individualgenerations or detection of defective generations is challenging. While recentstudies try to detect featuremap units that cause artifacts and evaluateindividual samples, these approaches require additional resources such asexternal networks or a number of training data to approximate the real datamanifold. In this work, we propose the concept of local activation, and devisea metric on the local activation to detect artifact generations withoutadditional supervision. We empirically verify that our approach can detect andcorrect artifact generations from GANs with various datasets. Finally, wediscuss a geometrical analysis to partially reveal the relation between theproposed concept and low visual fidelity. | [
"Deep Generative Neural Networks DGNNs are the models which approximate the input distribution given a target with neural networks. Representative architectures include variational autoencoder (VAE) {{cite:953a512ec3d6487eab6dea1e9f9ea7e3e690e104}}, neural language models {{cite:f4324d55d3c3b75105e772b4e8f21b8da88d3261}}, {{cite:d5b38d59559368006cf2d695dd1d85a5c53cf9bc}}, {{cite:96ffdf5c057e8febf40dd3674227e91fba7be56f}} and GANs. In particular, the adversarial training between a generator and a discriminator {{cite:a012851ac4394e7ebcd72a0241906d52c279927e}} has shown impressive performance in image generation {{cite:2028403fb27579d301d6780ff3993fda80bd639a}}, {{cite:6a9433b3599c38979d0b6989af6fc8bf8cbdb26c}}, {{cite:31067742637ac1b0811f29b0a8f4811532b5a36c}}, {{cite:8a90a5c54f3ccbf53ec2a5e69c5ac659b098259e}}.",
"Analysis for Interior Mechanism of GANs GAN Dissection {{cite:99806e32a813ef49f8a9a159975049484b85a341}} proposes a framework to investigate the generative role of each featuremap unit in GANs. It is shown that artifact generations can be improved by ablating units that are related to artifact generations. Another work {{cite:fe5156b747b38f2316124e233c04ae964172c628}} trains a linear classifier based on artifact-labeled data and removes artifacts by moving the latent code over the trained hyperplane. A sampling method with the trained generative boundaries was suggested to explain shared semantic information in the generator {{cite:3da844343d23c876b4f1a608541376cf316da986}}. Classifier-based defective internal featuremap unit identification was devised {{cite:530eb536427c66a11b6225037bfda786c1783b6b}}. The authors increase the visual fidelity by sequentially controlling the generation flow of the identified units.Analyses for latent space of the generator were also performed to manipulate the semantic of the generation {{cite:37daac52d08d8e4c657785540a3588e4a8991798}}, {{cite:4badf68d402440bf77e1eb83103367bb445c1a9d}}. Our work focuses more on the generation process and the relation between defective generation and the internal characteristics connected from the latent space.",
"Metric for Generative ModelVarious metrics have been proposed to evaluate the performance of generative models and each properties are well-summarized in {{cite:b19ad687e5e4ff0bf3fc7fe89342d8add03f0fa6}}. Although Fréchet Inception Distance (FID) {{cite:145b93fca82008c765559352eb00a3992dfc0ba2}} and Inception Score (IS) {{cite:6e4710d80597d98f85845fc02e6248abb32fa39f}} have shown robustness to image distortion, they sometimes assign high scores for generations with low visual fidelity. Precision and Recall (P&R) is a surrogate metric to quantify mode dropping and inventing based on training data {{cite:696436afb970671fa415c395efcf2a39106b9ff8}}, {{cite:a5730ea7bc729c258be1c81dbbe632da5e64e749}}. The authors also devised Realism Score (RS) to evaluate the visual fidelity of individual samples by comparing feature embeddings with training data. Perceptual path length (PPL) is another metric that quantifies the smoothness of the latent space with a hypothesis that the region in the latent space for defective generations has a small volume {{cite:6a9433b3599c38979d0b6989af6fc8bf8cbdb26c}}."
] | [] |
CONFIT: Toward Faithful Dialogue Summarization with
Linguistically-Informed Contrastive Fine-tuning | Factual inconsistencies in generated summaries severely limit the practicalapplications of abstractive dialogue summarization. Although significantprogress has been achieved by using pre-trained models, substantial amounts ofhallucinated content are found during the human evaluation. Pre-trained modelsare most commonly fine-tuned with cross-entropy loss for text summarization,which may not be an optimal strategy. In this work, we provide a typology offactual errors with annotation data to highlight the types of errors and moveaway from a binary understanding of factuality. We further propose a trainingstrategy that improves the factual consistency and overall quality of summariesvia a novel contrastive fine-tuning, called ConFiT. Based on ourlinguistically-informed typology of errors, we design different modularobjectives that each target a specific type. Specifically, we utilize hardnegative samples with errors to reduce the generation of factual inconsistency.In order to capture the key information between speakers, we also design adialogue-specific loss. Using human evaluation and automatic faithfulnessmetrics, we show that our model significantly reduces all kinds of factualerrors on the dialogue summarization, SAMSum corpus. Moreover, our model couldbe generalized to the meeting summarization, AMI corpus, and it producessignificantly higher scores than most of the baselines on both datasetsregarding word-overlap metrics. | [
"Multi-party dialogues are especially challenging to summarize using automated models, given that they often contain pauses, false starts, reconfirmations, hesitations, and speaker interruptions {{cite:3c890c904f1615cfde6f1897596a0d9528f6c089}}, {{cite:d853ef62b03560a43a54a66b707469be1cf1a2ec}}, {{cite:64ea7202f04334e3bf2dc0d8af3c249f59e2b602}}. Previous work in the field has addressed these challenges by incorporating semantic features, including keywords {{cite:c11d65c3997d9d2c510adbdb9880c06b105a1368}}, domain terminologies {{cite:3b1c1310988ce6a41cbecee9aac77d6afad40eba}}, topics {{cite:197ba4cd72741101a46a083e3a1b61240308b39a}}, {{cite:679395fbfeb6b6ba509b841e791522a7ee6e848f}}, entailment knowledge {{cite:d8acdbc1c9975cc7a646a5e701a3bedfbab6030c}}, and background knowledge {{cite:5e2195c4ad623f47bf51b0ed5ec325d2bef1a5d4}}. Other works exploit personal named entities {{cite:ae7a345ec63c062a056d1c0a445f643bcc027540}} and coreference information {{cite:7139cb1f5dbb78b263c34c95a8555151d046beac}} to learning to distinguish complex coreferent relationships expressed through personal pronouns (including the first person \"I\") in the conversation {{cite:edf5695a9ec18cc1efbe9be367139f06d399f71f}}.Researchers have also explored conversational structure {{cite:c97f0557cdb8f44edf8b1fc154bea4ce595ce343}}, utterance flow modelling {{cite:64ca2dc7d258ea181029f873083050706f12cace}}, syntactic structure {{cite:8db06629e0651a427e54b76b790f59129d943b0d}}, granularity control {{cite:b562fcbb6c85a980528f0a6302161757e76e5549}}, but they have not yet converged to a simple and practical solution.",
"Our proposed taxonomy of factual errors and annotations help us gain deeper insights into the causes of factual inconsistency in abstractive dialogue summarization outputs."
] | [] |
Asymptotically Optimal Golay-ZCZ Sequence Sets with Flexible Length | Zero correlation zone (ZCZ) sequences and Golay complementary sequences aretwo kinds of sequences with different preferable correlation properties.Golay-ZCZ sequences are special kinds of complementary sequences which alsopossess a large ZCZ and are good candidates for pilots in OFDM systems. KnownGolay-ZCZ sequences reported in the literature have a limitation in the lengthwhich is the form of a power of 2. In this paper, we propose two constructionsof Golay-ZCZ sequence sets with new parameters which generalize theconstructions of Gong et al. (IEEE Transaction on Communications 61(9), 2013)and Chen et al (IEEE Transaction on Communications 61(9), 2018). Notably, oneof the constructions results in optimal binary Golay-ZCZ sequences, while theother results in asymptotically optimal polyphase Golay-ZCZ sequences as thenumber of sequences increases. | [
"In this section we compare the proposed constructions to the previous works, specifically with the works of Gong et al. {{cite:9833e47adc70ed812169f0b3b953c9b1bbdbea2c}} and Chen et al. {{cite:fc980c0c6f18a0edce5d3f17a63d39e73504a72e}} and discuss the novelty of the proposed constructions.",
"In {{cite:9833e47adc70ed812169f0b3b953c9b1bbdbea2c}}, the authors only considered the complementary sequences of lengths of the form {{formula:e87e975a-f4c8-4980-9e02-a0f29ccc733c}} , and analysed their corresponding periodic zero auto-correlation zones. As compared to that, Theorem REF and Theorem REF considers complementary sequences of a more flexible form and analyses both periodic zero auto-correlation zone as well as zero crosscorrelation zone.In {{cite:fc980c0c6f18a0edce5d3f17a63d39e73504a72e}}, the authors proposed Golay-ZCZ sequence sets. However, the constituent sequences have lengths only of the form of {{formula:362fe371-6240-4aa2-bc5b-3f2685716f30}} . As compared to that, in Theorem REF we have proposed Golay-ZCZ sequence pairs of length {{formula:e727686a-8f5c-4fec-be32-7b0f5aeec4a2}} , where {{formula:78630baa-0f04-463f-88c5-b0339dd65bae}} is the length of a GCP. In Theorem REF , we have proposed Golay-ZCZ sequence sets, consisting sequences of lengths {{formula:6807ed6e-7ab8-4303-87ee-97f138018c2f}} , using an {{formula:32032b53-34ec-4ede-914c-cef84d103de1}} CCC. As we see the lengths are more flexible as compared to the results in {{cite:fc980c0c6f18a0edce5d3f17a63d39e73504a72e}}.The results reported in {{cite:fc980c0c6f18a0edce5d3f17a63d39e73504a72e}} only achieves optimality for the case of binary {{formula:ffc2528e-eb14-4b27-b5f6-c87a96cb36bc}} - Golay-ZCZ sequence pairs. As compared to that, based on the discussions in Section , we can see that Theorem REF results to binary {{formula:d0b85e3e-3bbf-4b39-a4db-af4ac31b1d5c}} - Golay-ZCZ sequence pair, which is optimal. Theorem REF results to polyphase {{formula:3b31a1d2-d1d0-41ad-9f8b-4e7cd946f65d}} Golay-ZCZ sequence set, which is also asymptotically optimal, as the number of sequence increases.Analysing Table 1 of {{cite:fc980c0c6f18a0edce5d3f17a63d39e73504a72e}}, we can observe that, as the number of sequences of the Golay-ZCZ sequence set increases, the ZCZ width decreases. As compared to that, in Theorem REF , as the number of sequences of the Golay-ZCZ sequence set increases, the ZCZ width also increases, and eventually we achieve asymptotic optimality.To increase the availability of CCCs of various parameters, which in turn increases the flexibility of the parameters of the proposed Golay-ZCZ sequence sets, we have proposed a new iterative construction of CCCs. We have also provided computer search results for binary CCCs with parameters {{formula:19196e5c-d7d1-496d-9a7f-07d53bcc2810}} , for various values of {{formula:d190f028-aecd-42bf-ba7f-fb04a3ea09f7}} given in Table REF . Moreover, using those as seed CCCs in {{cite:5c282f3e7b66602d326edd0405ab2f01d844c379}} and Theorem REF , we can construct several binary CCCs with new parameters, which are not reported before. As discussed in Remark REF , these CCCs are important in itself, since {{formula:962047ac-532b-4b1b-9cd5-5057aaa3f4b1}} is not a power-of-two."
] | [] |
Harnessing Cross-lingual Features to Improve Cognate Detection for
Low-resource Languages | Cognates are variants of the same lexical form across different languages;for example 'fonema' in Spanish and 'phoneme' in English are cognates, both ofwhich mean 'a unit of sound'. The task of automatic detection of cognates amongany two languages can help downstream NLP tasks such as Cross-lingualInformation Retrieval, Computational Phylogenetics, and Machine Translation. Inthis paper, we demonstrate the use of cross-lingual word embeddings fordetecting cognates among fourteen Indian Languages. Our approach introduces theuse of context from a knowledge graph to generate improved featurerepresentations for cognate detection. We, then, evaluate the impact of ourcognate detection mechanism on neural machine translation (NMT), as adownstream task. We evaluate our methods to detect cognates on a challengingdataset of twelve Indian languages, namely, Sanskrit, Hindi, Assamese, Oriya,Kannada, Gujarati, Tamil, Telugu, Punjabi, Bengali, Marathi, and Malayalam.Additionally, we create evaluation datasets for two more Indian languages,Konkani and Nepali. We observe an improvement of up to 18% points, in terms ofF-score, for cognate detection. Furthermore, we observe that cognates extractedusing our method help improve NMT quality by up to 2.76 BLEU. We also releaseour code, newly constructed datasets and cross-lingual models publicly. | [
"The two main existing approaches for the detection of cognates belong to the generative and discriminative paradigms. The first set of approaches is based on the computation of a similarity score between potential candidate pairs. This score can be based on orthographic similarity {{cite:f7c53d148899ae13adcaa294b54f93061891ecc4}}, {{cite:6e11c783154681971b54464c511cdc6be8bc561e}}, {{cite:45e44f78fadb9c469455159929561d5b772b54b0}}, phonetic similarity {{cite:6a7524d30325af2e9a9073e0c8a91c87646877c2}}, {{cite:62a78a59359d4e9a5274b197f4efe5a0456bc69b}}, {{cite:0c7f338e3abeacc94c47020512e57b699310b0fe}}, or a distance measure with the scores learned from an existing parallel set {{cite:47c971c57f2bc9429fc4f1e80c63d251e0d9ce34}}, {{cite:e4fab947a52c97b84d1cefb4bcc5b9f8a81fa3ad}}. The discriminative paradigm uses standard approaches to machine learning, which are based on (1) extracting features, e.g., character n-grams, and (2) learning to predict the transformations of the source word needed to {{cite:7f6d4f8fddaf19d7790bb2d97ddbdf055477d0e3}}, {{cite:cb31874c2e549cc9368d88711a53ae4a5403b55f}}.",
"Cognate Detection has been explored vastly in terms of classification methodologies. Previously, rama2016siamese employ a Siamese convolutional neural network to learn the phonetic features jointly with language relatedness for cognate identification, which was achieved through phoneme encodings. jager2017using use SVM for phonetic alignment and perform cognate detection for various language families. Various works on orthographic cognate detection usually take alignment of substrings within classifiers like SVM {{cite:883919abbcc5fb6fc8755e4b63ff25071eb49045}}, {{cite:09f1b082a5cc7601c74fe259dd293516162ffb77}} or HMM {{cite:b7af73ad62626e55ea418da6b0ac8929f0471a4f}}. ciobanu2014automatic employ dynamic programming based methods for sequence alignment. kanojia2019cognate perform cognate detection for some Indian languages, but a prominent part of their work includes manual verification and segratation of their output into cognates and non-cognates. kanojia2019utilizing utilize recurrent neural networks to harness the character sequence among cognates and non-cognates for Indian languages, but employ monolingual embeddings for the task. dijkstra2010cross show how cross-linguistic similarity of translation equivalents affects bilingual word recognition, even in tasks manually performed by humans. They discuss how the need for recognizing semantic similarity arises for non-identical cognates, based on the reaction time from human annotators. Similarly, merlo-andueza-rodriguez-2019-cross show that cross-lingual models exhibit the semantic properties of for bilingual lexicons despite their structural simplicities, which leads us to perform our investigation for low-resource Indian languages. uban-etal-2019-studying discuss the semantic change in languages by studying the change in cognate words across Romance languages using cross-lingual similarity. All of the previous approaches discussed above, lack the use of an appropriate cross-lingual similarity-based measure and do not work well for Indian languages as shown in this work. This paper discusses the quantitative and qualitative results using our approach and then, applies our output to different neural machine translation architectures.{{table:a720d96a-3453-4799-b43a-68d4d6cedb5d}}",
"{{table:38bee3e2-b124-4d4f-95e3-5ebbad2a5a4a}}"
] | [] |
Intelli-Paint: Towards Developing Human-like Painting Agents | The generation of well-designed artwork is often quite time-consuming andassumes a high degree of proficiency on part of the human painter. In order tofacilitate the human painting process, substantial research efforts have beenmade on teaching machines how to "paint like a human", and then using thetrained agent as a painting assistant tool for human users. However, currentresearch in this direction is often reliant on a progressive grid-baseddivision strategy wherein the agent divides the overall image into successivelyfiner grids, and then proceeds to paint each of them in parallel. Thisinevitably leads to artificial painting sequences which are not easilyintelligible to human users. To address this, we propose a novel paintingapproach which learns to generate output canvases while exhibiting a morehuman-like painting style. The proposed painting pipeline Intelli-Paintconsists of 1) a progressive layering strategy which allows the agent to firstpaint a natural background scene representation before adding in each of theforeground objects in a progressive fashion. 2) We also introduce a novelsequential brushstroke guidance strategy which helps the painting agent toshift its attention between different image regions in a semantic-aware manner.3) Finally, we propose a brushstroke regularization strategy which allows for~60-80% reduction in the total number of required brushstrokes without anyperceivable differences in the quality of the generated canvases. Through bothquantitative and qualitative results, we show that the resulting agents notonly show enhanced efficiency in output canvas generation but also exhibit amore natural-looking painting style which would better assist human usersexpress their ideas through digital artwork. | [
"The problem of teaching machines “how to paint” has been extensively studied in the context of stroke-based rendering (SBR), which focuses on the recreation of non-photorealistic imagery through appropriate positioning and selection of discrete elements such as paint strokes or stipples {{cite:aa53ec33e9b60290f39e8121504a5a578d62c162}}, {{cite:12f1934080866652f88fbd1ec73cac2ba92e7eac}}. Traditional SBR algorithms often require greedy search {{cite:28d69d5ab27b29b3a6fbde47e13cfe1bf9736b9d}}, {{cite:0dc3af91b7a4f8bb5180564a5afd35d9910dcbbd}}, optimization over devised heuristics {{cite:ce0aea5b3173dc17e3da01084dad7dc9d078fd42}}, or user input for selecting the position and nature of each brushstroke {{cite:60c98e9b3cd708707fce11e4dff5d9db29020c66}}, {{cite:cf8b9ff62532b021df247d5bcff7fdc3a2f0dabe}}. More recent solutions {{cite:bfdef7d71d264e034837e2799e6432a00aae0087}}, {{cite:0a07f2cbd18306503500542a42f7eed2cf885b57}} adopt the use of recurrent neural networks for computing optimal brushstroke decomposition.However, these methods require access to dense human brushstroke annotations, which limits their applicability to most real world problems. In another work, Zhao {{cite:ef24cc32e81c6d404608d9f6330ba3f4fab4c1ec}} use a conditional variational autoencoder framework for synthesising time-lapse videos depicting the recreation of a given target image. However, this requires access to painting time-lapse videos from real artists for training. Furthermore, the time-lapse outputs are generated at very low-resolution as compared to the high-resolution output sequences generated using our approach.",
"In recent years, there has been an increased focus on learning an unsupervised brushstroke decomposition without requiring access to dense human brushstroke annotations. For instance, recent works {{cite:ec2e8fda5977a5bbff5346962c61156024ecc8cf}}, {{cite:a3d0d03d0e6da8400dadbfe1d491836b382a159b}}, {{cite:1a4b7825e193133739ebaed4681b141393d4b233}}, {{cite:1741f65037c7542252851524e58811cc441cd39f}}, {{cite:2eb9098994227b4773caf1da08150ff66ec7d8cf}}, {{cite:657852b56d28d3222678e555a0b3d630654bdc1c}} use deep reinforcement learning and an adversarial training approach for learning an efficient brushstroke decomposition. Optimization-based methods {{cite:cd912ec55481733e41eccc8726043318b83ef856}} directly search for the optimal brushstroke parameters by performing gradient descent over a novel optimal-transport-based loss function. In another recent work, Liu {{cite:98ff9b57dd25a145fd9ce3a928399be8b256be0f}} propose a novel Paint Transformer which formulates the learning to paint problem as a feed-forward stroke set prediction problem.",
"While the above works show high proficiency in painting quality output canvases, the generation of the same invariably depends on a progressive grid-based division strategy. In this setting, the agent divides the overall image into successively finer grids, and then proceeds to paint each of them in parallel. Experimental analysis reveals that this not only reduces the efficiency of the final agent, but also leads to mechanical (grid-based) painting sequences which are not easily intelligible to human users."
] | [] |
GOSH: Task Scheduling Using Deep Surrogate Models in Fog Computing
Environments | Recently, intelligent scheduling approaches using surrogate models have beenproposed to efficiently allocate volatile tasks in heterogeneous fogenvironments. Advances like deterministic surrogate models, deep neuralnetworks (DNN) and gradient-based optimization allow low energy consumption andresponse times to be reached. However, deterministic surrogate models, whichestimate objective values for optimization, do not consider the uncertaintiesin the distribution of the Quality of Service (QoS) objective function that canlead to high Service Level Agreement (SLA) violation rates. Moreover, thebrittle nature of DNN training and prevent such models from reaching minimalenergy or response times. To overcome these difficulties, we present a novelscheduler: GOSH i.e. Gradient Based Optimization using Second Order derivativesand Heteroscedastic Deep Surrogate Models. GOSH uses a second-order gradientbased optimization approach to obtain better QoS and reduce the number ofiterations to converge to a scheduling decision, subsequently lowering thescheduling time. Instead of a vanilla DNN, GOSH uses a Natural ParameterNetwork to approximate objective scores. Further, a Lower Confidence Boundoptimization approach allows GOSH to find an optimal trade-off between greedyminimization of the mean latency and uncertainty reduction by employingerror-based exploration. Thus, GOSH and its co-simulation based extensionGOSH*, can adapt quickly and reach better objective scores than baselinemethods. We show that GOSH* reaches better objective scores than GOSH, but itis suitable only for high resource availability settings, whereas GOSH is aptfor limited resource settings. Real system experiments for both GOSH and GOSH*show significant improvements against the state-of-the-art in terms of energyconsumption, response time and SLA violations by up to 18, 27 and 82 percent,respectively. | [
"We list in Table REF the prior work, dividing such methods into four classes: Reinforcement learning (rows 1 and 2 of Table REF ), MaxWeight (row 3), stochastic modelling (rows 4 and 5) and gradient-based methods (row 6).",
"Reinforcement Learning models: Many recent scheduling approaches use various forms of reinforcement learning to adapt in hybrid edge-cloud environments. Predominantly, these approaches use policy-gradient methods, which use a neural network to model a policy and output a decision for each state (assuming Markovian dynamics). In a dynamic scheduling setup, at each scheduling interval, a reward signal that quantifies the QoS of that interval is used to update the parameters of the model. Examples include asynchronous-actor-critic (A3C) policy gradient approach {{cite:1e968eae03f8ca517fcb4812c7b4803baed9e97f}}. Such methods are shown to be robust and able to adapt in diverse scenarios {{cite:36cb444610537ddd23463e209f8528a3c61fd593}}, {{cite:1e968eae03f8ca517fcb4812c7b4803baed9e97f}}, {{cite:18d715283425956409ecfbabddd89d5a0985414f}}. However, such methods can be slow to adapt and fail to model volatile scenarios/workloads to be able to efficiently schedule in complex environments {{cite:71b9495e832b7402e6209a5b945045c5668d5fc8}}. Furthermore, they are known to perform poorly in heterogeneous setups or when the QoS is adaptive {{cite:c95ec548bb76b1a4488d9b4167494c7a57c39296}}.",
"MaxWeight based schedulers: Many prior scheduling approaches use the popular MaxWeight based techniques with robust theoretical guarantees. Such strategies aim to maximize the \"weight\" of job schedules where this weight is calculated using expected QoS metrics {{cite:05a52344ecf38ad415e4820ca562bc5e0f34a0e2}}, {{cite:0726c651edb7d3b8c5b0384a8fb8ba922484d8c9}}, {{cite:d59ac80331e3c335d0fb2c21fc2659c1a69b6d88}}. To compare with our method, we use the POND approach as a baseline due to its high performance compared to other methods in this class {{cite:05a52344ecf38ad415e4820ca562bc5e0f34a0e2}}. POND uses a pessimistic-optimistic Upper Confidence Bound (UCB) based optimization in maximization setup. UCB optimization of the weight of virtual queues for constrained online dispatch minimizes the violations and maximizes the expected average reward. However, such techniques are unable to leverage the edge-cloud diversity in energy consumption and latency to improve SLA {{cite:0c0db5ffe2ebdabde985a3f0523c53e3f040ca2d}}.",
"Stochastic Modelling: Prior work also uses stochastic surrogate models such as Heteroscedastic Gaussian Processes to optimize metrics like QoS by appropriate task scheduling {{cite:e1650c9cc6538acff9f28caa5a43cd4ce315c762}}, {{cite:67ab1d41a981c7cca1f157db52b54a7485ecdfbf}}, {{cite:40877b37f8aad2fcb5d82d4286d0503700ef0f4f}}, {{cite:0c0db5ffe2ebdabde985a3f0523c53e3f040ca2d}}. Other works use mean and variance estimates based on historical data to perform safe optimization {{cite:0c0db5ffe2ebdabde985a3f0523c53e3f040ca2d}}, {{cite:5fe75739c254a6345f8b4ab432a38a03baf4af14}}. These methods typically utilize the uncertainty information for robust or safe optimization {{cite:c5bc4909f7d96ea54f773da2a5102b9de98d0421}} or use error-based exploration {{cite:67ab1d41a981c7cca1f157db52b54a7485ecdfbf}}. Typically, due to poor modeling accuracy of Gaussian Processes, they are unable to perform well in complex environments like heterogeneous fog environments and hence more sophisticated models like NPNs are used in this work {{cite:f6805e7eb496e07c576ae1e74398a303dc6c6fae}}.",
"Gradient-based Optimization: A recently proposed class of methods (GOBI/GOBI*) use a deep neural network based surrogate model to approximate the QoS parameters like energy and response time {{cite:71b9495e832b7402e6209a5b945045c5668d5fc8}}. An extension of the GOBI scheduler for long-term QoS estimation is the MCDS approach {{cite:1f3ca402e08b406b9ad631e0a540bf6f1e7cb9da}} that utilizes Monte-Carlo simulations to estimate the reward at a future system state. Their predictions are aided by a co-simulated digital-twin of the target hosts which receive the tasks. These algorithms use gradient-based optimization approaches like AdamW {{cite:c61986191cdecf359e212e0f18aad894ba09fa0b}} with cosine annealing and restarts {{cite:364c955b81ee469d18d3dc50c530a9a89669afa8}}, {{cite:bbd978a3e63d0acfc1ffeda2e414afc43af9ea81}} to converge to an appropriate scheduling decision. Leveraging gradient-based directed search on neural approximator with coupled-simulation, GOBI* is able to adapt to non-stationary workloads. The neural approximator used in these methods is a Fully Connected Network (FCN) with input including the host and task utilization metrics of Instructions per second (IPS), RAM, Disk and Bandwidth with the scheduling decisions for each active task. GOBI*'s neural approximator also includes the objective score after the single-step execution of GOBI's scheduling decision on a simulated platform. The output of these neural approximators is an objective score that is minimized.",
"Although GOBI* provides distinct advantages compared to the state-of-the-art, it also faces drawbacks. First, the model require a neural approximator to be pre-trained on several hours of workload traces generated using a random allocation based scheduler. This might be infeasible for application scenarios where swift deployment is required. Second, the deployment environment might be very different from the setup used to generate the training data for the model. Even though the model is able to adapt to completely different setups, experiments show that GOBI* lacks the agility required for QoS efficient scheduling (Section REF ). Finally, even with techniques like momentum, our experiments show that the gradient-based optimization can take a large number of iterations to converge due to many saddle points in the QoS hyper-surface. Considering these shortcomings, there is a need for a model that is able to leverage state-space exploration to quickly adapt to setups different from the one used for training. To this end, we present the GOSH approach as part of this work."
] | [] |
Characterizing and addressing the issue of oversmoothing in neural
autoregressive sequence modeling | Neural autoregressive sequence models smear the probability among manypossible sequences including degenerate ones, such as empty or repetitivesequences. In this work, we tackle one specific case where the model assigns ahigh probability to unreasonably short sequences. We define the oversmoothingrate to quantify this issue. After confirming the high degree of oversmoothingin neural machine translation, we propose to explicitly minimize theoversmoothing rate during training. We conduct a set of experiments to studythe effect of the proposed regularization on both model distribution anddecoding performance. We use a neural machine translation task as the testbedand consider three different datasets of varying size. Our experiments revealthree major findings. First, we can control the oversmoothing rate of the modelby tuning the strength of the regularization. Second, by enhancing theoversmoothing loss contribution, the probability and the rank of <eos> tokendecrease heavily at positions where it is not supposed to be. Third, theproposed regularization impacts the outcome of beam search especially when alarge beam is used. The degradation of translation quality (measured in BLEU)with a large beam significantly lessens with lower oversmoothing rate, but thedegradation compared to smaller beam sizes remains to exist. From theseobservations, we conclude that the high degree of oversmoothing is the mainreason behind the degenerate case of overly probable short sequences in aneural autoregressive model. | [
"The issue of generating sequences that are shorter than the ground-truth one has been studied from various aspects including model parametrization, data collection, and decoding. Here we highlight some of these projects in the context of our work.",
"On the aspect of model parametrization,{{cite:a02611865a9d387a69aeec707a278c801fff74fd}} suggest using sparse transformation of the next-token distribution rather than the usual way of using softmax. Such a model is then able to assign zero probability to short sequences more readily and thereby reduce the oversmoothing rate. Their approach, however, does not explicitly encourage {{formula:8aaf404b-c590-487b-b664-8b1742135ca4}} tokens to be assigned zero probability, unlike ours where {{formula:c26d79e8-fe41-4999-9e0d-9683d633b309}} is treated specially.{{cite:34bb12879e4529df3aae1a5356926139d0523e09}} embed the {{formula:542c1a28-f2b8-4054-8aa7-ea12e1bd46d5}} token with a distinct vector at each position within the sequence. This was shown to help the probability of empty sequence, although they do not report its impact on translation quality at all.",
"On data collection, {{cite:090a0731f7d004cd403088c3a75d29881b8431a1}} analyze data collection and show that data augmentation techniques altering sequence length may address the issue of oversmoothing and improve translation quality. Their work is however limited to low-resource tasks. With respect to decoding, {{cite:c1d3861117d79fac3e946f0cf9883243f39b67af}} design a decoding algorithm that learns to correct the underestimated length. Alternative decoding algorithms, such as minimum Bayes risk decoding {{cite:c716885721dd5903348815f9ffad4f47a0436b26}}, {{cite:ef21791a31ddc58f7e38d061b6a0043bf466ee1c}}, have been shown to alleviate the length mismatch to a certain extent when compared to beam search.",
"These earlier approaches do not attempt at formally characterizing the cause behind the issue of oversmoothing. This is unlike our work, where we start by formalizing the issue of oversmoothing and propose a way to alleviate this issue by directly addressing this cause."
] | [] |
Degendering Resumes for Fair Algorithmic Resume Screening | We investigate whether it is feasible to remove gendered information fromresumes to mitigate potential bias in algorithmic resume screening. Using acorpus of 709k resumes from IT firms, we first train a series of models toclassify the self-reported gender of the applicant, thereby measuring theextent and nature of gendered information encoded in resumes. We then conduct aseries of gender obfuscation experiments, where we iteratively remove genderedinformation from resumes. Finally, we train a resume screening algorithm andinvestigate the trade-off between gender obfuscation and screening algorithmperformance. Results show: (1) There is a significant amount of genderedinformation in resumes. (2) Lexicon-based gender obfuscation method (i.e.removing tokens that are predictive of gender) can reduce the amount ofgendered information to a large extent. However, after a certain point, theperformance of the resume screening algorithm starts suffering. (3)General-purpose gender debiasing methods for NLP models such as removing gendersubspace from embeddings are not effective in obfuscating gender. | [
"Digitization of hiring has spawned an emerging line of research at the intersection of hiring discrimination and algorithmic bias.This research investigates how algorithmic hiring tools (e.g. resume screening) can be biased, and how that can lead to discrimination in hiring {{cite:b66b36a0c556ccdbd4fcedfe1d0e15c9d2d00b5a}}, {{cite:1ee23d529a8313c3bc23cddba858cd0741143314}}, {{cite:34baa0218881a373c7c667a507ff9d407ed0a687}}, {{cite:4a9bd110be1ce4b5121e95c8ed50ee6bf752104d}}.As more sophisticated NLP models get deployed in such tools, a cause for concern is that the gender representations learned in these models can propagate bias in downstream tasks.For example, research has shown that gendered wordings exist in job postings {{cite:0ec7d5b2b3a5385be629bf16c08c3b15a3fa84ca}}, {{cite:605b3bb197888816133210c6c9f615af60fba447}}.If a resume screening tool matches resumes to job descriptions using embeddings of the documents, male resumes may be more likely to be matched to job descriptions with masculine language.One approach to address this issue is to de-bias the NLP models to minimize potential bias in downstream tasks {{cite:3a53dc021e7762ea159f89bd194afa3ce87311dd}}.These general-purpose techniques, though not specific to resumes, include obfuscating/swapping gender {{cite:f4143dcc2dd60d4e24c165478f885eeb0b73216b}}, removing gender subspace in embedding models {{cite:1f2a70ac5b87412f58cdaa36a697778bc1a518df}} (See also {{cite:29591634c2d9bd9dc569c9b25dbb518f194d767d}} for the shortcomings of this approach), and training gender-neutral embeddings {{cite:38a186c255b1c850520f1a4fb937f11881087611}}.",
"Another cause for concern is that language models can learn gendered information from resumes, propagating any bias in the training data downstream like in the Amazon example.These concerns, while theoretically plausible, are largely empirically unfounded within the resume context.The extent to which models can use and learn gendered information from resumes depends on the amount of gendered information and the nature of gendered information in resumes, including gender differences in language use."
] | [] |
Slot-VPS: Object-centric Representation Learning for Video Panoptic
Segmentation | Video Panoptic Segmentation (VPS) aims at assigning a class label to eachpixel, uniquely segmenting and identifying all object instances consistentlyacross all frames. Classic solutions usually decompose the VPS task intoseveral sub-tasks and utilize multiple surrogates (e.g. boxes and masks,centres and offsets) to represent objects. However, this divide-and-conquerstrategy requires complex post-processing in both spatial and temporal domainsand is vulnerable to failures from surrogate tasks. In this paper, inspired byobject-centric learning which learns compact and robust object representations,we present Slot-VPS, the first end-to-end framework for this task. We encodeall panoptic entities in a video, including both foreground instances andbackground semantics, with a unified representation called panoptic slots. Thecoherent spatio-temporal object's information is retrieved and encoded into thepanoptic slots by the proposed Video Panoptic Retriever, enabling it tolocalize, segment, differentiate, and associate objects in a unified manner.Finally, the output panoptic slots can be directly converted into the class,mask, and object ID of panoptic objects in the video. We conduct extensiveablation studies and demonstrate the effectiveness of our approach on twobenchmark datasets, Cityscapes-VPS (\textit{val} and test sets) and VIPER(\textit{val} set), achieving new state-of-the-art performance of 63.7, 63.3and 56.2 VPQ, respectively. | [
"Panoptic Segmentation (PS).Unifying semantic and instance segmentation on the image level,the PStask {{cite:28b3f106450784c28cf33d96f836cb14ee5ba232}}, {{cite:fd5af7667a2b1961a504f14bd341288dc9d10c78}}, {{cite:dfbc86e79d5738904c44a568c65402b349973169}}, {{cite:eb852291793bdd5307174afcd4edbe95c47d8301}}, {{cite:83e7efe9a143f282fe09e8883c5736decb6da21e}}, {{cite:9ac9f3cf7f7ec8bb7138b7d37e1270d5502f17f9}} requires assigning a class label for all pixels and uniquely segmenting all object instances.Early attempts {{cite:28b3f106450784c28cf33d96f836cb14ee5ba232}} in PS follow the decomposition pipeline,separately predicting the semantic and instance segmentation results and then adopting the things-stuff fusion process in later stages.Several workstry to simplify the process and improve the accuracy throughreplacing the post things-stuff fusion with the parameter-free {{cite:dfbc86e79d5738904c44a568c65402b349973169}} or trainable {{cite:6001f6e9ad424769396d56a035c8bdf715a0e696}} panoptic head.Furthermore, more researchers {{cite:2412492fe5fa4fb0ef5d988ed247aa5e28f92902}}, {{cite:d71f9021b8456d239fa4cdbe328fb5d8feafd54d}} try to abandon the separated branches andbuild an end-to-end unified framework.A recent trend in PS task {{cite:d6be5246c717c5dcd1e78420b3cb8b94ec74c494}}, {{cite:ab8fd299da7a50949236f990701d44e51bec96ca}}, {{cite:7148270a7e22218c9f324555850ad49ea17cf7f0}} is regarding this task as the set prediction problem and try to construct the concise end-to-end networks with the help of transformers.However, note that all these unified methods only consider the unification of things and stuff on the spatial domain, excluding the temporal domain.",
"Video Panoptic Segmentation (VPS).As a direct extension of the PS task into video domain, previous approaches in the VPS task{{cite:ecd6873686c8f318eaae41216e83c2e8cda9a8ce}}, {{cite:78107bcd1aeea5eaded517ab35409b9dbd0fda0d}} usually apply the image-level approach {{cite:dfbc86e79d5738904c44a568c65402b349973169}}, {{cite:83e7efe9a143f282fe09e8883c5736decb6da21e}} to each video frame and try to associate the result of all frames by using an extra temporal association head (track head {{cite:ecd6873686c8f318eaae41216e83c2e8cda9a8ce}}, {{cite:7ea4c275eb6b057f3be5b582df7be5512ccf6a04}}, temporal center regression head {{cite:78107bcd1aeea5eaded517ab35409b9dbd0fda0d}}), .However, all these methods represent objects with multiple representations (boxes and masks, centers and offsets), requiring multiple separate networks to handle sub-tasks of VPS and complex post-processing (NMS, things-stuff fusion, similarity fusion for tracking)on both spatial and temporal domains.To the best of our knowledge, this is the first time that the unification of things and stuff in both spatial and temporal domains are discussed simultaneously for this task.",
"Transformers.Inspired by the success of Transformers {{cite:26baabf31da8c761e7ca2fa26d2993197e32cfd5}} in Natural Language Processing (NLP) tasks,numerous researches have also been investigated in the computer vision community,Object Detection {{cite:d6be5246c717c5dcd1e78420b3cb8b94ec74c494}}, {{cite:ccc7b8b6ee73078de767e12328f37594e6b094ca}},Panoptic Segmentation {{cite:ab8fd299da7a50949236f990701d44e51bec96ca}}, {{cite:7148270a7e22218c9f324555850ad49ea17cf7f0}}, andVideo Instance Segmentation {{cite:0d3c8107812f14693c7e29b32b7cc1d7224e18a6}}, {{cite:9f9b473eb79be160a8512d6bf5024d1f2eee0481}}, {{cite:04f1a6c509c8777418b732bcc626ad0add0d5fe3}}, .In these pipelines, objects are represented as a set of vectors containing all related information such as location, appearance, Attention in transformers is responsible for localizing and segmenting objects, and the bipartite-matching based mechanisms help separate different objects.With the help of these technologies,above tasks can be converted into direct set prediction problem without many hand-designed components.Different from all these works, we introduce the slots competing mechanism into the learning process to enhance the discriminability of objects in both spatial and temporal domains.Jointly representing stuff and things on the video level with panoptic slots, we propose the fully unifiedend-to-endframework in which all operations in both spatial and temporal domains are fulfilled based on the panoptic slots.",
"Object-centric learning.Object-centric representation learning {{cite:6fdaefa80a048ed322dd954c1f82e99c9d3b2bee}}, {{cite:027d0aad4dd58ff68e9aa7dfde8cc1f6f99dee92}}, {{cite:94cfe95a9c01de34ea4abf599c5b0608cd7d5064}}, {{cite:d83055d9555566c3c3d8bfc08fb9ce9d898089c8}}, {{cite:b75a42625774a12216424b0bc384864cb587abf8}} mainly focuses on learning robust, generalizable object representations in various scenarios such as unsupervised object discovery, novel-viewpoint prediction, In {{cite:94cfe95a9c01de34ea4abf599c5b0608cd7d5064}}, Francesco propose the Slot Attention to predict a set of task-dependent abstract representations called slots from the perceptual representations such as the output of a convolutional neural network.Slots are exchangeable and can bind to any object in the input.Different from the attention in transformer, the Slot Attention lets the randomly sampled slots compete with each other in the iterative learning process, which will further help the discrimination of objects.However, the Slot Attention is mainlyapplied to synthetic dataset scenarios,assuming a normal distribution for the slots.In this paper, motivated by the object-centric representations and the competing mechanism, we introduce the panoptic slots and the VPR module to successfullybring object-centric representation learning to real-world data, and set up new state-of-the-art for the VPS task ontwo datasets."
] | [] |
Addressing Adversarial Machine Learning Attacks in Smart Healthcare
Perspectives | Smart healthcare systems are gaining popularity with the rapid development ofintelligent sensors, the Internet of Things (IoT) applications and services,and wireless communications. However, at the same time, several vulnerabilitiesand adversarial attacks make it challenging for a safe and secure smarthealthcare system from a security point of view. Machine learning has been usedwidely to develop suitable models to predict and mitigate attacks. Still, theattacks could trick the machine learning models and misclassify outputsgenerated by the model. As a result, it leads to incorrect decisions, forexample, false disease detection and wrong treatment plans for patients. Inthis paper, we address the type of adversarial attacks and their impact onsmart healthcare systems. We propose a model to examine how adversarial attacksimpact machine learning classifiers. To test the model, we use a medical imagedataset. Our model can classify medical images with high accuracy. We thenattacked the model with a Fast Gradient Sign Method attack (FGSM) to cause themodel to predict the images and misclassify them inaccurately. Using transferlearning, we train a VGG-19 model with the medical dataset and later implementthe FGSM to the Convolutional Neural Network (CNN) to examine the significantimpact it causes on the performance and accuracy of the machine learning model.Our results demonstrate that the adversarial attack misclassifies the images,causing the model's accuracy rate to drop from 88% to 11%. | [
"Several studies discuss adversarial attacks and their impact on machine learning models. In Table REF , we summarise the existing works and show their comparison with our work. For instance, proposal {{cite:33a283a9d794ef62a30594d9b30aa242093d1dfc}} provides an analysis of adversarial attacks by utilising five adversarial algorithms that are used to examine certain attack models. However, the dataset used in {{cite:33a283a9d794ef62a30594d9b30aa242093d1dfc}} is generated from device data. Moreover, unlike our proposal, they do not use a CNN to test the model's accuracy.",
"The authors in {{cite:bd6f5c2740af1d69b33731fbb06e5c0c4167ff6a}} explore the consequences of adversarial attacks and show how adversarial training can support therobustness of supervised models. They use confusion matrices andclassification performance of J48 and Random Forest modelstrained on adversarial samples generated. Notably, {{cite:bd6f5c2740af1d69b33731fbb06e5c0c4167ff6a}} addresses the unique challenges of implementing adversarial attacks in the cybersecurity domain. Although the paper discusses the FGSM attack, unlike our proposal, it does not use CNN.",
"Proposal {{cite:d8878ae7eff548be9a1120fa5761dde90ba415a9}} presents an overview of different adversarial attacks, including FGSM. However, unlike our motivation, they use the MNIST and CIFAR10 datasets, not a custom dataset. Proposal {{cite:a975f232d0020d5417f13e81709568c0d9d5f300}} discusses theoretical foundations, algorithms, and applications of adversarial attack techniques along with defence techniques to combat them and protect systems. Unlike our proposal, it does not consider an implementation detail. Proposals {{cite:202532019bd839b597cc0d315a8d7876f8261196}} and {{cite:8cb90f81138f13118a0b61d67ee1581da9a959a7}} discuss the issues of machine learning models misclassifying, the FGSM attack, and adversarial training. But once gain, unlike our approach, they do not consider CNN.",
"In {{cite:0e2fb5af2155b9d855358e4b13d4dc5e6b5876fb}}, authors discuss adversarial machine learning from a smart healthcare perspective. However, unlike ours, the proposal does not consider FGSM and CNN, nor any detailed design is stated. While proposals {{cite:6e029828968b9dfe9aa9baabb209136f036a47c4}} and {{cite:780103bb98c8ba5f0da110043ad8ea1ccf60d639}} present some technical details on implementing adversarial attacks on machine learning, unlike ours, they do not use a custom dataset. Further, they do not focus on the smart healthcare perspective.",
"In summary, we note that the above-mentioned proposals donot discuss the implementation of a CNN. While a few proposals discuss the importance of CNN in adversarial machine learning attacks, the discussion does not lead to an actual execution that we provide in our paper. That said, we focus on implementing an FGSM attack on CNN. We also note there is a trend to use an existing dataset, e.g., MNIST and CIFAR-10. In our paper, we avoid such datasets to show the actual implementation of the model using a medical health dataset."
] | [] |
Hybrid Classical-Quantum Autoencoder for Anomaly Detection | We propose a Hybrid classical-quantum Autoencoder (HAE) model, which is asynergy of a classical autoencoder (AE) and a parametrized quantum circuit(PQC) that is inserted into its bottleneck. The PQC augments the latent space,on which a standard outlier detection method is applied to search for anomalousdata points within a classical dataset. Using this model and applying it toboth standard benchmarking datasets, and a specific use-case dataset whichrelates to predictive maintenance of gas power plants, we show that theaddition of the PQC leads to a performance enhancement in terms of precision,recall, and F1 score. Furthermore, we probe different PQC Ans\"atze and analysewhich PQC features make them effective for this task. | [
"Pang et al. {{cite:d58c4b5af61ea4d39efa44614b90f9deefe8b3bc}} provide an extensive surveyof current classical methods for flagging anomalies using deep learning methodologies. One popular approach makes use of Autoencoders (AEs) {{cite:5835f8d95f92efd501bcbcb16e5644910b320dcd}}, {{cite:611ad9fcc792833b04219bb6e4f10094f6bdd620}} or its more powerful variant - Variational Autoencoders (VAEs) {{cite:696a46ba02485bd015ddebfa46e78125059b0c9e}}, {{cite:78f427af145bcbdd59105bbcfd22af34ec2ac903}}, {{cite:0bed6cb3d3c2220cc8a28633d7ff34c9624c8cfa}}. This type of models serve as a foundation to this project.",
"Initial work on quantum autoencoders focused on utilizing their bottleneck architecture for compression of quantum data {{cite:a2a39a954c8ad8faefe84859b00589d2f20a328d}}. The authors showed an application of this idea in the context of quantum simulations. However, in the present work, we would like to concentrate on classical data being processed by a quantum algorithm.",
"Herr et al. {{cite:8b565c9dc33a104a64af0f4de7d10621763ba779}} applied quantum generative models for anomaly detection on classical data. In their pioneering work, the authors utilized a network architecture similar to AnoGAN {{cite:06b4b2e7ca4aee1762706506d576e15b17005032}} and introduced a hybrid classical-quantum structure for the generator network. However, in their work, classical data is only passed into the discriminator, therefore a pre-processing layer for classical data in the generator is not needed. Even though this approach did not show any advantage in performance compared to its classical counterpart on the chosen dataset, it laid an important groundwork for further research.",
"One of the first studies to investigate the correlation between quantum model classification accuracy and expressivity descriptor as proposed Sim et al. {{cite:8dfea0cc57e3434983ea723dfeccefdacb4f1bd1}} was conducted by Hubregtsen et al. {{cite:8ba6c408322f686d133fc49bc4538bdfb28307cd}}. The experiment was performed on purely quantum models and on artificial classical data, which showed strong correlation between classification accuracy and expressivity measure. In our work, we build up on this study and test these descriptors in a hybrid classical-quantum scenario on a real-world dataset {{cite:a174933922e8fe0628ba1e8e440c491d507a3341}}, {{cite:e5596cd4e2f139b0f5435c51e73b4d89ced13b6c}}, {{cite:5532dfe3cba6e5ef51b6fc46ebac855316f99363}}."
] | [] |
Joint Analog and Digital Transceiver Design for Wideband Full Duplex
MIMO Systems | In this paper, we propose a wideband Full Duplex (FD) Multiple-InputMultiple-Output (MIMO) communication system comprising of an FD MIMO nodesimultaneously communicating with two multi-antenna UpLink (UL) and DownLink(DL) nodes utilizing the same time and frequency resources. To suppress thestrong Self-Interference (SI) signal due to simultaneous transmission andreception in FD MIMO systems, we propose a joint design of Analog and Digital(A/D) cancellation as well as transmit and receive beamforming capitalizing onbaseband Orthogonal Frequency-Division Multiplexing (OFDM) signal modeling.Considering practical transmitter impairments, we present a multi-tap widebandanalog canceller architecture whose number of taps does not scale with thenumber of transceiver antennas and multipath SI components. We also propose anovel adaptive digital cancellation based on truncated singular valuedecomposition that reduces the residual SI signal estimation parameters. Tomaximize the FD sum rate, a joint optimization framework is presented for A/Dcancellation and digital beamforming. Finally, our extensive waveformsimulation results demonstrate that the proposed wideband FD MIMO designexhibits higher SI cancellation capability with reduced complexity compared toexisting cancellation techniques, resulting in improved achievable rateperformance. | [
"For narrowband or frequency-flat FD MIMO systems, full-tap analog cancellers connecting all TX outputs to RX inputs are usually employed, where the number of taps increases with the number of TX/RX RF chains {{cite:c7ac5317ecc1263d9fa57142e60aac8c39824eb8}}, {{cite:fc0e8a246447b1341ae1a2c7a171455ebbdb496e}}. To reduce this hardware complexity, analog SI canceller designs exploiting AUXiliary (AUX) TX structures and/or joint design of TX/RX beamformers were studied in {{cite:70842b91b37b1ae54e9ae815517c88feb58645e1}}, {{cite:b5062149096284798c2204d985f521f77b20ac6a}}, {{cite:a2b291d6113e7fd8cabad4fe91b19fa740b4f106}}, where the analog cancellation signals were injected into each of the RX inputs using separate TX RF chains. Although these techniques reduce the analog canceller hardware complexity, they are unable to suppress the nonlinear SI components due to the non-ideal RF front-end hardware rendering the RX chains into saturation {{cite:e798ae91fd85682a5ffd9a96a31c9089cb907195}}. In our previous work {{cite:607d41403602b9182809106f4c03db57371d1dbc}}, we presented a unified low complexity Analog and Digital (A/D) cancellation for narrowband FD MIMO systems.For wideband FD MIMO systems in {{cite:f1377991e33f62fd42caaf41eff56820bad8e98f}}, {{cite:d21f839422362386355d81687fb1954d160039c7}}, full-tap analog cancellers with adaptive filters were utilized to provide sufficient analog SI cancellation. In {{cite:406a10f942ab2078659e9c6a0b458a0eabdf35fc}}, a full-tap wideband FD MIMO RF canceller was presented with a tunable time delay circuit, which employed reflected type phase shifters to emulate the true time delays of the SI channel. Recently, a full-tap beam-based RF cancellation approach was introduced in {{cite:74cd0ebb2bd3267e8e3938da519d4ad26810ee17}}, which employed analog Least Mean-Squared (LMS) loops as the adaptive filters for SI mitigation in FD massive MIMO systems. Those analog LMS loops include time delay generators, down-converters, Low-Pass Filters (LPFs), and up-converters. An integrated LMS adaptive wideband FD MIMO RF canceller was proposed in {{cite:05303479c93574dbecfcf138af0f627e22935a62}}, where the time delay of the cancellation was generated using an {{formula:bc9d14ce-3937-45df-bdab-e1dc8b16052d}} -path filter.However, the hardware complexity of all the above full-tap RF cancellers scales with the number of TX and RX RF chains as well as the number of SI multipath components, rendering the practical implementation of the analog SI cancellation unit a core design bottleneck.",
"Alleviating the need for analog SI cancellation, spatial suppression techniques were presented in {{cite:c7ac5317ecc1263d9fa57142e60aac8c39824eb8}}, {{cite:06ce9c356dc9c2c24ed6bc392414d62bcacd4074}} for narrowband FD MIMO systems, where the SI suppression was solely handled by the digital TX/RX beamformers. However, those spatial suppression techniques were unable to cancel the SI in high TX power and often resulted in reductions of the data rates for both the UL and DL signals of interest. This stemmed from the fact that some of the available spatial DoFs were devoted to mitigating SI {{cite:009aca8c1124188f0ca973f33e132f8194b8324b}}. To avoid such issues, digital cancellation techniques exploiting SI signal modeling were utilized in practice to supplement the analog canceller in suppressing the SI signal. To achieve sufficient SI suppression, existing digital cancellation approaches capitalize on models for the PA impairments {{cite:d21f839422362386355d81687fb1954d160039c7}} and IQ mixer image effect {{cite:ec811e60642ff64eec5a1432d11fff48c9b6a5a5}}, or rely on cascaded SI designs taking into account both nonidealities {{cite:a8af50678333ec5f3d5ac25c0db47b7f3822205c}}. However, the number of estimation parameters of those models grows with the number of TX/RX RF chains and SI channel components.To reduce the number of parameters for the FD MIMO system, a digital canceller based on Principle Component Analysis (PCA) was provided in {{cite:1482bfdf9ad36c05c6e66355c9994979775c148a}}. Furthermore, in {{cite:7c5481c7a11b8682d658ed7fa8969d6ff91298b3}}, {{cite:505755f3c22b8c5f1352f3a8e4579291e5b04add}}, {{cite:3a5462a02bc4d96243d5e39766e70e609b60b3b2}}, the authors considered FD MIMO OFDM signal modeling to design rate maximizing TX/RX beamformers. However, these techniques assumed full-tap RF cancellers to achieve certain SI suppression levels."
] | [] |
An Interactive Approach for Identifying Structure Definitions | Our ability to grasp and understand complex phenomena is essentially based onrecognizing structures and relating these to each other. For example, anymeteorological description of a weather condition and explanation of itsevolution recurs to meteorological structures, such as convection andcirculation structures, cloud fields and rain fronts. All of these arespatiotemporal structures, defined by time-dependent patterns in the underlyingfields. Typically, such a structure is defined by a verbal description thatcorresponds to the more or less uniform, often somewhat vague mental images ofthe experts. However, a precise, formal definition of the structures or, more generally,concepts is often desirable, e.g., to enable automated data analysis or thedevelopment of phenomenological models. Here, we present a systematic approachand an interactive tool to obtain formal definitions of spatiotemporalstructures. The tool enables experts to evaluate and compare differentstructure definitions on the basis of data sets with time-dependent fields thatcontain the respective structure. Since structure definitions are typicallyparameterized, an essential part is to identify parameter ranges that lead todesired structures in all time steps. In addition, it is important to allow aquantitative assessment of the resulting structures simultaneously. Wedemonstrate the use of the tool by applying it to two meteorological examples:finding structure definitions for vortex cores and center lines of temporarilyevolving tropical cyclones. Ideally, structure definitions should be objective and applicable to as manydata sets as possible. However, finding such definitions, e.g., for the commonatmospheric structures in meteorology, can only be a long-term goal. Theproposed procedure, together with the presented tool, is just a first approachto facilitate this long and arduous way. | [
"Concepts are the building blocks of thoughts and are necessary for the representation of knowledge as well as for mental processes like categorization, inference, decision making, learning and communication.This view has become widely accepted in cognitive science and philosophy of science, see e.g., {{cite:6c8ba0834a5041ab5c8ca8b715610d6e1335dae2}}, {{cite:c535c942f3e11d6419a8f3d7d924084a35e80814}}.Here we perceive concepts as mental constructs (as opposed to abstract objects or skills) and we use a naive notion of concepts based on concrete examples, thus avoiding complex philosophical questions (see, e.g., {{cite:9cb0c45147ade30b0236103043f4d6798a8fda4b}} for an introduction on this).",
"The extraction of structures from data has a long history in data visualization.The structures are often mathematical features.Most prominent examples are mathematical features of spatial or spatiotemporal fields, like extremal structures, level sets, critical points and skeletons of stable manifolds.Therefore, the term “feature extraction” has become established in visualization.When dealing with application-oriented structures that carry semantics, such as jet streams, precipitation fields or tropical storms, it is better to speak of “structures” or “concepts”. These are characterized by certain properties, which may include also mathematical features.This terminology corresponds to that of many other sciences, but differs from that traditionally used in visualization, where application-oriented structures are also often called “features”.",
"For a survey on the substantial history of structure extraction and visualization, see reviews on geometric-topological objects {{cite:8e3a9093dbc25091485995b3e8262cf8bfee5066}}, {{cite:c639ca09b17fb5f707444825305d02f4517e4e24}}, on coherent set detection {{cite:806e6affef7fe83fd3c05e05e1ddb9107a016f1d}}, on feature tracking in meteorological contexts {{cite:ff4e0b5be3c2614be1e1d117fada72f7d7b8db74}}, {{cite:bc33b3bae29b84acc27b6b1750e487bef1d9c860}}, and on the related problem of image segmentation {{cite:c5b85c7f02a0fdc54267cfa629cad3ea8768d882}}.Most approaches in data visualization refer to structures defined by single scalar or vector fields, e.g., level-sets or ridges of Lyaponov exponents or, more application-related, the centerlines of vortices or jet streams {{cite:105919448aac1051c52013fbccd352baa121955d}}.",
"Data visualization often focuses on isolating application-defined regions of interest in spatiotemporal data. In a multidimensional “attribute space”, defined by the value ranges of given and derived data, one searches interactively for subregions containing the value combinations of interest.Related procedures have been developed for image segmentation and volume rendering, where transfer functions map data characteristics to optical qualities, like color and opacity, to highlight regions of interest {{cite:e4b3f8ab8fbc7e536d1bf0059839674f359db144}}, {{cite:4faea34add55c392693f2390739e7b18c1055c91}}.",
"An essential prerequisite for finding suitable structure definitions with visual control is interactivity. Here we can build on the research thread that began with the work by {{cite:c6ae00c6a8a13d13d832abd17706ded5e0c98d8c}}.In this, the definition of subregions in multivariate range space is greatly facilitated by multiple coordinated views (histograms, scatter plots, parallel coordinates, or function graphs), representing different variables side by side. When data points are interactively selected (“brushed”) in one view, the associated data items are immediately highlighted in all linked views.This “linking” is particularly powerful when attribute views are combined with 3D views, highlighting the corresponding preimages in the domain.This technique can also be extended to time-dependent data {{cite:ad6fb2b604def6f24c1c595e4868b1ce675bf3e4}}, {{cite:94542a365969d17fc36bc860e874d2ac722a403f}} and has been applied to simulation data of Hurricane Isabel {{cite:20942d041a54da7b3699dfe14e763fa7a4c5dfff}}.While in this work all structures are defined by ranges of multivariate quantities, it is not concerned with finding more general and also parameter-dependent definitions of structures.",
"A tool for interactive exploration of sets of parameter-dependent 3D geometries has been presented in {{cite:7a7c2a23cb4e09b5b7c429eae07d74adc080bad7}}. However, this work also did not aim at obtaining structural definitions.To the best of the authors' knowledge, there is no tool to date that supports the explicit goal of obtaining mathematical definitions of mentally and/or verbally given structures by using visual data analysis techniques.",
"Visual comparison of similar, but different spatial data has been more deeply explored in the context of ensemble visualization, see, e.g., {{cite:a2895ad19044db70eb8e3fed376be922f5ff7d19}} and references therein.",
"The difficulty of finding a generally accepted, precise definition of concepts that is as objective as possible (i.e., independent of the person observing) is best illustrated by the example of flow vortices. Data visualization, together with flow research, has contributed a great deal to this. There is not enough space here, to sketch the nearly half-century-long research thread; instead we refer to the current state of research {{cite:e5164f8e14288bc8d0061a843b4a360f2380a32e}} and the papers cited therein.In recent work {{cite:ffb335a4797310c9aa889ecd488f56e600d1d9f5}}, a formal approach to the definition and persistence of meteorological structures has been taken. This is exactly the kind of research we want to support and facilitate with the proposed procedure and interactive tool."
] | [] |
Interference Suppression Using Deep Learning: Current Approaches and
Open Challenges | In light of the finite nature of the wireless spectrum and the increasingdemand for spectrum use arising from recent technological breakthroughs inwireless communication, the problem of interference continues to persist.Despite recent advancements in resolving interference issues, interferencestill presents a difficult challenge to effective usage of the spectrum. Thisis partly due to the rise in the use of license-free and managed shared bandsfor Wi-Fi, long term evolution (LTE) unlicensed (LTE-U), LTE licensed assistedaccess (LAA), 5G NR, and other opportunistic spectrum access solutions. As aresult of this, the need for efficient spectrum usage schemes that are robustagainst interference has never been more important. In the past, most solutionsto interference have addressed the problem by using avoidance techniques aswell as non-AI mitigation approaches (for example, adaptive filters). The keydownside to non-AI techniques is the need for domain expertise in theextraction or exploitation of signal features such as cyclostationarity,bandwidth and modulation of the interfering signals. More recently, researchershave successfully explored AI/ML enabled physical (PHY) layer techniques,especially deep learning which reduces or compensates for the interferingsignal instead of simply avoiding it. The underlying idea of ML basedapproaches is to learn the interference or the interference characteristicsfrom the data, thereby sidelining the need for domain expertise in suppressingthe interference. In this paper, we review a wide range of techniques that haveused deep learning to suppress interference. We provide comparison andguidelines for many different types of deep learning techniques in interferencesuppression. In addition, we highlight challenges and potential future researchdirections for the successful adoption of deep learning in interferencesuppression. | [
"To the best of our knowledge, the current literature lacks a comprehensive study on deep learning-based interference suppression. As at the time of writing this paper, we have found only two surveys which consider DL techniques in the context of interference suppression. In {{cite:b183b99f633a09ea132dda3882e365450146871e}}, the authors discussed deep learning techniques for radar applications in general and briefly highlighted interference techniques that fell into the deep learning categories. One key distinction of our work from theirs is that the interference types they considered includes only jamming and clutter. We note that this investigation is very important and our work goes a step further by investigating more DL-based interference suppression applications including radars, optical communications, IoT networks, GPS navigation, full duplex systems and wireless sensor networks. Further, we consider many other types of interference such as self-interference, adjacent channel interference, homogeneous and heterogeneous technology interference. Furthermore, the jamming suppression techniques considered in {{cite:b183b99f633a09ea132dda3882e365450146871e}} focused on target signal classification and recognition. Our work is fundamentally different from theirs because we focused more on interference suppression techniques and not interference classification techniques. In another work, Chen et al. {{cite:dd858087ebe07e318206f3300fe2eb066800ca6a}} surveyed both coexistence and interference mitigation techniques, including traditional (non-DL) and DL approaches for wireless personal area network (WPAN) and WLAN applications. While this is also another important contribution, the scope of the application considered is different in comparison to the one presented in this survey. Also, we found that their work focused more heavily on non-DL interference suppression and briefly discussed DL-based suppression techniques. Furthermore, in our own work, we discuss several critical interference characteristics and why DL techniques are suitable to suppress them. Also, we point the community to publicly available dataset for interference suppression.",
"Many of the other surveys {{cite:c8ba572824b5e79886c830ed08467c104d5d87a5}}, {{cite:8a421b48575b3231cd122bed48fc033b675595b6}}, {{cite:bc445c70c0895815afa7a24fb234454984bbfd5a}}, {{cite:bfe916f5b1e7606b29c2424f54dec4f36abfe63e}}, {{cite:ef1da107c039c371dc2653594ad08aa9ad18cb37}} on interference suppression reviewed papers that utilized signal processing techniques and even non-ML techniques to avoid interference. For instance, Azmi et. al. {{cite:c8ba572824b5e79886c830ed08467c104d5d87a5}} discussed technical papers that used techniques such as transmit power control, dynamic frequency selection, interferer classification schemes {{cite:8a421b48575b3231cd122bed48fc033b675595b6}}, time slotted channel hopping schemes amongst others to suppress interference in wireless sensor networks (WSN). Further, Lu et al. reviewed interference mitigation schemes that utilized adaptive array processing in GPS systems. Le et al. also surveyed interference mitigation schemes used in wireless body area sensor networks. Kunert et. al. {{cite:bc445c70c0895815afa7a24fb234454984bbfd5a}} surveyed six basic techniques that researchers have used to suppress radar interference. They sorted the suppression techniques into domains such as polarization, time, frequency, coding and space. While the techniques discussed in this survey paper typically require domain expertise to be successful, it is not clear if they will be scalable in applications such as IoT and 5G where there are multiple devices sharing the same frequency channel. Furthermore, one of the approaches discussed in the paper is \"communicate and avoid\" in various domains including time and frequency. While this is a straightforward approach, it involves multiple steps to achieve and the time to implement in a latency-sensitive applications such as 5G New Radio (NR) or vehicle to everything (V2X) may be unrealistic. Some of these limitations can be tackled using deep learning techniques surveyed in this paper.",
"For automotive radar suppression algorithms, Toth et. al. {{cite:68a0fa2b106ab02b9b9fe6d9c2362e7dbb7cc149}} investigated state-of-the-art interference suppression techniques including zeroing, ramp filtering, IMAT and time domain parametric interpolation (IVMip). Their study did not include ML or DL approaches that could be leveraged to suppress interference. Also, their application focused primarily on automotive radar interference, whereas our survey encompasses many other application areas. In the same vein, Nabil et. al. {{cite:216bdf77a4507cb9d402285b9d365ebe613119a9}}, discussed advances in self and mutual interference suppression techniques for wireless networks. They surveyed redundancy techniques such as MIMO technology to suppress mutual interference in scenarios when the receiver does not have complete knowledge of the interfering signal. They also discussed interference cancellation techniques such as Balun transformer, isolation of the transmitter and receiver paths, phase shifting and spatial techniques (enabling an appropriate distance between the transmit antennas).",
"Our work is different from all the other surveys on interference suppression because we explored suppression approaches that utilized deep learning for many application areas and interference characteristic types while also providing a comparison between them and the major challenges with using deep learning to suppress interference. To the best of our knowledge, this is the first paper that explores deep learning based interference suppression techniques for multiple application areas and various interference characteristics. There is a lot of promising results that show that DL based techniques can be very useful in quickly suppressing interference in many wireless communications applications.{{table:eb6ab9da-e4af-4f61-8269-7123527c4397}}"
] | [] |
Decentralized Mean Field Games | Multiagent reinforcement learning algorithms have not been widely adopted inlarge scale environments with many agents as they often scale poorly with thenumber of agents. Using mean field theory to aggregate agents has been proposedas a solution to this problem. However, almost all previous methods in thisarea make a strong assumption of a centralized system where all the agents inthe environment learn the same policy and are effectively indistinguishablefrom each other. In this paper, we relax this assumption aboutindistinguishable agents and propose a new mean field system known asDecentralized Mean Field Games, where each agent can be quite different fromothers. All agents learn independent policies in a decentralized fashion, basedon their local observations. We define a theoretical solution concept for thissystem and provide a fixed point guarantee for a Q-learning based algorithm inthis system. A practical consequence of our approach is that we can address a`chicken-and-egg' problem in empirical mean field reinforcement learningalgorithms. Further, we provide Q-learning and actor-critic algorithms that usethe decentralized mean field learning approach and give stronger performancescompared to common baselines in this area. In our setting, agents do not needto be clones of each other and learn in a fully decentralized fashion. Hence,for the first time, we show the application of mean field learning methods infully competitive environments, large-scale continuous action spaceenvironments, and other environments with heterogeneous agents. Importantly, wealso apply the mean field method in a ride-sharing problem using a real-worlddataset. We propose a decentralized solution to this problem, which is morepractical than existing centralized training methods. | [
"MFGs were first proposed in {{cite:afc8fdbb7c98abfb7c9a70984c1f07af4cd2d6e4}}, while a comprehensive development of the system and principled application methods were given later in {{cite:880734ceec21ae48a982ef3ed44c042f0b66d362}}. Subsequently, learning algorithms were proposed for this framework. {{cite:cb98a0bb4eb56870de34c1c2803204e056100221}} introduce a restrictive form of MFG (known as stationary MFG) and provide a model-free policy-gradient {{cite:5bb44954a4b2087ca55b331f7c57b2070e7dfba6}}, {{cite:9eccae20f8774ea2bced24a3f1bae4757ed98157}} algorithm along with convergence guarantees to a local Nash equilibrium. On similar lines, {{cite:17279446c46f04e1c9d361177f8f96fd0ab7b676}} provide a model-free {{formula:2ee850e1-a902-4382-bf12-a9a3d1250f36}} -learning algorithm {{cite:a4d8b1e6afe84ad45e1e824bbcb6d664f0f59a8c}} for solving MFGs, also in the stationary setting. The assumptions in these works are difficult to verify in real-world environments. Particularly, {{cite:17279446c46f04e1c9d361177f8f96fd0ab7b676}} assume the presence of a game engine (simulator) that accurately provides mean field information to all agents at each time step, which is not practical in many environments.Further, other works depend on fictitious play updates for the mean field parameters {{cite:8918de9a0d61ca65e4f00937d02ccca48742586e}}, {{cite:e4b7a4a0d0f5abd1b3701e174fd758a895db428c}},which involves the strong assumption that opponents play stationary strategies. All these papers use the centralized setting for the theoryand the experiments.",
"Prior works {{cite:70452ee0a1b999f408c5f9c6adea45e237d1812d}}, {{cite:7af19474a0d1ab60c8eaff9ab82ff244d0bca5cc}}, {{cite:147841f4d665394681ef49e795ae6c8be39e02dd}} have established the existence of a (centralized) mean field equilibrium in the discrete-time MFG under a discounted cost criterion, in finite and infinite-horizon settings. Authors have also studied the behaviour of iterative algorithms and provided theoretical analysis for learning of the non-stationary (centralized) mean field equilibrium in infinite-horizon settings {{cite:c4a254a9d43c615f6291a08649e32adf2e81bbc5}}, {{cite:6ce17a2444e9e25152d845a3c0c194155860c50a}}, {{cite:9d73ef51ccf706eaf0ffefc7db6cb5fd46198623}}. We provide similar guarantees in the decentralized setting with possibly heterogeneous agents.",
"{{cite:c22be1a12647e2e68692fc14a2693004d3253865}} introduces MFRL that uses a mean field approximation through the empirical mean action, and provides two practical algorithms that show good performance in large MARL settings. The approach is model-free and the algorithms do not need strong assumptions regarding the nature of the environment. However, they assume access to global information that needs a centralized setting for both theory and experiments.MFRL has been extended to multiple types {{cite:db2591aeaeaaaa197811744e252182cba4c75f93}} and partially observable environments {{cite:092837d7b945450f71cdd95d70f91f98bc247602}}. However, unlike us, {{cite:db2591aeaeaaaa197811744e252182cba4c75f93}} assumes that agents can be divided into a finite set of types, where agents within a type are homogeneous.{{cite:092837d7b945450f71cdd95d70f91f98bc247602}} relaxes the assumption of global information in MFRL, however, it still uses the Nash equilibrium as the solution concept.Further, that work contains some assumptions regarding the existence of conjugate priors, which is hard to verify in real-world environments. Additional related work is provided in Appendix .",
"Mean field games have been used in the inverse reinforcement learning paradigm, where the objective is to determine a suitable reward function given expert demonstrations {{cite:33adf30e94aafd7caab55811b86aa94c9562ff7f}}. Model-based learning solutions have also been proposed for mean field games {{cite:953c2750c6a529427730e0129f53784d4a65541e}}, {{cite:258acb190c97c916e0eeca8eab91a4a50273bc99}}, though these methods suffer from restrictive assumptions on the model and the analysis does not generalize to other models. Methods such as {{cite:2b553d0729d7de981b924b63690148cc18ed6588}} perform a very computationally expensive computation using the full knowledge of the environment, which does not scale to large real-world environments. The work by {{cite:aed090dad546a8ee893b30e6f6474e233cc28626}} provides a mean field actor-critic algorithm along with theoretical analysis in the linear function approximation setting, unlike other works which only analyze the tabular setting {{cite:17279446c46f04e1c9d361177f8f96fd0ab7b676}}, {{cite:c22be1a12647e2e68692fc14a2693004d3253865}}. However, it is not clear if this algorithm is strong empirically since it does not contain empirical experiments that illustrate its performance. Further, this work considered the linear-quadratic setting that contains restrictions on the type of reward function. {{cite:1253eaf45717c3943e56931fbf9a83f8a654aa00}} explore the connection between MARL and mean field games in the model-free setting.",
"The mean field setting under the cooperative case can be handled using a mean field control model {{cite:6df218243b34f7d413d70705ede37af4159b1a18}}. In the linear-quadratic setting, {{cite:6df218243b34f7d413d70705ede37af4159b1a18}} prove that policy-gradient methods converge to local equilibrium. In further work, the same authors prove that {{formula:07072351-3f1a-4bad-a3e2-9281e727625f}} -learning algorithms also converge {{cite:29f2e2633093aeb0e4cb1bda5f76959c6dd68439}}.",
"In the experiments, we consider a real-world application of our methods on the Ride-Pool Matching Problem (RMP) as originally defined in {{cite:0a974e68b218f09a164cc3632d8da80f95022fc0}}. This is the problem considered by top ride-sharing platforms such as UberPool and Lyft-Line. The problem studies the efficiency of accepting ride requests by individual vehicles in such a way that the vehicles make more money (cater to more requests) per trip and the ride-sharing platform improves its ability to serve more orders. Previous approaches to solving the RMP problem have used standard optimization techniques {{cite:7f19af1935376fe2d918e407098488e530d01ef5}}. However, these methods are not scalable to large environments. One example from this class of methods is the zone path construction approach (ZAC) {{cite:1ed68be77e139fe42c3ddb9b921b7e73c34731d6}}. The ZAC solves for an optimization objective where the environment is abstracted into zones and each zone path represents a set of trips. The available vehicles are assigned to suitable zone paths using the ZAC algorithm. Another proposed solution was to make greedy assignments {{cite:0a974e68b218f09a164cc3632d8da80f95022fc0}}, which considers maximizing the immediate returns and not the long term discounted gains traditionally studied in RL. Prior work has also considered RL approaches for this problem {{cite:1a238495bf17148aea9d656dc3fd8e398f464f23}}, {{cite:1d1dc8fe45052c5892f7c60ec330872acb6dee52}}. However, these works consider very restrictive settings, which do not model multiagent interactions between the different vehicles. The work by {{cite:992805ca40f23d09880dd28a6161f72d25ab54b3}} used a mean field approach for order dispatching, however, it assumes vehicles serve only one order at a time. The recent work by {{cite:b48453bcb51c84ce94091ac21198eafd4a629571}} introduced a very general formulation for the RMP where vehicles of arbitrary capacity are designed to serve batches of requests, with the whole solution scalable to thousands of vehicles and requests. They introduced a Deep {{formula:a8df983b-1554-4dbe-be92-490c0df5be57}} -Network (DQN) {{cite:2027f8354bfaae1c6e4ff3265d2c449689756978}} based RL approach (called neural approximate dynamic programming or NeurADP) that learns efficient assignment of ride requests. However, the proposed approach assumes a set of identical vehicles and uses centralized training, where all vehicles learn the same policy using centralized updates. This is not practical in real-world environments, where the individual vehicles are typically heterogeneous (many differences in vehicular capacity and preferences). We propose a mean field based decentralized solution to this problem, which is more practical than the approach by {{cite:b48453bcb51c84ce94091ac21198eafd4a629571}}."
] | [] |
Hierarchical Clustering: $O(1)$-Approximation for Well-Clustered Graphs | Hierarchical clustering studies a recursive partition of a data set intoclusters of successively smaller size, and is a fundamental problem in dataanalysis. In this work we study the cost function for hierarchical clusteringintroduced by Dasgupta, and present two polynomial-time approximationalgorithms: Our first result is an $O(1)$-approximation algorithm for graphs ofhigh conductance. Our simple construction bypasses complicated recursiveroutines of finding sparse cuts known in the literature. Our second and mainresult is an $O(1)$-approximation algorithm for a wide family of graphs thatexhibit a well-defined structure of clusters. This result generalises theprevious state-of-the-art, which holds only for graphs generated fromstochastic models. The significance of our work is demonstrated by theempirical analysis on both synthetic and real-world data sets, on which ourpresented algorithm outperforms the previously proposed algorithm for graphswith a well-defined cluster structure. | [
"Our work fits in a line of research initiated by Dasgupta {{cite:3cdb42f895f43f93a62231e898c9b4d7f3ae8480}}, who introduced a cost function to measure the quality of an HC tree.Dasgupta proved that a recursive application of the algorithm for the Sparsest Cut problemcan be used to construct an {{formula:7abeea80-8877-45d2-afc2-8bd4ef516ccd}} -approximate HC tree. The approximation factor was first improved to {{formula:770a5580-2856-48eb-8652-4be54f715d6c}} byRoy and Pokutta {{cite:0058143d9733143f62a2d8f9270afd4a2f417d40}}.Charikar and Chatziafratis {{cite:0056bb69043bd2f78c6cfb016cefbeba5940e154}} improved Dasgupta's analysis of the recursive sparsest cut algorithm by establishing the following black-box connection: an {{formula:8c03a801-44e9-454f-9bab-edbf0942fddb}} -approximate algorithm for the Sparsest Cut problem can be used to construct an {{formula:b53eb444-05a4-429d-95b2-006a75ae59d6}} -approximate HC tree according to Dasgupta's cost function. Hence, an {{formula:34fdde5f-a8ad-4217-a006-5f220071a320}} -approximate HC tree can be computed in polynomial time by using the celebrated result of {{cite:09606854139a7d28aba6e85c1865b05eb0edfe17}}.For general input instances, it is known to be NP-hard to find an optimal HC tree {{cite:3cdb42f895f43f93a62231e898c9b4d7f3ae8480}}, and SSEH-hard to achieve an {{formula:7a1f0d5b-cd29-4e7f-a82e-86452d67e100}} -approximation with respect to Dasgupta's cost function {{cite:0058143d9733143f62a2d8f9270afd4a2f417d40}}, {{cite:0056bb69043bd2f78c6cfb016cefbeba5940e154}}.",
"Cohen-Addad et al. {{cite:e3665822316d78bcdc5334ca8b14d943400f3550}} analysed the performance of several linkage heuristics (e.g., Average Linkage) for constructing HC trees, and showed that such algorithms could output an HC tree of high cost in the worst case.Moving beyond the worst-case scenario, Cohen-Addad et al. {{cite:0ce18570f1a318acaa2cab7df12a2d182cf76222}} studied a hierarchical extension of the SBM and showed that, for graphs generated according to this model, a certain SVD projection algorithm {{cite:a5b3741aad3e627fa4f367a60415d9c74ece8dc0}} together with several linkage heuristics can be applied to construct a {{formula:cb96503b-0322-4ce0-9802-8aa5fa796ccb}} -approximate HC tree with high probability. We emphasise that our notion of well-clustered graphs generalises the SBM variant studied in {{cite:0ce18570f1a318acaa2cab7df12a2d182cf76222}}, and does not assume the rigid hierarchical structure of the clusters.",
"For another line of related work, Moseley and Wang {{cite:38c0fa9399efd56d43176a51808418b8348db169}} studied the dual objective function and proved that Average Linkage achieves a {{formula:55b01b6f-38f1-463c-ad68-ee89d52cbcee}} -approximation for the new objective. Notice that, although this has received significant attention very recently {{cite:4463fc723bdeacba0305a24035f4334d0ee3a663}}, {{cite:951c6b136f6463ca6b270d0dffd32a1338f28241}}, {{cite:caef1dbd4b910bf4eaa70fa4d359ab21aa7942e4}}, {{cite:851f02ad5f4032856a1959ea43be97dd185d34f5}}, {{cite:47f755c43ac9ed2ca6fe46a7b9f568e9303114c3}}, achieving an {{formula:6092035f-b715-4775-9434-1a7930b2e468}} -approximation is tractable under this alternative objective. This suggests the fundamental difference on the hardness of the problem under different objective functions, and is our reason to entirely focus on Dasgupta's cost function in this work."
] | [] |
On Optimizing Interventions in Shared Autonomy | Shared autonomy refers to approaches for enabling an autonomous agent tocollaborate with a human with the aim of improving human performance. However,besides improving performance, it may often also be beneficial that the agentconcurrently accounts for preserving the user's experience or satisfaction ofcollaboration. In order to address this additional goal, we examine approachesfor improving the user experience by constraining the number of interventionsby the autonomous agent. We propose two model-free reinforcement learningmethods that can account for both hard and soft constraints on the number ofinterventions. We show that not only does our method outperform the existingbaseline, but also eliminates the need to manually tune a black-boxhyperparameter for controlling the level of assistance. We also provide anin-depth analysis of intervention scenarios in order to further illuminatesystem understanding. | [
"Shared Autonomy.",
"In shared autonomy, the control of a system is shared by human and agent to accomplish a common goal. Earlier works {{cite:4cfa6b84088917663744bc298ea095916a3634b0}}, {{cite:cf8a4032abc3d64d1e9a9d487b3122852f7a5c67}}, {{cite:ac0a1713370e8ab85e387e5bca13dc25f61f2ef0}} assumed that the agent is aware of the common goal. However, this is a strong assumption in new and previously unseen applications, and as a result, more recent works have proposed agents that can infer the user's intent and predict the user's goal from the user's action and environmental dynamics. They do so by formulating this problem as a partially observable Markov decision process (POMDP). Bayesian inference {{cite:db703e712b0c261874157ea26821c9ec6c06a01c}}, {{cite:880c77722b1b6982471e4900fcfe141d9ac1031b}}, {{cite:f0fb9b14749492bd6fbef51cec0077ced6930094}}, {{cite:47e209f14557bb7c20f9edcaa9774cfca958bb8d}}, inverse reinforcement learning {{cite:b39929d04dfcc87891131328087ce6c02e850435}}, {{cite:3d368594644959e6244e6d2000059bb118d6caf4}}, {{cite:7cad9894c860460b7613d4b3b9809bf0e3d75b56}} and hindsight optimization {{cite:923e952599ea1610250792d69175beacaff6595e}}are the most common approaches to predicting the user's goal in this formulation. These approaches are based on an assumption that the environment dynamics, the goal space, and the user’s policy are known to the agent, which is often violated in practice.{{cite:e2dd3b7cd7b8bb19d9ce92388132552d15a5bb21}} proposed a model-free shared autonomy framework to relax these assumptions and formulate the problem as a Markov decision process (MDP). This method does not require prior knowledge of environment dynamics, the human's policy, the goal space, and goal representation. Instead, it uses human-in-the-loop reinforcement learning with functional approximation to formulate a policy for the agent learned only from environmental observation and user input. Based on the aforementioned work, {{cite:b27f37bc3906914cb7c11ce692d42500517b4aaa}} proposed Stochastic Q Bumpers. Rather than simply focusing on improving the current performance of a human, they consider the additional objective of helping a human learn, or optimizing future performance.",
"Intervention optimization. In the setting of user assistance, there are only a few works that emphasise the influence of the assistive agent on the user. An agent that intervenes too often will reduce a user's advocacy and ultimately affect their experience in the system.To tackle this challenge, {{cite:e2dd3b7cd7b8bb19d9ce92388132552d15a5bb21}} employed deep Q-learning to train the assistive agent. Instead of taking the action with the highest q-value, they first sort the actions in terms of similarity to the user provided action, in descending order, using an action-similarity function. A final action is then selected as the one that is closest to the user's action while ensuring its q-value is not significantly worse than the optimal action. This action selection addresses the intervention problem implicitly.",
"By tuning the tolerance of the system to suboptimal human suggestions, which is a black-box hyperparameter, this method can provide different levels of assistance. However, designing such an action-similarity function makes this approach an environment-dependent problem.{{cite:ae4811cc003110ef3075c9d1a08d2f24d576b130}} proposed to minimize the interventions in a similar way. They accept the user's action only when it is close enough to the optimal action as determined by an optimal controller. However, developing such an optimal controller may require complete knowledge of the environment dynamics and the user's goal. Similar to {{cite:ae4811cc003110ef3075c9d1a08d2f24d576b130}},{{cite:3f3146a0cd240511da8113eea58bdaf9a17533cf}} formulate the problem as a constrained MDP where they consider minimizing the amount of interventions an agent makes, subject to the returns of the agent's learned policy being greater than a given constant. In contrast, our work describes a method to maximize returns of the human-AI collaboration, subject to the number of interventions taken being less than a given constant. Additionally, their method only works with on-policy RL algorithms and it needs to set a target return as the constraint, which requires a deep understanding of the environment and user. {{cite:19a53f4af792e37ed3f5f1cbffaabffbcdc65273}} and {{cite:c404d9355075dd70c9cf91a38b4ab20cd6482bb5}} used imitation learning and supervised learning to predict action discrepancy or the novelty and the risk of the state to decide when to intervene, which have shown promising results on robotic tasks."
] | [] |
Ensembling Off-the-shelf Models for GAN Training | The advent of large-scale training has produced a cornucopia of powerfulvisual recognition models. However, generative models, such as GANs, havetraditionally been trained from scratch in an unsupervised manner. Can thecollective "knowledge" from a large bank of pretrained vision models beleveraged to improve GAN training? If so, with so many models to choose from,which one(s) should be selected, and in what manner are they most effective? Wefind that pretrained computer vision models can significantly improveperformance when used in an ensemble of discriminators. Notably, the particularsubset of selected models greatly affects performance. We propose an effectiveselection mechanism, by probing the linear separability between real and fakesamples in pretrained model embeddings, choosing the most accurate model, andprogressively adding it to the discriminator ensemble. Interestingly, ourmethod can improve GAN training in both limited data and large-scale settings.Given only 10k training samples, our FID on LSUN Cat matches the StyleGAN2trained on 1.6M images. On the full dataset, our method improves FID by 1.5x to2x on cat, church, and horse categories of LSUN. | [
"Improving GAN training. Since the introduction of GANs {{cite:bf168b19c5d2c37556c821b84a4471a1e4a2f8e6}}, significant advances have been induced by architectural changes {{cite:3facdc5df9896a752cf8cce3eec28f655cc935b7}}, {{cite:1e665feaffeeb777dc3033a5de01ae7673068d19}}, {{cite:95dcdf1ef360725edfccdae1c3778f4fa33c4755}}, training schemes {{cite:80a4954ee222ed4a98faf7ee4081c0701513cdf9}}, {{cite:9290ad40d674d5240184419cb80aa07a8c51aa71}}, as well as objective functions {{cite:18f4d8a1d29c8333a5844a65b740faffb11f210a}}, {{cite:bca0d71703bf0137fdbef989bdd1e0db07deec40}}, {{cite:db2004f2daec607a9f202c692c8ab33c66802953}}, {{cite:0648c22b6fd0f483aebef0b0b38a3eb5b7e74690}}, {{cite:9f85fd2463c1490475afcec6b4b24fe5809eb882}}. In previous works, the learning objectives often aim to minimize different types of divergences between real and fake distribution. The discriminators are typically trained from scratch and do not use external pretrained networks. As a result, the discriminator is prone to overfit the training set, especially for the limited data setting {{cite:6387e4ea4d115a4761e7be633341cf8cf08f8aa4}}, {{cite:8429e3eb250e188d721b0c1d123ac6a0ea10632b}}.",
"Use of pretrained models in image synthesis. Pretrained models have been widely used as perceptual loss functions {{cite:2a65c9bf2e154eaeed7f2e289cdea60827723ba9}}, {{cite:12009a3cfe138ea8be493bfcadf57bf1bf3d4ab0}}, {{cite:4aee180d04a8c5511e5bebcfb4f05f55291c0b44}} to measure the distance between an output image and a target image in deep feature space. The loss has proven effective for conditional image synthesis tasks such as super-resolution {{cite:9d0a074fe5ce2032a2c412945ad475e4f39aa78f}}, image-to-image translation {{cite:35e1a17259254dd9c6bb34f62adf76661bf36e5d}}, {{cite:9cde801433be8b27fcdf53f4ab3d779128d54e7e}}, {{cite:84a370488de192f189c745f9193e22ec368bdcf3}}, and neural style transfer {{cite:12009a3cfe138ea8be493bfcadf57bf1bf3d4ab0}}. Zhang et al. {{cite:bf18b10fd5541da5039924db0dfe08827e2c5bcb}} show that deep features can indeed match the human perception of image similarity better than classic metrics. Sungatullina et al. {{cite:c4cdbffa7205895c80729fecf08223cbb310ba13}} propose a perceptual discriminator to combine perceptual loss and adversarial loss for unpaired image-to-image translation. This idea was recently used by a concurrent work on CG2real {{cite:a64505cca254401481e521e0c66bce1a4af0b9f2}}. Our work is inspired by the idea of perceptual discriminators {{cite:c4cdbffa7205895c80729fecf08223cbb310ba13}} but differs in three ways. First, we focus on a different application of unconditional GAN training rather than image-to-image translation. Second, instead of using a single VGG model, we ensemble a diverse set of feature representations that complement each other. Finally, we propose an automatic model selection method to find models useful for a given domain. A concurrent work {{cite:03da7d0409ac74942463d7273fdf9374a1ba4d6a}} propose to reduce overfitting of perceptual discriminators {{cite:c4cdbffa7205895c80729fecf08223cbb310ba13}} using random projection and achieve better and faster GAN training.",
"Loosely related to our work, other works have used pretrained models for clustering, encoding, and nearest neighbor search during their model training. Logo-GAN {{cite:c9776193066202901950dfde1f52b16da54fcc8d}} uses deep features to get synthetic clustering labels for conditional GAN training. InclusiveGAN {{cite:1bae482704a9c18021d8d838ff796957a2b50293}} improves the recall of generated samples by enforcing each real image to be close to a generated image in deep feature space. Shocher et al. {{cite:01c8b72417eea7e8ed5f966fa8d0eb8c7caa6d2a}} uses an encoder-decoder based generative model with pretrained encoder for image-to-image translation tasks. Pretrained features have also been used to condition the generator in GANs {{cite:d09849fde72a3cba48f58426652361d714b4ecae}}, {{cite:a8d84e9ded8e72b236e2215553f0bffa16f503e9}}. Different from the above work, our method empowers the discriminator with pretrained models and requires no changes to the backbone generator.",
"Use of pretrained models in image editing. Pretrained models have also been used in image editing once the generative model has been trained. Notable examples include image projection with a perceptual distance {{cite:3b561aa2dd394f341c2d0700f4c186d195e62244}}, {{cite:5cb45bad4ef43a01878e5d0d05abbdcbdb5e4f1c}}, text-driven image editing with CLIP {{cite:65c5ab88e706993a37df01e99f69696f0840e7bc}}, finding editable directions using attribute classifier models {{cite:856c04f2160752799a1d13532857c966bef8cf94}}, and extracting semantic editing regions with pretrained segmentation networks {{cite:89a872a5487bb7b6bb71b1ccb416fa4b76686385}}. In our work, we focus on using the rich knowledge of computer vision models to improve model training.",
"Transfer learning. Large-scale supervised and self-supervised models have learned useful feature representations {{cite:70d95efa43d8e5d35af5f902f55ac9e95f2ac979}}, {{cite:66591f82a853cdc91cdd2a707f306f2ab2d98079}}, {{cite:290eab947661d7c5e31454e805d3ddd3ebdd2e92}}, {{cite:4e1fe520c786be8a00e736adfc8f05a390660d24}}, {{cite:9b6c873016c3536da0d5da484b6332a313dbb681}}, {{cite:7cb13ded6c1e18055938341f432463a044757bb9}} that can transfer well to unseen tasks, datasets, and domains {{cite:e51f8c1a4fa21aee5e40a162b6197a505354d999}}, {{cite:e69e68ea24adfbdf867077b4a304adfa43c69015}}, {{cite:1eff355a626532b909d1c20fde4204a71edec4b3}}, {{cite:bd126cebda4340fca456a6befdc38ac3cf35a51f}}, {{cite:26327cb59329245ed8a518fce0962a5bb4471d57}}, {{cite:759665ba9bf4676ef57259c8e5838aa8193e4d8d}}, {{cite:f86ae2bc281a2b01b54b9c2f1dffd96fa982e145}}, {{cite:cfe643fea1fd60fee94c8b48180fbadaea768334}}, {{cite:001be9e7a9e556bb69e032f4e9b11f5f4810526e}}.In generative modeling, recent works propose transferring the weights of pretrained generators and discriminators from a source domain (e.g., faces) to a new domain (e.g., portraits of one person) {{cite:d76c86589280a02649c2a947687f2b8493c3189d}}, {{cite:cc10bf53ceaf4f5d649e22e085710a3b9dd38414}}, {{cite:9a71cdd916c9bc20859a784a4bcf07a7517fc9cf}}, {{cite:9e1c67ea084dc6f094ae19c1109e0001734ade93}}, {{cite:60ca3b2b65788aa2b73de3fcf1213054c61d21ec}}, {{cite:ea57f01831466873d873af5211ac6b9b52b4afd7}}, {{cite:69c5ccc547f62e3f038186879aaebc72379a2c8d}}, {{cite:0a25f72f611ebae73c1fd26da893837e2d4452ec}}. Together with differentiable data augmentation techniques {{cite:8429e3eb250e188d721b0c1d123ac6a0ea10632b}}, {{cite:54a5cb3bf3e8ce0bb17bbc24901d875bdb9f9b66}}, {{cite:5a3636c620aeeccaa5aa80eb1a901e811fd6694a}}, {{cite:6387e4ea4d115a4761e7be633341cf8cf08f8aa4}}, they have shown faster convergence speed and better sampling quality for limited-data settings.Different from them, we transfer the knowledge of learned feature representations of computer vision models. This enables us to leverage the knowledge from a diverse set of sources at scale."
] | [] |
MIDI-DDSP: Detailed Control of Musical Performance via Hierarchical
Modeling | Musical expression requires control of both what notes are played, and howthey are performed. Conventional audio synthesizers provide detailed expressivecontrols, but at the cost of realism. Black-box neural audio synthesis andconcatenative samplers can produce realistic audio, but have few mechanisms forcontrol. In this work, we introduce MIDI-DDSP a hierarchical model of musicalinstruments that enables both realistic neural audio synthesis and detaileduser control. Starting from interpretable Differentiable Digital SignalProcessing (DDSP) synthesis parameters, we infer musical notes and high-levelproperties of their expressive performance (such as timbre, vibrato, dynamics,and articulation). This creates a 3-level hierarchy (notes, performance,synthesis) that affords individuals the option to intervene at each level, orutilize trained priors (performance given notes, synthesis given performance)for creative assistance. Through quantitative experiments and listening tests,we demonstrate that this hierarchy can reconstruct high-fidelity audio,accurately predict performance attributes for a note sequence, independentlymanipulate the attributes of a given performance, and as a complete system,generate realistic audio from a novel note sequence. By utilizing aninterpretable hierarchy, with multiple levels of granularity, MIDI-DDSP opensthe door to assistive tools to empower individuals across a diverse range ofmusical experience. | [
"Note Synthesis. Existing neural synthesis models allow either high-level manipulation of note pitch, velocity, and timing {{cite:ea152fe4c6a8fec48c3be8f63d16af39ce346489}}, {{cite:261ebbccc4fdb005549a912f48ac1c3fc65ecc40}}, {{cite:3027da2d06cc6f02171e2b0885911a00f4d761e9}}, {{cite:bc3112271189e44280bca972a65949c636d1fe9f}}, or low-level synthesis parameters {{cite:cedbcd4fa7bd49b9c5d9846e3ff157b105afe58b}}, {{cite:d896c0f70e414fbde6f15b4a05273cb5cac7dbec}}, {{cite:6b0df4713d80413efbb48439c701934ff4810eca}}. MIDI-DDSP connects these two approaches by enabling both high-level note controls and low-level synthesis manipulation in a single system.",
"Most related to this work is MIDI2Params {{cite:d896c0f70e414fbde6f15b4a05273cb5cac7dbec}}, a hierarchical model that autoregressively predicts frame-wise pitch and loudness contours to drive the original DDSP autoencoder {{cite:beacc4d2a6b97d79a83c15e64a576a0ca5b459dd}}. MIDI-DDSP builds on this work by adding an additional level of hierarchy for the note expression, training a new more accurate DDSP base model, and explicitly modeling the synthesizer coefficients output by that model, rather than the pitch and loudness inputs to the model. We extensively compare to our reimplementation of MIDI2Params as a baseline throughout the paper.",
"Hierarchical Audio Modelling. Audio waveforms have dependencies over timescales spanning several orders of magnitude, lending themselves to hierarchical modeling. For example, {{cite:18f39c580fc402fa390e3362d8846ebd46710320}} and {{cite:37f36bb9024109cc881af2adefd6b921f9d52d4b}} both choose to encode audio as discrete latent codes at different time resolutions, and apply autoregressive models as priors over those codes. MIDI-DDSP applies a similar approach in spirit, but constructs a hierarchy based on semantic musical structure (note, performance, synthesis), allowing interpretable manipulation by users.",
"Expressive Performance Analysis and Synthesis.Many prior systems pair analysis and synthesis functions to capture expressive performance characteristics {{cite:ced9fdb80ef3bdbdfdc89fa1fac8c9f1fea0f3fe}}, {{cite:fad7c7cb576e898070d817b54fcf67331aee7abe}}, {{cite:0a154fa43775e7a5c57502368f149a39d91d80b7}}. Such methods often use heuristic functions to generate parameters for driving synthesizers or selecting and modifying sample units.MIDI-DDSP similarly uses feature extraction, but each level is paired with a differentiable neural network function that directly learns the mapping to expression and synthesis controls for more realistic audio synthesis."
] | [] |
Incomplete Knowledge Graph Alignment | Knowledge graph (KG) alignment - the task of recognizing entities referringto the same thing in different KGs - is recognized as one of the most importantoperations in the field of KG construction and completion. However, existingalignment techniques often assume that the input KGs are complete andisomorphic, which is not true due to the real-world heterogeneity in thedomain, size, and sparsity. In this work, we address the problem of aligningincomplete KGs with representation learning. Our KG embedding frameworkexploits two feature channels: transitivity-based and proximity-based. Theformer captures the consistency constraints between entities via translationpaths, while the latter captures the neighbourhood structure of KGs viaattention guided relation-aware graph neural network. The two feature channelsare jointly learned to exchange important features between the input KGs whileenforcing the output representations of the input KGs in the same embeddingspace. Also, we develop a missing links detector that discovers and recoversthe missing links in the input KGs during the training process, which helpsmitigate the incompleteness issue and thus improve the compatibility of thelearned representations. The embeddings then are fused to generate thealignment result, and the high-confidence matched node pairs are updated to thepre-aligned supervision data to improve the embeddings gradually. Empiricalresults show that our model is more accurate than the SOTA and is robustagainst different levels of incompleteness. | [
"Knowledge Graph Alignment.Traditional cross-lingual KG alignment approaches often rely on various domain-independent features to discover cross-lingual links {{cite:cbea053c2b4d812415a1c151cb54dd54ec7e5baa}}. Modern cross-lingual KGs entity aligners exploit the emergence of graph embedding techniques and show promising results {{cite:f2b60839cfc27f8451cf6e73dbf15257d7aea1e3}}, {{cite:cf3c14d354647db0b0a23141a7c8d64ed6f400ad}}. From the two input monolingual KGs, these techniques first embed the entities into low-dimensional vector spaces. The identical entities are then retrieved based on the similarity of the learned representations. The earlier techniques of this paradigm, including MTransE {{cite:26760b0e0b8696ed5182920f0d3f291dae942373}}, JAPE {{cite:a9e6a76528077be8ccf042dbd6d821af0e6dcd5c}}, ITransE {{cite:dd97a5a58203b3d587f92ca7374769a153501417}}, and BootEA {{cite:a4c4f5fcc47f6f92eba5af2970d89db72b3a0971}}, often employ shallow embedding techniques to embed the entities and the relations using the translation constraint {{cite:e0a9345873f6e605fe23ac6128b53bfd7a2a208f}}, which assumes that the embedding distance between two any entities is equal to the connecting relation.",
"Due to the lack of information integration of shallow models, recent embedding based alignment techniques employ graph convolutional network (GCN) {{cite:38a042e0d6c61861bb4009d83b3938223134befe}} to exploit the complex nature of KGs better. GCN-Align {{cite:8555b2654ac51fdebfc49eebaf6bcf765640a2f3}} employs GCNs to capture at the same time the relational and the individual entity footprint. RDGCN {{cite:16a4cae1de01f31ad8af24cefee15958bdf0cd13}} introduces a two-layers GCN with highway gates to extract attentive interactions between each KG and its dual relation. MUGNN {{cite:fee08e21a170606feec07570808a03fc20a99da2}} preprocesses the input KGs with grounding rules before forwarding them to a dual-channel GCNs model. KDcoE {{cite:813948c0cf61f86389eb1a90a29b7662a4889e05}} leverages co-training of a KG embedding and a literal description embedding to enhance the semi-supervised learning of multilingual KG embeddings. MultiKE {{cite:8e66643fcd88e6236a47fb8a342baf1ad11e160e}} unifies multiple views of entities to learn embeddings for entity alignment. KG-matching {{cite:8ce2420365d68010440d043d3144b81adc36c207}} constructs a relation network namely topic entity graph to encode the entities' name by a word-based LSTM. REA {{cite:52c18ca64c9619407a6272a6ea3c985e6816d5ae}} proposes a two-component model to enhance the noise robustness of the aligner. AliNet {{cite:dceb31d258fc2d7d4184189ebaf661e9d9746203}} employs entities' distant neighbours to expand their neighbour overlapping {{cite:e2ca76733aa8749cf4d45162722e7ffaeec6d835}}.",
"Our work goes beyond the state-of-the-art by enhancing the GNN properties with attention mechanism and entity-relation composition to capture the relational correlation effectively. We also combine the GNN with a translation-based channel to guarantee the knowledge transfer between the input KGs and thus mitigate the incompleteness issue. Similar to our work are GCN-Align and MultiKE, which also use a multi-view scheme, but our GNN channel leverages relation embeddings that better capture relational information. Also, unlike the existing models such as RDGCN and GCN-Align that stacks GNN models, our framework utilizes a non-GNN channel to cover the GNN-based channel and overcome the GNN weaknesses, such as information dilution and noise amplification issue.",
"Knowledge Graph Completion. The Knowledge Graph Completion (KGC) techniques aim to predict the missing triples in the KGs automatically. Existing KGC techniques often embed entities and relations into low-dimensional vectors, then design a score function to check the candidate triplets' plausibility. The earlier methods often employ shallow embedding models with translation-based scoring function {{cite:e0a9345873f6e605fe23ac6128b53bfd7a2a208f}}, {{cite:c5f9c066fb92d6b3dc135baed94d3dd4285293df}}. Recently, KGC techniques using deep embedding model and complex score have been proposed, including CNNs based model {{cite:062a65425942f9bc31aa5ecd6204786e0573ff13}}, {{cite:30a776056a8f1f54a989c204e2f2f9bb1ebc5740}}, RNNs based models {{cite:1e45ae72452ab1cae5fe92d169f6e2134f0092c4}}, and GCNs based model {{cite:621f766657984915cef1dff3679c5659630fee9a}}. However, they are surprisingly found to be unstable across different datasets due to the overloading of unrelated objectives {{cite:a7a57957561c44762f5f54095cbe15eb98df1f84}}. Thus, we choose the simple yet powerful translation-based model to supply the main GNN-based channel in our framework. Our work also goes beyond the existing KGC framework by answering at the same time two questions: (i) given any two entities, whether there is a triple connecting them and (ii) which is the relation type of the triple between the two entities; while the existing frameworks often answer only the latter question. It is worth noting that the former question is challenging, as the number of disjoint entities outweighs that of connecting ones and thus might cause low recall in detecting the missing triples. We tackle this challenge by exchanging the useful patterns of high-confidence entities retrieved from the alignment of the two input KGs, thus making the information of the KGs complete for each other."
] | [] |
Neural Architectures for Biological Inter-Sentence Relation Extraction | We introduce a family of deep-learning architectures for inter-sentencerelation extraction, i.e., relations where the participants are not necessarilyin the same sentence. We apply these architectures to an important use case inthe biomedical domain: assigning biological context to biochemical events. Inthis work, biological context is defined as the type of biological systemwithin which the biochemical event is observed. The neural architectures encodeand aggregate multiple occurrences of the same candidate context mentions todetermine whether it is the correct context for a particular event mention. Wepropose two broad types of architectures: the first type aggregates multipleinstances that correspond to the same candidate context with respect to eventmention before emitting a classification; the second type independentlyclassifies each instance and uses the results to vote for the final class, akinto an ensemble approach. Our experiments show that the proposed neuralclassifiers are competitive and some achieve better performance than previousstate of the art traditional machine learning methods without the need forfeature engineering. Our analysis shows that the neural methods particularlyimprove precision compared to traditional machine learning classifiers and alsodemonstrates how the difficulty of inter-sentence relation extraction increasesas the distance between the event and context mentions increase. | [
"The problem of relation extraction (RE) has received extensiveattention {{cite:6fa7370951d860f9449368ac78c4e645ab088b3c}}, {{cite:801426b8bab0abd12b568265268cdb1f245fa920}}, including within the biomedicaldomain {{cite:32743cef0339849ac5a15cece696cc4d5ff809ef}}, {{cite:6a763d2d3b9008d908040c3a58fafd590bfe8acd}}, with recent promising resultsincorporating distant supervision {{cite:54e9202e1a513c27ace774be4ccad075d667cccc}}. However, most of the work focuses on identifying relations among entities within thesame sentence. In the biological context association problem, the entities are potentially located in different sentences, making the context association task an instance of an inter-sentence relation extraction problem.",
"Previous work in inter-sentence relation extraction includes {{cite:9edc0aa3774f66440dcbdd4573fb36fe63e6c637}},which combined within-sentence syntactic features with an introduced dependencylink between the root nodes of parse trees from different sentencesthat contain a given pair of entities. {{cite:e5b5abf1e8994782b7572948f0a7209c8c5094bc}} proposes an inter-sentence relation extraction model that builds a labeled edge graph convolutional neural network model on a document-level graph. There have also been efforts to create language resources to foster the development of inter-sentence relation extraction methods. {{cite:db56cffff14e7693f99fa70d55cb92cacc3ae98b}} propose an open domain data set generated from Wikipedia to Wikidata. {{cite:bf8627a515a05358c27cfa6cb09cfef779013d9e}} propose an inter-sentence relation extraction data set constructed using distance supervision.Modeling inter-sentence relation extraction using transformer architectures require processing potentially long sequences. Long input sequences are problematic because computing the self-attention matrix has quadratic runtime and space complexity relative to the its length. This observation has motivated research efforts to generate efficient approximations of self-attention. {{cite:b08bd9439440b0ced31dd6bd8aaa3b2017b36bf4}} proposes a sparse, drop-in replacement for the self-attention mechanism with linear complexity that relies on sliding windows and selects domain-dependent global attention tokens from the input sequence. {{cite:28116a06d43390c37d0eb302d8ba9a7b0a055261}} proposes a lower-rank approximation of the self-attention matrix to linearize the complexity. {{cite:ac3292808d5291f18202490fd52b3f5316b85160}} ommits the pair-wise dependencies between the input tokens and then factorizes the attention matrix to reduce its rank. Other approaches {{cite:20fc3bfd200fa7ae81d6a6497d642e1ee410642a}} rely on kernel functions to compute approximations with linear time and space complexity. {{cite:4ed847dfb87644d1e3fd572e615436fb2211bdfa}} takes this approach further by using relative position encodings, instead of absolute ones.",
"Prior work has specifically studiedthe contextualization of information extraction in the biomedical domain. {{cite:0c733d7fd3d2a8a45daaf12a4a21728edfe7d375}} associates anatomical contextual containers with event mentions that appear in the same sentence via a set of rules that considers lexical patterns in the case of ambiguity and falls back to token distance if no pattern is matched. {{cite:2c8be6a9211e4a940a9859b440f5d150684cd05b}} elaborates on the same idea by incorporating dependency trees into the rules instead of lexical patterns, as well as introducing a method to detect negations and speculative statements.",
"{{cite:a842bb8d7b28b18d8a9e7eb0de1c1cc9823576f9}} previously studied the task of context association for the biomedical domain and framed it as a problem of inter-sentence relation extraction.This work presentsset of linguistic and lexical features that describe the neighborhood of the participant entities and proposes an aggregation mechanism that results in improved context association.",
"Previous work relied upon feature engineering to encode the participants and their potential interactions. State-of-the-art NLP research leverages large language models to exploit transfer learning. Models such as {{cite:d66d6f3d5458edf7c0e360b71ef5e5a54f1d1822}}, and similar transformer based architectures {{cite:5ee7aad401b97d063fd7fea7200dba60751e1502}} better capture the semantics of text based on its surrounding context with unsupervised pre-training over extremely large corpora. Specialized models, such as {{cite:9d6b1265ee35a13cc6094924989d3e9638f1e556}}, {{cite:12860c44c04bf4786e4a30a24aa20cfa17c4bd32}}, {{cite:fd1d4b9d1ce61d54837f916c2063584c406f15eb}} refine language models by continuing pre-training with in-domain corpora.",
"To the best of our knowledge, the work presented here is the first to propose and analyze deep-learning aggregation and ensemble architectures for many-to-one, long-distance relation extraction."
] | [] |
A Robust Optimization Approach to Deep Learning | Many state-of-the-art adversarial training methods leverage upper bounds ofthe adversarial loss to provide security guarantees. Yet, these methods requirecomputations at each training step that can not be incorporated in the gradientfor backpropagation. We introduce a new, more principled approach toadversarial training based on a closed form solution of an upper bound of theadversarial loss, which can be effectively trained with backpropagation. Thisbound is facilitated by state-of-the-art tools from robust optimization. Wederive two new methods with our approach. The first method (Approximated RobustUpper Bound or aRUB) uses the first order approximation of the network as wellas basic tools from linear robust optimization to obtain an approximate upperbound of the adversarial loss that can be easily implemented. The second method(Robust Upper Bound or RUB), computes an exact upper bound of the adversarialloss. Across a variety of tabular and vision data sets we demonstrate theeffectiveness of our more principled approach -- RUB is substantially morerobust than state-of-the-art methods for larger perturbations, while aRUBmatches the performance of state-of-the-art methods for small perturbations.Also, both RUB and aRUB run faster than standard adversarial training (at theexpense of an increase in memory). All the code to reproduce the results can befound at https://github.com/kimvc7/Robustness. | [
"Over the last two decades, robust optimization has become a successful approach to solve optimization problems under uncertainty. For an overview of the primary research in this field we refer the reader to {{cite:761b0592477debf9ba2d1601c7cb1f4c86a4b895}}. Areas like mathematical programming and engineering have long applied these tools to develop models that are robust against uncertainty in the parameters, which may arise from rounding or implementation errors. For many applications, the robust problem can be reformulated as a tractable optimization problem, which is referred to as the robust counterpart. For instance, for several types of uncertainty sets, the robust counterpart of a linear programming problem can be written as a linear or conic programming problem {{cite:6f0e8867ad885a6513370f7602be519b0efe97e5}}, which can be solved with many of the current optimization software. While there is not a systematic way to find robust counterparts for a general nonlinear uncertain problem, multiple techniques have been developed to obtain tractable formulations in some specific nonlinear cases.",
"As shown in {{cite:6f0e8867ad885a6513370f7602be519b0efe97e5}}, the exact robust counterpart is known for Conic Quadratic Problems and Semidefinite problems in which the uncertainty is finite, an interval or an unstructured norm-bounded set. More generally, it is shown in {{cite:7b6d13d3ec75dac6f05e239c76f6998470dfe05a}} that for problems in which the objective function or the constraints are concave in the uncertainty parameters, Fenchel duality can be used to exactly derive the corresponding robust counterpart. While the result does not necessarily have a closed form, the authors show that it yields a tractable formulation for the most common uncertainty sets (e.g. polyhedral and ellipsoidal uncertainty sets).",
"The problem becomes significantly more complex when the functions in the objective or in the constraints are instead convex in the uncertainty {{cite:5eab523fa9bcbc6ed1c36d84d4dc2282b7e120ec}}. Since obtaining exact robust counterparts in these cases is generally infeasible, safe approximations are considered instead {{cite:45ee378ed43d42cdcd998b4eb1bf3982ea868722}}. For instance, {{cite:87a33ee024d5b9ca37bcbb9c39cedb1203fb02e0}} develop safe approximations for the specific cases of second order cone and semidefinite programming constraints with polyhedral uncertainty. These techniques are generalized in {{cite:1b7da919c6c27b066aca7e4ad702399dd72b2101}}, where the authors convert the robust counterpart to an adjustable robust optimization problem that produces a safe approximation for any problem that is convex in the optimization variables as well as in the the uncertain parameters.",
"Even though the approaches mentioned above consider uncertainty in the parameters of the model as opposed to uncertainty in the input data, the same techniques can be utilized for obtaining robust counterparts in the latter case. In fact, the robust optimization methodologies have recently been applied to develop machine learning models that are robust against perturbations in the input data. In {{cite:ed4889433ccb099b9fdc5bca245da2154d660a24}}, for example, the authors consider uncertainty in the data features as well as in the data labels to obtain tractable robust counterparts for some of the major classification methods: support vectormachines, logistic regression, and decision trees. However, due to the high complexity of neural networks as well as the large dimensions of the problems in which they are often utilized, robust counterparts or safe approximations for this type of models have not yet been developed. There are two major challenges with applying robust optimization tools for training robust neural networks:",
"Neural networks are neither convex nor concave functions: As mentioned earlier, robust counterparts are difficult to find for a general problem. Although plenty of work has been done to find tractable reformulations as well as safe approximations, all of them rely on the underlying function being a convex or a concave function of the uncertainty parameters. Unfortunately, neural networks don't satisfy either condition, which makes it really difficult to apply any of the approaches discussed above.The robust counterpart needs to preserve the scalability of the model: Neural networks are most successful in problems involving vision data sets, which often imply large input dimensions and enormous amount of data. For the most part, they can still be successfully trained thanks to the fact that back propagation algorithms can be applied to solve the corresponding unconstrained optimization problem. However, the robust optimization techniques for both convex and concave cases often require the addition of new constraints and variables for each data sample, increasing significantly the number of parameters of the network and making it very difficult to use standard machine learning software for training.",
"A straightforward way to overcome both of these difficulties would be to replace the loss function of the network with its first order approximation. However, this loss function is usually highly nonlinear and therefore the linear approximation is very inaccurate. Our method aRUB explores a slight modification of this approach that significantly improves adversarial accuracy by considering only the linear approximation of the network's output layer.",
"Alternatively, a more rigorous approach to overcome problem (i) would be to piece-wise analyze the convexity of the network and apply the robust optimization techniques in each piece separately, but this approach would introduce additional variables that are in conflict with requirement (ii). For our RUB method we then develop a general framework to split the network by convexity type, and we show that in the specific case in which the uncertainty set is the {{formula:f69aee06-0fb2-4dbe-8530-79a3d8c920c7}} norm bounded sphere, we can solve for the extra variables and obtain an unconstrained problem that can be tractably solved using standard gradient descent techniques."
] | [] |
Joint Device Detection, Channel Estimation, and Data Decoding with
Collision Resolution for MIMO Massive Unsourced Random Access | In this paper, we investigate a joint device activity detection (DAD),channel estimation (CE), and data decoding (DD) algorithm for multiple-inputmultiple-output (MIMO) massive unsourced random access (URA). Different fromthe state-of-the-art slotted transmission scheme, the data in the proposedframework is split into only two parts. A portion of the data is coded bycompressed sensing (CS) and the rest is low-density-parity-check (LDPC) coded.In addition to being part of the data, information bits in the CS phase alsoundertake the task of interleaving pattern design and channel estimation (CE).The principle of interleave-division multiple access (IDMA) is exploited toreduce the interference among devices in the LDPC phase. Based on the beliefpropagation (BP) algorithm, a low-complexity iterative message passing (MP)algorithm is utilized to decode the data embedded in these two phasesseparately. Moreover, combined with successive interference cancellation (SIC),the proposed joint DAD-CE-DD algorithm is performed to further improveperformance by utilizing the belief of each other. Additionally, based on theenergy detection (ED) and sliding window protocol (SWP), we develop a collisionresolution protocol to handle the codeword collision, a common issue in the URAsystem. In addition to the complexity reduction, the proposed algorithmexhibits a substantial performance enhancement compared to the state-of-the-artin terms of efficiency and accuracy. | [
"Next generation multiple access (NGMA) aims to support massive connectivity scenarios for many envisioned Internet of Things (IoT) applications, e.g. manufacturing, transportation, agriculture, and medicine, to be efficiently and flexibly connected {{cite:51f549c42f8e86ec7fed129f304793fac2a47376}}, {{cite:fe13778a3c39d42bb9fddc9eae772c04d4138e5e}}, {{cite:63c41d245b3a84f66af0a3636785808f9168121e}}, {{cite:4ddd89f69345aeb3eea4d83013a455165da60d2a}}, {{cite:a826360bf1ae42555d34b12663fd290d39409328}}, {{cite:158a9db93fc71b6b16b6497fb670bb43bea2a046}}. A generic scenario for IoT connectivity involves a massive number of machine-type connections {{cite:fe13778a3c39d42bb9fddc9eae772c04d4138e5e}}, known as the massive machine-type communication (mMTC), which is one of the main use cases of the fifth-generation (5G) and beyond {{cite:158a9db93fc71b6b16b6497fb670bb43bea2a046}}. Compared to human-type communications (HTC), MTC has two distinct features: sporadic traffic and short packet communication. The sporadic access means that for any given time, for the sake of long battery life and energy saving, only a small fraction within a huge number of devices are active for data transmission {{cite:63c41d245b3a84f66af0a3636785808f9168121e}}. While the small data payloads result in a fall of spectrum efficiency and an increase of latency when applied with traditional multiple access protocols. In this regard, a number of scalable random access schemes have been investigated in the content of massive connectivity. Conceptually, there are two lines of work for that end, namely, individual and common codebook-based approaches, respectively.",
"In the spirit of the individual codebook based approach, each device is equipped with a unique codebook for the identification conducted by the base station (BS) {{cite:63c41d245b3a84f66af0a3636785808f9168121e}}, {{cite:ca1b3f5818f6eac7eb75099d04b39ede58714a2c}}. Since the BS is interested in the IDs of devices, this framework can be referred to as sourced random access (SRA). A typical transmission process includes two phases in SRA, namely, training and data transmission, respectively. In the training phase, the BS conducts the task of device activity detection (DAD) and channel estimation (CE) based on the pre-allocated unique pilot sequences transmitted by active devices, while the data transmission is performed in the following phase. The task of DAD-CE can be modeled as the compressed sensing (CS) problem and there have been quantities of algorithmic solutions for this issue. Among these CS recovery techniques, the approximate message passing (AMP) algorithm, which is first proposed in {{cite:f98d5529f35409a299c42c84062bd12962839705}} for the single measurement vector (SMV) based problem, exhibits great performance in terms of accuracy and complexity. Since the publication of {{cite:f98d5529f35409a299c42c84062bd12962839705}}, there have been vast variants based on the AMP algorithm, such as multiple measurement vector (MMV) AMP {{cite:63c41d245b3a84f66af0a3636785808f9168121e}}, generalized MMV (GMMV) AMP {{cite:0018dc51c57538e0e96b23aa686627c8c4f5a877}}, orthogonal AMP (OAMP) {{cite:569697e7f85094f07a2276123700dcc572aca5b0}}, generalized AMP (GAMP) {{cite:fa21aea798d25ce4794c982f9e1004dbe1469c0c}} and many other works. Another line of work is based on the belief propagation (BP) algorithm {{cite:74f47ac4f0196b1ba9ff4120218cd3cf42bcc2fa}}, which models the problem as a graph and iteratively calculates the messages among different nodes {{cite:1b941f64add6d02ed3230d9b8d94aa27fca9bb04}}, {{cite:e51ec82269f89da567c6b499420849fff301d2c2}}, {{cite:a64c4525254d53e7b8949ab5e4f4db15eba1f516}}, {{cite:6e71dfd19dfd4f70fe166f9a01373698696d2999}}. Thanks to the consistency of the message updating rules, the messages can always be jointly updated within iterations to obtain improved performance. However, the premise of these algorithms is that each device should have a unique pilot sequence. As we jump out of these algorithmic solutions and go back to the essence of the individual codebook, we find this is inapplicable to assign each device a unique pilot in the mMTC scenario with a huge amount of potential devices. Meanwhile, it is a waste of resources especially with sporadic traffic. To mitigate this issue, the study on the framework of common codebooks is rapidly developed.",
"As opposed to the case of individual codebooks, active devices choose codewords from a common codebook for their messages. In this regard, the task of BS is only to produce the transmitted messages regardless of the corresponding IDs, thus leading to the so-called unsourced random access (URA). The main difference between URA with other grant-based and grant-free random access protocols is that the BS does not perform device identification but only decodes the list of active device messages up to permutations {{cite:158a9db93fc71b6b16b6497fb670bb43bea2a046}}. Conceptually, there are two distinct advantages for this novel framework: {{formula:c681a571-70ff-403c-ab2c-17eaa357f1c4}} the BS is capable of accommodating massive devices, since all devices share the common codebook instead of being assigned unique preamble sequences as in traditional schemes. {{formula:927f6d2c-e8a2-40d8-b44f-96de20df0a92}} No device identification information is needed to be embedded in the transmitted sequence. As such, the overhead will be reduced, contributing to improved efficiency. The study on URA is first reported in {{cite:068d117929c5e20bb805e83e8c15100b58043860}}, in which Polyanskiy considered the scenario of a massive number of infrequent and uncoordinated devices and discussed a random coding existence bound in the Gaussian multiple-access channel (GMAC). Since the publication of {{cite:068d117929c5e20bb805e83e8c15100b58043860}}, there have been many works devoted to approaching that bound {{cite:89b00e7f0eedaa5870d9cada01e30b9a7d1abc84}}, {{cite:f5992fc45c331736b49f347c50c6a3edf7ef858e}}, {{cite:520ad9ad7c6ae93a7637e571891a1f630e425308}}, {{cite:26fb857f2baa23912e61f20ffc16ffe870429870}}, {{cite:c8d84665fe919657da4e799b056234300838fed7}}, {{cite:52822f3b8cf1f2539ce37727fd9fce0b50364a4e}}, {{cite:03c5c2ce44a60769c3f9bd564f2fdfc8e6a832a6}}, {{cite:5f0efb831b0976aaa3bca7aff88f39bc7902bb98}}. A low complexity coding scheme is investigated in {{cite:89b00e7f0eedaa5870d9cada01e30b9a7d1abc84}}. Based on the combination of compute-and-forward and coding for a binary adder channel, it exhibits an improved performance compared with the traditional schemes, such as slotted-ALOHA and treating interference as noise (TIN). However, the size of the codebook increases exponentially with the blocklength, resulting in difficulty in achieving that underlying bound. To mitigate this issue, a slotted transmission framework is proposed in {{cite:f5992fc45c331736b49f347c50c6a3edf7ef858e}}, referred to as the coded compressed sensing (CCS) scheme. This cascaded system includes inner and outer codes. The outer tree encoder first splits the data into several slots with parity bits added. The CS coding is then employed within each slot to pick codewords from the codebook. The blocklength, as well as the size of the codebook in each slot, are greatly reduced, thus leading to a relaxation of the computational burden. On the receiver side, the inner CS-based decoder is first utilized to recover the slot-wise transmitted messages within each slot. The outer tree decoder then stitches the decoded messages across different slots according to the prearranged parity. This structure is inherited by the later works in {{cite:520ad9ad7c6ae93a7637e571891a1f630e425308}}, {{cite:26fb857f2baa23912e61f20ffc16ffe870429870}}, where the authors exploit the concept of sparse regression code (SPARC) {{cite:50903d524fbc9a27c956321bf8f31c416a8f6244}} to conduct the inner encoding and decode it with the AMP algorithm. Some other coding schemes for URA are also reported in {{cite:c8d84665fe919657da4e799b056234300838fed7}}, {{cite:52822f3b8cf1f2539ce37727fd9fce0b50364a4e}}, {{cite:03c5c2ce44a60769c3f9bd564f2fdfc8e6a832a6}}, {{cite:5f0efb831b0976aaa3bca7aff88f39bc7902bb98}}. The polar coding schemes based on TIN and successive interference cancellation (SIC) are investigated in {{cite:c8d84665fe919657da4e799b056234300838fed7}}, {{cite:52822f3b8cf1f2539ce37727fd9fce0b50364a4e}}. In {{cite:03c5c2ce44a60769c3f9bd564f2fdfc8e6a832a6}}, {{cite:5f0efb831b0976aaa3bca7aff88f39bc7902bb98}}, the author consider a hierarchical framework, where the device's data is split into two parts. A portion of the data is encoded by a common CS codebook and recovered by the CS techniques, while the rest is encoded by the low-density parity check code (LDPC) {{cite:72b953d186c31d02afd55d5e2d6bf07ae4900d65}} with the key parameters conveyed in the former part. Besides the above works considering the GMAC, there are also works in the fading channel {{cite:b9814fadb526d699dcbd8acb94dc205645e7df9e}}, {{cite:a676927f277c26b8725d28d3f7afa55742b4daa4}}, {{cite:755ade9916fbc05fe7f62caddba5e6e4c53386fa}}.",
"Besides the above works considering a single receive antenna at the BS, the study of massive URA in multiple-input multiple-output (MIMO) systems has also drawn increasing attention. The BS equipped with multiple antennas provides extra dimensions for the received signal, thus offering more opportunity for signal detection and access of a massive number of devices. A coordinated-wise descend activity detection (AD) algorithm is proposed in {{cite:a24bcf5ec5c759b93c19a9b8b9e003e6355b879c}}, {{cite:835ae737164d4c2088fe73f25631af1a4bbe0fcb}}, which finds the large-scale fading coefficients (LSFCs) of all the devices (coordinates) iteratively. The authors adopt a concatenated coding scheme with the aforementioned tree code as the outer code, and a non-Bayesian method (i.e., maximum likelihood (ML) detection) is leveraged as the inner decoder based on the covariance of the received signal, referred to as the covariance-based ML (CB-ML) algorithm. However, the performance of the CB-ML algorithm degrades dramatically when the number of antennas is less than that of the active devices. There are also some improvements to this algorithm in terms of complexity or accuracy {{cite:4ddd89f69345aeb3eea4d83013a455165da60d2a}}, {{cite:a0b8972cb12af815f6021d535fcf78f02035d9aa}}. The slotted transmission framework proposed in {{cite:4ddd89f69345aeb3eea4d83013a455165da60d2a}} eliminates the need for concatenated coding. That is, no parity bit is added within each slot and the slot-wise messages can be stitched together by clustering the estimated channels. However, the channels in {{cite:4ddd89f69345aeb3eea4d83013a455165da60d2a}} are assumed to be identically distributed over all the slots, which is difficult to hold in practice. Besides, the collision in codewords (more than one device chooses the same codeword) will lead to a superimposition of the estimated channels, resulting in a failure of the stitching process. Currently, the design of an efficient collision resolution scheme in URA remains missing. In {{cite:a0b8972cb12af815f6021d535fcf78f02035d9aa}}, a more efficient coordinate selection policy is developed based on the multi-armed bandit approaches, leading to a faster convergence than the CB-ML algorithm. Nevertheless, these slotted transmission frameworks, such as CB-ML and its variants, all map a piece of data within each slot to a long transmitted codeword based on the CS coding, resulting in low efficiency. In this regard, the algorithmic design for the MIMO massive URA with high efficiency and accuracy remains missing."
] | [] |
The strong component structure of the barely subcritical directed
configuration model | We study the behaviour of the largest components of the directedconfiguration model in the barely subcritical regime. We show that with highprobability all strongly connected components in this regime are either cyclesor isolated vertices and give an asymptotic distribution of the size of the$k$th largest cycle. This gives a configuration model analogue of a result of{\L}uczak and Seierstad for the binomial random digraph. | [
"The study of the giant component in random graph models was initiated by the seminal paper of Erdős and Rényi {{cite:f756425b25126c1914f4356f06fa838b06a6ff96}} regarding the giant component in {{formula:12a131e4-fe8e-4139-9426-3d6d1122c199}} .Since then the appearance of a giant component in various models has remained an active topic of study in the area.",
"When working with directed graphs, there are a number of types of connected component which are of interest.In this paper we concern ourselves with the strongly connected components.The study of strongly connected components in random digraphs began in the model {{formula:6f485438-5986-443e-8c10-f8a6fc2ad2dd}} where we include each possible edge with probability {{formula:8456d68f-22cc-4427-91b8-65a54d058c58}} independently.Karp {{cite:6844a630cdc662b469feb8b2ceec8006a9aa38e9}} and Łuczak {{cite:6268df4f4eaa035949d47cd2b67d3e975f3cfc95}} independently showed that when {{formula:943670f5-c2c7-476f-b66d-b9dddf565d88}} , then {{formula:8b6d53fb-14fd-4466-9d8b-dff87ed76f4f}} has all strongly connected components of size {{formula:dc371f69-524a-49f5-8118-e34021dac902}} if {{formula:eebc3b19-2eb3-48e6-a488-b4f9a61e845b}} .If instead {{formula:83b237e1-d18c-459e-8c2e-8452044de9df}} they showed that there exists a unique strongly connected component of linear order with all other components of size {{formula:1a934235-a05e-4e66-8903-48e5c9500bdc}} .The case {{formula:f92b90f2-1419-46b6-aa62-d1e7ab64d6f2}} with {{formula:2e714500-e8ac-4d3d-ad8b-d043b69ce608}} and {{formula:8961a7fb-aa01-4ffa-95b1-bf276fe234d1}} was studied by Łuczak and Seierstad {{cite:32feb045353f45ff8156f94f4be1fbf11af97e59}}.They showed that if {{formula:f809e73a-27f4-44d4-8611-c1f6063e6d75}} , there is a unique strongly connected component of size {{formula:d7c2b34c-bb97-423e-a59f-15f1df2bbf23}} and other components all of size {{formula:3863bebd-145a-46bf-9755-e22834646010}} .When {{formula:f633dc55-5bcf-4777-ab5a-d8e6359f74a2}} they showed an analogue of thm:subcritmainthm:",
"Theorem 1.4 Let {{formula:063c0487-f74e-4b35-9f3f-0bf215d8bc0d}} where {{formula:d174f52d-c8ec-4439-9cef-8106283ba552}} but {{formula:ec98d1c2-5665-4277-95ec-8dfd1802c9e9}} . Assume {{formula:36a18392-183d-4459-b9ff-d06285bca64a}} is a constant and let {{formula:12cd6bf9-c687-4d7b-9b51-c98ca78998f3}} denote the size of the {{formula:ad927fad-bad1-4b38-a524-15ed07d0c937}} th largest strongly connected component. Then, asymptotically almost surely, {{formula:96d4cb95-7f9a-4f70-81bb-f119c90d3e09}} contains no complex components and{{formula:cf997049-cf64-468e-9016-f176d8b95d23}} ",
"where {{formula:b7f5adce-5d40-425c-99a7-2bf4effbd3ab}} .",
"The so-called critical window, when {{formula:a1dbda26-bee9-4ce5-b38d-1ea209ecdc7b}} has also been the subject of some study.In {{cite:d2f053c31130ff650cec9ab1e2135160219f0d30}} the author showed bounds on the size of the largest strongly connected component in this regime which are akin to bounds obtained by Nachmias and Peres for the undirected case {{cite:08e34254c9570a3a5a7fce558f5f3dae51d2bbea}}.Moreover, Goldschmidt and Stephenson {{cite:9a04073f142de7ea8d36c26d5e9b8d0cd5e60177}} gave a scaling limit result for the largest strongly connected components in the critical window.",
"The directed configuration model has also been studied previously.It was first studied by Cooper and Frieze who showed that provided the maximum degree {{formula:bc7f051f-3d2c-40cd-af22-901847ba417a}} then if {{formula:05076a3c-99d1-4f5a-aa62-fc6cb9880f0c}} , there is no all strongly connected components are small and if {{formula:bd879b40-0239-4fd7-b40a-ed8494030877}} there is a giant strongly connected component of linear size.The assumptions on the degree sequence have subsequently been relaxed, Graf {{cite:5f236a8037b81eff5119c517ae4816c6f331b2ab}} showed {{formula:1ce69892-781a-43b4-8898-8c0937b3683a}} is enough to draw the same conclusion and Cai and Perarnau {{cite:f85e9ca66ee41b0329bcca337941715abcc661c2}} improved this further to only require bounded second moments.A scaling limit result has been recently obtained at the exact point of criticality, {{formula:6760fd68-a078-4ef5-a993-842101c83dcc}} by Donderwinkel and Xie {{cite:51e09269112303dc35ccbf8cc7a9be2d29a17180}} in a very closely related model where vertices' degrees are sampled from a limiting degree sequence."
] | [] |
Contrastive Explanations for Comparing Preferences of Reinforcement
Learning Agents | In complex tasks where the reward function is not straightforward andconsists of a set of objectives, multiple reinforcement learning (RL) policiesthat perform task adequately, but employ different strategies can be trained byadjusting the impact of individual objectives on reward function. Understandingthe differences in strategies between policies is necessary to enable users tochoose between offered policies, and can help developers understand differentbehaviors that emerge from various reward functions and traininghyperparameters in RL systems. In this work we compare behavior of two policiestrained on the same task, but with different preferences in objectives. Wepropose a method for distinguishing between differences in behavior that stemfrom different abilities from those that are a consequence of opposingpreferences of two RL agents. Furthermore, we use only data on preference-baseddifferences in order to generate contrasting explanations about agents'preferences. Finally, we test and evaluate our approach on an autonomousdriving task and compare the behavior of a safety-oriented policy and one thatprefers speed. | [
"In the recent years various methods for explaining either one decision or the entire behavior of RL system have been proposed {{cite:40fd4cf7e7120259de2c39ffa24de0bc1a6da5ca}}.However, despite the fact that research shows humans tend to seek contrasting explanations when reasoning about an event {{cite:90c563169018f8aab1b4e31efa980eee905fce3f}}, there is still limited work in the field of comparing alternative behaviors of reinforcement learning systems.",
"{{cite:72f461625089b5db2e6b342a60892c21107e1646}}, {{cite:318a1816d51d1adea9e1ac1d4c769424233e9cf9}} approached the problem of explainability from a causal perspective and proposed a method for generating local explanations that contrast alternative actions in a specific state. The approach however requires a hand-crafted causal model of the environment, which may be difficult to obtain and requires expert knowledge. Additionally, authors focus on generating local explanations, while we interpret and compare global behavior of agents.",
"Summarisation methods, one of the most notable global approaches for condensing and explaining the agent behavior have also been used to compare multiple RL policies. {{cite:e6ea7af993ecbc39d372bb47c56e3be6b4e998f1}} generated contrasting summaries of agents' behavior in order to highlight the differences in their capabilities. Similar to our work, {{cite:019a7f4887f3ccd9a84c8f44e4da2fd30e110638}} used the notion of disagreement between policies to detect and analyse situations where two policies pick different actions, but opted for explaining policies' differences through contrasting summaries. However, both approaches focus only on extracting discrepancies between agents' abilities, and disregard the potential difference in their strategies. Additionally, applicability of summarisation methods is limited to tasks with visual input.",
"Most relevant to our work, {{cite:67dde3f88e311b06eb9537e6ed8cc02ce5230253}} compare the outcomes of following different policies from a specific state to justify agent's choice of action. However, their work focuses only on local explanations, and requires manual encodings of states and outcomes. In contrast, in our work we aim to provide global comparisons of policies and do not rely on hand-crafted interpretable features."
] | [] |
Deep Spatially and Temporally Aware Similarity Computation for Road
Network Constrained Trajectories | Trajectory similarity computation has drawn massive attention, as it is corefunctionality in a wide range of applications such as ride-sharing, trafficanalysis, and social recommendation. Motivated by the recent success of deeplearning technologies, researchers start devoting efforts to learning-basedsimilarity analyses to overcome the limitations (i.e., high cost and pooradaptability) of traditional methods. Specifically, deep trajectory similaritycomputation aims to learn a distance function that can evaluate how similar twotrajectories are via neural networks. However, existing learning-based methodsfocus on spatial similarity but ignore the time dimension of trajectories,which is suboptimal for time-aware applications. Besides, they tend todisregard the embedding of trajectories into road networks, restricting theirapplicability in real scenarios. In this paper, we propose an effectivelearning-based framework, called ST2Vec, to perform efficient spatially andtemporally aware trajectory similarity computation in road networks. Finally,extensive experimental evaluation using three real trajectory data sets showsthat ST2Vec outperforms all the state-of-the-art approaches substantially. | [
"We review related work on trajectory similarity in terms of non-learning-based methods and learning-based methods.",
"Non-learning-based methods {{cite:f6abe23374ec58cbf713b93b14fd37044c8560a7}}, {{cite:b3d3b0f01d8f6eac071f1d9c1ebd98eeb5bced62}}, {{cite:363c292e19a8f0353b83bfc11d5ea4f0ff6557d6}}, {{cite:ac42d904058aed18c69d6ef1161963a3e6b1ad95}}, {{cite:870ccaca636495aadd97c78e310f8c17e6d6d1b5}}, {{cite:c16448ca0df1461084e5db7614c42eff8cd5b26c}}, {{cite:a11ca3cdf2519a1c15ac063a1ecc9f8308805af2}} of trajectory similarity computation rely on well defined similarity measures and associated acceleration techniques. Here, we focus on popular similarity measures, while a comprehensive coverage of free-space based similarity measures is available elsewhere {{cite:ff610136d3907a2ee6a8add341f7b1a9a5b91876}}, {{cite:738d9142a2946dcfc1610c403eb896f294b60ad7}}. Network-aware similarity computation techniques first map original trajectories into road-network paths that consist of vertices or segments. Then, they define similarity measures based on classic distance measures such as Hausdorff {{cite:cb001c0c1f4690ec5c3f5461c1a9fcd6f8d5405a}}, DTW {{cite:be8d05f0200730b6dc5da282fc4f034ce25add5e}}, LCSS {{cite:c4ee88fdf0a1bcf7e7b189b178d28a0a317e4dbe}}, and ERP {{cite:e0e7c4af0686684252ca693f077417594c527f97}}, generally by aggregating the distances between road vertices or segments of two trajectories. For example, Koide et al. {{cite:a11ca3cdf2519a1c15ac063a1ecc9f8308805af2}} propose NetERP by aggregating shortest path-distances between the vertices of two trajectories. Based on LCSS, Wang et al. {{cite:ac42d904058aed18c69d6ef1161963a3e6b1ad95}}, {{cite:870ccaca636495aadd97c78e310f8c17e6d6d1b5}} propose the Longest Overlapping Road Segments (LORS) for trajectory similarity computation and clustering. Similarly, Yuan et al. {{cite:c16448ca0df1461084e5db7614c42eff8cd5b26c}} propose the direction-aware Longest Common Road Segments (LCRS). These methods typically have quadratic computational complexity, as they rely on computations for aligned point pairs. Moreover, non-learning-based methods rely on hand-crafted heuristics, failing to exploit information hidden in trajectories.",
"Learning-based methods {{cite:5365f54a1b87c6ea5df93479de8097958346f521}}, {{cite:e86b6c7559f57d01d3afab77f8c67a473939fe99}}, {{cite:46afc1ee75b74ae6f6690d5d9b389c1186a879ac}}, {{cite:df77aeb2e25cd906dd40c68733705c7df7877979}}, {{cite:ae8767b89a0b963c85a8d9208a2e58cdf67617b2}} are becoming increasingly popular in recent years, as they feature the success of deep learning technologies, i.e., powerful approximation capability. The learning-based methods learn distance functions via neural networks that embed input trajectories and approximate given distance measures. This way, trajectory embeddings are generated that enable fast trajectory similarity computation and downstream analyses. Li et al. {{cite:5365f54a1b87c6ea5df93479de8097958346f521}} propose t2vec that addresses the high computational cost of traditional methods while taking into account low sampling rates and the influence of noisy points. Nevertheless, t2vec was designed for trajectory representation learning, not similarity computation. Yao et al. {{cite:e86b6c7559f57d01d3afab77f8c67a473939fe99}} propose NEUTRAJ, which employs metric learning to approximate trajectory similarity for different free-space based distance measures. Further, Zhang et al. {{cite:46afc1ee75b74ae6f6690d5d9b389c1186a879ac}} propose Traj2SimVec, which considers sub-trajectory similarity in the learning process. Zhang et al. {{cite:df77aeb2e25cd906dd40c68733705c7df7877979}} propose T3S, which utilizes attention function to improve the performance. Despite the efforts of these studies, they all target spatial trajectory similarity in free space and cannot model the complex dependence of road networks. Recently, Han et al. {{cite:ae8767b89a0b963c85a8d9208a2e58cdf67617b2}} develop GTS, which takes spatial trajectory similarity learning into road network context and achieves the state-of-the-art performance. Nevertheless, GTS is spatial-oriented similarity learning while ignoring the temporal aspect of trajectories. More specifically, GTS is designed for POI-based spatial trajectory similarity computation. For trajectories that share the same or neighbor POIs but with totally different traveling paths, GTS treats them as similar to each other. Besides, GTS only learns a single type of distance measure (i.e., TP {{cite:b3d3b0f01d8f6eac071f1d9c1ebd98eeb5bced62}}, an extension of Hausdorff distance), while ST2Vec accommodates a series of popular measures including TP, DITA, LCRS, and NetERP."
] | [] |
Data Efficient Language-supervised Zero-shot Recognition with Optimal
Transport Distillation | Traditional computer vision models are trained to predict a fixed set ofpredefined categories. Recently, natural language has been shown to be abroader and richer source of supervision that provides finer descriptions tovisual concepts than supervised "gold" labels. Previous works, such as CLIP,use InfoNCE loss to train a model to predict the pairing between images andtext captions. CLIP, however, is data hungry and requires more than 400Mimage-text pairs for training. The inefficiency can be partially attributed tothe fact that the image-text pairs are noisy. To address this, we propose OTTER(Optimal TransporT distillation for Efficient zero-shot Recognition), whichuses online entropic optimal transport to find a soft image-text match aslabels for contrastive learning. Based on pretrained image and text encoders,models trained with OTTER achieve strong performance with only 3M image textpairs. Compared with InfoNCE loss, label smoothing, and knowledge distillation,OTTER consistently outperforms these baselines in zero shot evaluation onGoogle Open Images (19,958 classes) and multi-labeled ImageNet 10K (10032classes) from Tencent ML-Images. Over 42 evaluations on 7 differentdataset/architecture settings x 6 metrics, OTTER outperforms (32) or ties (2)all baselines in 34 of them. | [
"Zero-Shot Learning in Computer Vision: Zero-shot learning (ZSL) studies the generalization of knowledge to unseen classes. Previous methods for zero-shot recognition in computer vision mainly follow three paradigms. The first type, including DeViSE {{cite:dd26c81f1f8bfda09d0a390381cd465fa1c1234f}} and ConSE {{cite:2beea6333c53ce69c1137dfbd171585515d66e2b}}, uses pretrained word embedding vectors to represent different categories and implicitly model their relationships.However, word embedding is a preliminary and limited representation of class relationships, which hurts performance. The second paradigm, including GCNZ {{cite:662d23e7b9e6e609dee7cb91803b4a94308e95c9}}, DPGZ {{cite:8fcbafe64c481c72ae54370998afe1184f5319c1}}, and HZSL {{cite:0d9ce6d0c2176a2f0a23b61c0ce4d670930ed15f}}, explicitly models class relationships as a graph, and uses a graph convolutional network (GCN), or a predefined class hierarchy, such as WordNet {{cite:02a203966f47f15fbee626566a62637ad6059d5e}}, to learn the knowledge propagation between classes.However, real-world class relationships are complicated and simple graph structures such as WordNet are too limited to model such relationships. Lastly, {{cite:d621c4abcd05a17f6909aea6ff0704f199de2f87}}, {{cite:42ffa4de58ea3249e64ba83c17bc0f43e4f67d47}}, {{cite:9c9b992470a21dbec848c53cb0ea0f5aa878f491}} rely on human-labeled attributes to model semantics of classes. The scalability of these methods are limited by the need for attribute annotations. More recently, CLIP {{cite:f5086c1ed4b0587dd07f9077225a5b4b698f8dd9}} applies language-supervision to ZSL by training on image caption pairs. Our work is based on CLIP and we generalize the InfoNCE loss to improve its data efficiency.",
"Vision and Language: Natural language has long been used as a source of supervision in fields like image-text retrieval {{cite:27eb10536ca1ba99c647df6936290a2516f2a2f3}}, object classification {{cite:812692bf38ab004b181c57f90b5eaf2610f16d9c}}, and video understanding {{cite:55a9e44532fd649eebb9ccebe0d387434abd75e2}}. {{cite:e5c00224e4b66f799ed41b328280950d94bf7557}}, {{cite:b77f92f6c7de9b6a8c6351339ced1fd8191a96b8}}, {{cite:791ac4b398e645028f32f6e4cbae038d4b506264}}, {{cite:d8b3a059a0ced6ed079f1ec4066691b91b3f8966}} have proposed methods of learning visual and language representations in a joint embedding space. More recently, {{cite:660ed5609b966887380bde4f7775a89abb701e63}}, {{cite:322873e54e938adfc3a83e826a6a7d6391252405}}, {{cite:53385fe3f0dd0545c42bcb642ed9d304f98171ed}} propose using a cross-modal attention mechanism to increase performance in image-text matching. In addition, {{cite:ec6dd95e32b644f02eb0a46e2523e0a9a67b238a}}, {{cite:bb977348d24aff0d919a0c3e4aa6c5488a662046}}, {{cite:a76670e1b96ae7fdcbd866091e0f88d3584685d9}}, {{cite:9a3584aac30e1c3a3a8244e2ecdc623216b9195d}} demonstrate that good visual representations can be learned by predicting image captions. To scale up vision-language joint training, CLIP {{cite:f5086c1ed4b0587dd07f9077225a5b4b698f8dd9}} and ALIGN {{cite:87f3bb3fba778a419e05402b8d640b9cf69f86d0}} both collect their own image-text datasets with 400M and 1B image-caption pairs.",
"Optimal transport (OT) is a theory that enables comparison of two probability distributions whose supports may not overlap. OT has been applied to many areas such as domain adaptation {{cite:3c93f49cd967c7818d76a55e09f0befeb3875448}}, generative models {{cite:ccc7a51c050652495dbbc7872ed32f7b8bc3b0d1}}, and self-supervised vision representation learning {{cite:d66934df05ab208430fcfd88b9a860001703d52b}}, {{cite:0e4645276d2750cefb6c42fbd90f60d47391afef}}.In vision and language, {{cite:01589fb238c6653dc1f59341167e2e8c4448c678}} uses OT to align objects in images and words in texts. The problem formulation of our work is similar to {{cite:21aa9478a1f1e5abe9601ee6da00c443713b6cac}}, where OT is used to mitigate the label noise in remote-sensing data under supervised learning. In our paper, we extend the method from supervised learning to contrastive learning, where OT is a natural way to estimate pairings between images and texts. In another related work, {{cite:d8698da7e45b0fb16fde1e093c046dd8a10ab152}} adds an additional OT-based Wasserstein loss to contrastive representation distillation {{cite:09783654d58d5ae1d12c0dda33a0d5de63aba0e8}}. The loss matches student representations to teacher representations in a batch. {{cite:d8698da7e45b0fb16fde1e093c046dd8a10ab152}} is different from our method since it directly minimizes the Wassertein loss between two models' representations, while our method uses OT to estimate the pairing probability and use the probability for knowledge distillation. Directly minimizing Wasserstein loss between image/text embeddings in our case will lead to collapsed representations, where models generate constant output regardless of inputs.",
"Other related works: Our work is also related to areas including learning with noisy labels, contrastive learning, and knowledge distillation. Specifically, our method uses OT to estimate the probability of unpaired image/text being the correct match for a given text/image. This is reminiscent to the problem of estimating label transition probability under noisy labels {{cite:b7b1d523e44609ab75d1e305712817fc25aa4001}}. Our method is based on contrastive learning, which is commonly used in self-supervised visual representation learning {{cite:ed53fd428e5e0e0dfd2aa2447fe8fc3103cf9228}}, {{cite:95eb051b7dc7ae2c5724584ebbcbf1c88a43f57e}}. For vision representation learning, {{cite:c197bc4085e0bb9aef6c1fa8497a7919537a2ac8}} argues that sampling hard negative pairs can improve learning efficiency. For language-supervised representation learning, however, it is important to mitigate the noise of widely spread hard negative samples, since positive image-text pairs are usually only loosely correlated. Our method is also an extension to knowledge distillation (KD) {{cite:6e4fca5e2b10393015246feac9c13a11fc02630f}}. Typical KD directly relies on a teacher model to directly generate a target distribution {{cite:8a942e7977098bb3a5600bb637187ab3e01b437a}}, {{cite:f7b1f5a3d3cabd67914d86ecbde162464b7426c7}}, {{cite:09671190e287a59ed007fbc05b27f8c3f5b6d607}}. Our method is differentsince our target distribution is computed by OT based on the pairwise similarity estimated by a teacher model. Experiments show that this works better for image-text contrastive learning."
] | [] |
Knowledge graph enhanced recommender system | Knowledge Graphs (KGs) have shown great success in recommendation. This isattributed to the rich attribute information contained in KG to improve itemand user representations as side information. However, existing knowledge-awaremethods leverage attribute information at a coarse-grained level both in itemand user side. In this paper, we proposed a novel attentive knowledge graphattribute network(AKGAN) to learn item attributes and user interests viaattribute information in KG. Technically, AKGAN adopts a heterogeneous graphneural network framework, which has a different design between the first layerand the latter layer. With one attribute placed in the corresponding range ofelement-wise positions, AKGAN employs a novel interest-aware attention network,which releases the limitation that the sum of attention weight is 1, to modelthe complexity and personality of user interests towards attributes.Experimental results on three benchmark datasets show the effectiveness andexplainability of AKGAN. | [
"Our work is highly related with the knowledge-aware recommendation, which can be grouped into three categories.",
"Embedding-based Methods {{cite:75b05944294faff5c286e0f4d731947581a97416}}, {{cite:a743073fbb5a7b2e827f7dc42b0329399f5714ba}}, {{cite:a4bd4ecefdc0fc5b050d493d80feccb1e6e24e68}}, {{cite:d05d9a729398e499818f798f8e04c19825384555}}, {{cite:9ab6510bb6c7fd1aa0f6fb206de6a2522318f2bf}}, {{cite:ec2a9e6e800c536434e56473611d110e63e18530}} hire KG embedding algorithm (i.e., TransE {{cite:80cb3543dc80065fb612d8ab0ea178906a21ee52}} and TransH {{cite:af7fdcc8b3b3269c5e44b0230f7b58e98cde4146}}) to model prior representations of item, which are used to guide the recommender model. For example, DKN {{cite:ec2a9e6e800c536434e56473611d110e63e18530}} learns knowledge-level embedding of entities in news content via TransD {{cite:a91fd0443b6e3d92bbfe984e18b0c4b8c59d7a45}} for news recommendation. KSR {{cite:d05d9a729398e499818f798f8e04c19825384555}} utilizes knowledge base information learned with TransE as attribute-level preference to enhance the sequential recommendation.",
"Path-based Methods {{cite:b8464ab21f52adabc9ad6c16c184099521fb43e8}}, {{cite:5bb3077a4f91856b497abc09504f815f27d0c756}}, {{cite:21694c787e4e46aa1693eb9bf585a0511c1dfa98}}, {{cite:029f52870f26329d64287de4cc58cd130e5a7977}}, {{cite:655fe48b19bd0492d79253f07ae55fe369168692}} usually predefines a path scheme (i.e., meta-path) and leverages path-level semantic similarities of entities to refine the representations of users and items. For example, MCRec {{cite:8614b85aa9174cbc9e10d39fe01f2a5b54a54b1a}} learns the explicit representations of meta-paths to depict the interaction context of user-item pairs.RKGE {{cite:39ebe98909e4e05f4e469442678203d9179c641d}} mines the path relation between user and item automatically and encodes the entire path using a recurrent network to predict user preference towards this item.",
"GNN-based Methods {{cite:f1970067e62caad6c894ae2b9958ac24e6d8f541}}, {{cite:c6f9ca2ee435d675a3b7607d0cdf0805d0380a85}}, {{cite:84a1e3278ebe8076e05730b751d87f59f3ca87bd}}, {{cite:f953f2e2d2b04d2c45a7859fdb97c23a281c2fd0}}, {{cite:5b3a75fd02e9b1a36d7ebc6f7c767be1b92cab06}}, {{cite:d290d0dd55247ad9239c2aa110325be37ba9cafb}} utilizes the message-passing mechanism in graph to aggregate high-order attribute informations into item representation for enhanced and explainable recommendation. KGAT{{cite:f1970067e62caad6c894ae2b9958ac24e6d8f541}} regards user-item interaction as a new relation added to KG and then employees attentive mechanism to propagate attribute information. IntentGC {{cite:5b3a75fd02e9b1a36d7ebc6f7c767be1b92cab06}} reconstructs user-to-user relationships and item-to-item relationships based on KG and proposes a novel graph convolutional network to aggregate the attribute information from neighbors. KGIN {{cite:d290d0dd55247ad9239c2aa110325be37ba9cafb}} learn user interest via an attentive combination of attributes and integrates relationalinformation from multi-hop paths to refine the representations."
] | [] |
Self-attention based anchor proposal for skeleton-based action
recognition | Skeleton sequences are widely used for action recognition task due to itslightweight and compact characteristics. Recent graph convolutional network(GCN) approaches have achieved great success for skeleton-based actionrecognition since its grateful modeling ability of non-Euclidean data. GCN isable to utilize the short-range joint dependencies while lack to directly modelthe distant joints relations that are vital to distinguishing various actions.Thus, many GCN approaches try to employ hierarchical mechanism to aggregatewider-range neighborhood information. We propose a novel self-attention basedskeleton-anchor proposal (SAP) module to comprehensively model the internalrelations of a human body for motion feature learning. The proposed SAP moduleaims to explore inherent relationship within human body using a tripletrepresentation via encoding high order angle information rather than the fixedpair-wise bone connection used in the existing hierarchical GCN approaches. ASelf-attention based anchor selection method is designed in the proposed SAPmodule for extracting the root point of encoding angular information. Bycoupling proposed SAP module with popular spatial-temporal graph neuralnetworks, e.g. MSG3D, it achieves new state-of-the-art accuracy on challengingbenchmark datasets. Further ablation study have shown the effectiveness of ourproposed SAP module, which is able to obviously improve the performance of manypopular skeleton-based action recognition methods. | [
"Action recognition based on skeleton data has received lots of attention due to its robustness against variation of background and compactness to the RGB based represent. Handcrafted features were used in early approaches, where features could be manually designed based on joint angles {{cite:50269aa86fa7cd1c6458d358a4b70bcb2257b53f}}, kinematic features{{cite:3f989f06b8aadf47e61dbe4cb873c6c60db2af12}}, trajectories {{cite:f975f679aaba44bef4ea29b21d7b62302407de01}} or their combinations {{cite:d61759ad5d16cf96a9240674d86e246b2396de5b}}. With the development of deep learning, Many CNN or RNN based data-driven methods that could automatically learn the action patterns have attracted much attention. For instance, the skeleton action could either be treated as motion images in CNN approaches {{cite:3bd31e6f69e13fe23783e00f14c2dde9c630701f}}, {{cite:f19dadddfd17c87ff664c321cecdeb4a9234bd35}} or modeled as sequences of coordinates in RNN approaches{{cite:f37709da2fb972afd4a498c8178dfea9dfffa7ef}}, {{cite:289fd597fd140eb871fa6d406fbfb76dac08df96}}.",
"Graph-based methods have drawn much attention in recent skeleton based action recognition methods due to its ability of modeling relationship between body joints. The first GCN model for skeleton based action recognition is ST-GCN proposed by {{cite:5ba2fd1c8e3c22a81af29bc1f481207f95a26784}}, where the skeleton is treated as a graph, with joints as nodes and bones as edges. Some data dependent methods have been developed to further capture the relationship between distant joints {{cite:59002d745250c55800b3c39b9bce2e6eb456b8fb}}, {{cite:7407c76dece9b5a134965342e1e4d2f2d0cb455d}}, {{cite:dbd7e52ae4214a3a7d7eb676b5883d451b443060}}. In addition, the multi-scale structural feature representation method have been developed via higher order polynomials of the skeleton adjacency matrix. For instance, a multiple-hop module is used to break the limitation of representational capacity ca caused by first-order approximation {{cite:3b16e471331083cc74226b146e782f27d5c65be1}}, {{cite:ff2d71cf29c608d1f06451cfe626a1a61ff4dbea}}, {{cite:bb8c691246083d5430e99d055f9d84628f575645}}. Inspired by {{cite:ff2d71cf29c608d1f06451cfe626a1a61ff4dbea}}, a sub-graph convolution cascaded by residual connection with enrich temporal receptive field is proposed by {{cite:f9da1c505ca574b5c6d34ae96776b174a02bacac}}. A combination approaches called Efficient GCN is designed by {{cite:e17ee174aae38992e836ae548c2f749f77813ca5}}. An multi-granular GCN based methods on temporal domain is designed by {{cite:8642d6c612b363cbba1ca4ad7aa60c7fc30b9fb4}}. Angle information extracted from manually specified joint groups is fused to GCN model in {{cite:e6f9480717f24de43fb5cd1b11aee28fd24817dc}}.{{figure:686eb6b1-fcbc-4b79-9725-66484aab64e5}}"
] | [] |
Hashing It Out: A Survey of Programmers' Cannabis Usage, Perception, and
Motivation | Cannabis is one of the most common mind-altering substances. It is used bothmedicinally and recreationally and is enmeshed in a complex and changing legallandscape. Anecdotal evidence suggests that some software developers may usecannabis to aid some programming tasks. At the same time, anti-drug policiesand tests remain common in many software engineering environments, sometimesleading to hiring shortages for certain jobs. Despite these connections, littleis actually known about the prevalence of, and motivation for, cannabis usewhile programming. In this paper, we report the results of the firstlarge-scale survey of cannabis use by programmers. We report findings about 803developers' (including 450 full-time programmers') cannabis usage prevalence,perceptions, and motivations. For example, we find that some programmers doregularly use cannabis while programming: 35% of our sample has triedprogramming while using cannabis, and 18% currently do so at least once amonth. Furthermore, this cannabis usage is primarily motivated by a perceivedenhancement to certain software development skills (such as brainstorming orgetting into a programming zone) rather than medicinal reasons (such as painrelief). Finally, we find that cannabis use while programming occurs at similarrates for programming employees, managers, and students despite differences incannabis perceptions and visibility. Our results have implications forprogramming job drug policies and motivate future research into cannabis usewhile programming. | [
"We provide context about the legality and uses of cannabis, focusing on cannabis use the United States as we scoped our study to US-based developers. However, we note that recent cannabis legality shifts in the US align with global trends. We close with a discussion of prior work at the intersection of programming and cannabis.",
"General Cannabis Background—In the past century, cannabis legality has shifted dramatically in the US. While many cannabis preparations were listed in the US pharmacopoeia in the early 20th century, cannabis was officially criminalized under the Controlled Substances Act in 1970. Since 1996, however, 36 states have legalized cannabis for medical purposes, and 17 for adult use. The policy whiplash has been largely driven by politics and popular opinion rather than science, with the re-emergence of cannabis occurring partially due to: 1) acknowledgement of the cruel excesses and ineffectiveness of the War on Drugs {{cite:0b5d8538a604cc0b00c3a3fc657732988b538051}}, {{cite:53ed30829d9e09fd45bc8305d39146d820397478}};2) increasing acknowledgement of cannabis's therapeutic benefits {{cite:dc35d39a5955a51a293c2d1f266aba3bddea5e13}};and 3) the ongoing opioid crisis, which has driven home the need for alternative, less harmful pain medicines {{cite:93d99aba359699d58f8a98ebba4b70a36a386dda}}.In conjunction with liberalizing policies, past-year cannabis use has increased among Americans 12 or older: 11% in 2002 to 17.5% in 2019 {{cite:b7f50bfde7845f2f889bf40b65f84230db55b7fa}}.",
"Cannabis is used for many reasons, including medical (e.g., pain, nausea) and recreational (e.g., social or perceptual enhancement, altered consciousness) purposes {{cite:ec938d4c28465a3b610e558922843a365b9ad5fc}}, {{cite:4d56af9b934d413d969f7d586e7c0bcc6d836a7b}}.This wide variety of uses is enabled by the pharmacopeia of compounds in cannabis, which include more than 100 active cannabis-derived compounds (cannabinoids) such as cannabidiol (CBD) and {{formula:831e82e6-74f5-4fc3-a862-da339994fa6f}} -9-tetrahydrocannabinol (THC) as well as terpenes and flavanoids, which are responsible for taste and odor {{cite:dc35d39a5955a51a293c2d1f266aba3bddea5e13}}.The most common medical reason for cannabis use is chronic pain, demonstrated by patient surveys {{cite:a1355a7b373825f636f2f30e885d0828cf595213}}, {{cite:57de4f03a0bcc2a3054d6815bbce6ca4da108939}}and registry data for medical cannabis patients {{cite:3397656c2e8b922042c1fda79faf7a6b18c7043a}}. Within broad categories of medical vs. recreational use, however, there is subtle shading of motivations. Lee et al. developed the comprehensive marijuana motives questionnaire, in which they identified factors associated with cannabis use, including enjoyment, coping, experimentation, altered perception, sleep/rest, and availability {{cite:4d56af9b934d413d969f7d586e7c0bcc6d836a7b}}. These factors are quite broad, and may encompass recreational, medical or other use patterns such as for creative work or performance enhancement.",
"Cannabis and Programming—Within this complex environment, cannabis interacts with software development. For instance, cannabis use in Silicon Valley appears to be high, with area dispensaries reporting that around 40% of their clientele are tech workers {{cite:3df3113764c92e0f35d051df7b65b0df51b2d55a}}. Similarly, a qualitative study of coding bootcamps identified “lots and lots of [weed]” as one key element of support {{cite:4048ef42b0a17b6c9eeef7b9f5fc109c7fcdb332}}.",
"No previous empirical studies have directly investigated the intersection of software and cannabis. Reports either combine cannabis with other illegal substances or include programmers only as a sub-population {{cite:319a8831c3e88bc286ceac674d4f83b4c065185f}}, {{cite:be52f790e5e4ecc448e36bb197fc1819822578b9}}. For instance, one study has a table of the percentage of workers in various fields reporting illegal drug use {{cite:319a8831c3e88bc286ceac674d4f83b4c065185f}}. Programmers appear in the table, with 10.4% reporting usage in the last year. However, this is the only mention of programmers, cannabis is not separated from other drugs, and the data is from the early 1990s, well before cannabis was legalized in numerous states. Thus, questions of cannabis prevalence, usage motivations, and culture in software engineering remain unanswered.",
"However, cannabis use has been studied in the context of creative problem solving, a key component of many software engineering tasks {{cite:bb59ecfeb3caf4624f6d1630676bd26e560c9bc6}}, {{cite:862313a79a5f2f451549b63e5f606fa4a6689df9}}, {{cite:42c376771fd92c6491e5bc6de961333199ac2674}}. Generally, cannabis is seen as a creativity enhancer: in one study, 54% of participants believed it increased their creativity {{cite:e834effe916dcc63a9a191b97e17058915a27b43}}. Similarly, Steve Jobs, who was regularly using cannabis when he founded Apple, stated that “the best way [he] would describe the effect of the marijuana ...is that it would make [him] relaxed and creative” {{cite:c6bfc6257e27865b8e86e79b6c305f3f08c980d6}}. From a cognitive perspective, the link between creativity and cannabis use is more tenuous. LaFrance et al. found that cannabis users were more creative than non-cannabis users, however, this difference may be explained by underlying personality {{cite:f08d1ac429d20405df3a0ffe30712178701c3d59}}. On the other hand, Kowal et al. found that highly potent cannabis can actually impair divergent thought, a type of creative thinking {{cite:e857ce319b4b4a99d43d044c376cf78dd8567190}}. These conflicting results linking cannabis and creative problem solving may motivate future observational studies of cannabis and programming.",
"Apart from creativity, cannabis has a range of other cognitive effects which could impact software development; long-term cannabis use can impair memory and attention control {{cite:32deb36169540f72bf46e3a622d669aef97f4050}}. However, even for general cognitive effects, the current scientific literature is often insufficient; evidence regarding cannabis's long-term effects on skills used in software engineering such as decision making, motor ability, and working memory is inconsistent (see Kroon et al. {{cite:32deb36169540f72bf46e3a622d669aef97f4050}} for a review of the current research on cannabis and cognition).",
"Programming and Other Mind-Altering Substances–While the intersection between programming and cannabis remains unexplored, there exists research into connections with other mind-altering substances such as alcohol and LSD. One study of IT professionals in India linked high levels of work-related stress to increased rates of alcohol abuse and depression {{cite:0b59788f886000e6cff118275b5d3df69e8a6216}}. Jarosz et al. linked alcohol and creativity, finding that intoxicated subjects completed more creative problem solving tasks than sober counterparts {{cite:097a8c8caa613813882aeb0cdca5f8e40a5b8f7c}}. As for LSD, psychedelics have historical cultural connections with software development {{cite:ec3bc4d2a20a3724f7a3faf4cd10de4bd653d9d3}}, {{cite:844509307339874ee0301c64042cde8d6694960f}}: LSD was commonly reported by many early PC and internet developers as a source of innovation and creativity {{cite:ec3bc4d2a20a3724f7a3faf4cd10de4bd653d9d3}}, a phenomenon likely connected to 1960s counterculture {{cite:844509307339874ee0301c64042cde8d6694960f}}. However, despite preliminary evidence that micro-dosing on psychedelics may increase creativity {{cite:c6abe62a3b6da89d5f6e2f868ba74893acc2e5cd}}, connections remain cultural and anecdotal rather than causal, further motivating future studies of mind-altering substances and programming."
] | [] |
Contrastive Vision-Language Pre-training with Limited Resources | Pioneering dual-encoder pre-training works (e.g., CLIP and ALIGN) haverevealed the potential of aligning multi-modal representations with contrastivelearning. However, these works require a tremendous amount of data andcomputational resources (e.g., billion-level web data and hundreds of GPUs),which prevent researchers with limited resources from reproduction and furtherexploration. To this end, we propose a stack of novel methods, whichsignificantly cut down the heavy resource dependency and allow us to conductdual-encoder multi-modal representation alignment with limited resources.Besides, we provide a reproducible baseline of competitive results, namelyZeroVL, with only 14M publicly accessible academic datasets and 8 V100 GPUs.Additionally, we collect 100M web data for pre-training, and achieve comparableor superior results than state-of-the-art methods, further proving theeffectiveness of our methods on large-scale data. We hope that this work willprovide useful data points and experience for future research in contrastivevision-language pre-training. Code is available athttps://github.com/zerovl/ZeroVL. | [
"Mainstream multi-modal pre-training methods are divided into two classes: single-encoder and dual-encoder.",
"Single-encoder methods {{cite:d7f0b80835559a245316ab64da641a733c97ecec}}, {{cite:adac3f92a309f9e918820de60f26ccb1643377f7}}, {{cite:cb9654b97f0c97a2da650ef3c8c83d0507946ec6}}, {{cite:0fced4b72131c7a5e5a333f07d3eb71fa144075a}}, {{cite:7b44ba7e1060d1756f1814bea5ac5221dce40d94}}, {{cite:38fb51881319d26dcfc606afc7a784500e37fb86}}, {{cite:9c8f3ddf7ad88a0a37eee2356ae756d59ddfa4ca}}, {{cite:4513f5a5e7e0236187f030dd19a464f2d32aa6b1}}, {{cite:44933a0296de919a6e4540adbfaffd256ff41735}}, {{cite:39d7d9d63aaebd88c1227001974b3bddb5c568bf}} typically map image and text representations with a shared encoder, , BERT{{cite:e487191c5f7427cd3d3048c01993dad9fad9ba9c}}. However, jointly mapping representations from multiple modals exponentially increases the computation complexity, and these methods could be impractical for industrial cross-modal retrieval tasks {{cite:fda979ad76cb7474c18639d7862197ecbf5add1b}}.",
"Dual-encoder methods {{cite:c7acdc4046167b921c421550a541656f5a70eb3d}}, {{cite:2a5cd5dd226df75e8b7d4c711a75dbd76087faa5}}, {{cite:ed86aafac46dd09b7081936c18394728328522d5}}, {{cite:0559229416bae3fae57ad42df5c5aaa9e94ab6c9}} are flexible and inherently suitable for cross-modal retrieval tasks. This paper closely relates to dual-encoder methods CLIP {{cite:c7acdc4046167b921c421550a541656f5a70eb3d}} and ALIGN {{cite:2a5cd5dd226df75e8b7d4c711a75dbd76087faa5}}, which have demonstrated the potential of aligning multi-modal representations with the contrastive loss. Although great improvements have been made, these works require a tremendous amount of data and computational resources, , billion-level web data and hundreds of GPUs. Such experimental environments will challenge reproductions and further explorations for researchers with limited data or computational resources. This paper aims to address this problem by providing a training guideline and a reproducible strong baseline with limited resources."
] | [] |
Point2Cyl: Reverse Engineering 3D Objects from Point Clouds to Extrusion
Cylinders | We propose Point2Cyl, a supervised network transforming a raw 3D point cloudto a set of extrusion cylinders. Reverse engineering from a raw geometry to aCAD model is an essential task to enable manipulation of the 3D data in shapeediting software and thus expand their usages in many downstream applications.Particularly, the form of CAD models having a sequence of extrusion cylinders-- a 2D sketch plus an extrusion axis and range -- and their booleancombinations is not only widely used in the CAD community/software but also hasgreat expressivity of shapes, compared to having limited types of primitives(e.g., planes, spheres, and cylinders). In this work, we introduce a neuralnetwork that solves the extrusion cylinder decomposition problem in ageometry-grounded way by first learning underlying geometric proxies.Precisely, our approach first predicts per-point segmentation, base/barrellabels and normals, then estimates for the underlying extrusion parameters indifferentiable and closed-form formulations. Our experiments show that ourapproach demonstrates the best performance on two recent CAD datasets, FusionGallery and DeepCAD, and we further showcase our approach on reverseengineering and editing. | [
"Primitive fitting to point clouds.In vision literature, the primitive fitting and object decomposition have been investigated for decades with diverse types of primitives.The simplest forms of primitives are planes, which have attracted significant attention as they are omnipresent in our environments {{cite:6df7342698e74c4e5e3647faec9d54be85d55364}}, {{cite:c6d8f273387d23012f9e776fb664bb4ff33d20c7}}, {{cite:0c20a1f239d45fd34d7b481afb568b7720872676}}, {{cite:52fce48cf12a6947f02094be37725b2a9b38e2f1}}, {{cite:5a9cb4aa960c2d01fab598ae1b1d46eea92cd733}}.More general types of primitives were also explored in the decomposition with RANSAC {{cite:7686a8e838459fcae17d76a53074424ab4b810ca}}, {{cite:da8d8b99841112f254a5bcfb2882013625bb506f}}, {{cite:a093d49a8ece8c8ec508edd331bd3e31411e44b5}} and region-growing {{cite:640144f75204793dbf4d2ab1ccd065057ff0d768}} approaches.The local Hough transform of Drost and Ilic {{cite:acf99a17ed93030a4a04ff2aef3c34715e4a92d2}} showed how the detection of primitives can be made more efficient by considering the local voting spaces. This idea is quickly extended to handle a wider range of primitives {{cite:c1657f9dad1c97e1f6524530df61e43da2aadd3d}}, {{cite:63b76a69bfae31539edbe0f653633629af3757e9}}. Conic sections are also of special interest as they allow for an infinite set of variations {{cite:067cc6a606b682ae8c8a938f20269cf0885fc207}}, {{cite:bfa21d5141c411f07d6b9b50d18088942238ed90}}, {{cite:feb75dfae93d85f149f26ea3bc9ea5b065f216f4}}, {{cite:7eb5afd475057a588260fb8385a7853214e402c7}}.",
"The proliferation of deep learning has steered researchers to data-driven frameworks. Li {{cite:5676ae0fcc7720de044c7f4b92bdf941a6c2457d}} is the first to use a supervised learning method for the primitive fitting, and Paschalidou {{cite:314db3640074ffd946ce55f57895b82aead55f89}} and Sharma {{cite:6c5777a713cbfaac7a2648dafa4a81da2334b754}} extended the idea to fit superquadrics and B-spline surfaces to a point cloud, respectively. However, all these methods assume a fixed set of primitives and none of them handle our Extrusion Cylinders in their fittings.",
"Reverse engineering in CAD world.In CAD and graphics communities,reverse engineering is a well studied problem with many traditional algorithms explored over the years {{cite:763f5f191eedf2c79a83003f28db9010c8ba34d5}}, {{cite:5e6b58aaaa35ccf4b9b551d2b24c647a425cdcf5}}. Benko {{cite:39dc0e6f5a2f78edcd3b335583b1590b6199db3c}} described a commonly used procedure for segmenting triangle meshes by analysing the pattern which the normal vectors form on the Gaussian sphere. This allows planes, cylinders, cones and doubly curves surfaces to be identified. Following segmentation, these primitives can be fittedusing geometric constraints {{cite:59c87ee85b097a4b8e42d8b78c5616aa7e706a83}}.",
"It is only with the recent availability of large B-Rep, sketch and construction sequence datasets {{cite:2de4895afd61f04e52a81ed16cd5fc9eb5989527}}, {{cite:c4e55f5ccd7fa72a091c7de5ef1921b7b9ad852f}}, {{cite:c19e7682bfe7a31b53b3bf8a41580235e479269a}}, {{cite:721ca3a53fac3cf2cff06c9153c0ec0a2e4c628f}} that the machine learning community has become interested in the generation of CAD data {{cite:6c5777a713cbfaac7a2648dafa4a81da2334b754}}, {{cite:2b4056bf80250a31d719d8682627d81c46320d8b}}, {{cite:9190d959b686970f2f138fef1d8505c3da78f7fc}}, {{cite:c4e55f5ccd7fa72a091c7de5ef1921b7b9ad852f}}, {{cite:7c92fc5bb31ec85134e633c856a3524f56459260}}, {{cite:c19e7682bfe7a31b53b3bf8a41580235e479269a}}, {{cite:c213d43e7a828000e26c33e047ce0027ec570627}}, {{cite:d7e0f86067480081f1558d1c21b011b377347736}}, {{cite:39c0b8aac8f6960908ba1318bd8c4fda0aa92bdc}}, {{cite:721ca3a53fac3cf2cff06c9153c0ec0a2e4c628f}}, {{cite:1794a55e0933f82166e9673799f8032cd8393896}}. Wang {{cite:2b4056bf80250a31d719d8682627d81c46320d8b}} showed results for identifying keypoints in a point cloud suitable for fitting with parametric curves. Sharma {{cite:6c5777a713cbfaac7a2648dafa4a81da2334b754}} decomposed a point cloud into patches suitable for fitting with parametric surfaces. In both cases the difficult task of combining individual curves or surface to build watertight solid models which can be worked with in CAD modelers is not addressed.Ganin {{cite:c213d43e7a828000e26c33e047ce0027ec570627}} use a transformer model to generate 2D sketches. While their model can be conditioned on 2D image data, this data must be a sequence raster images of the individual curves which build up the model – a sequential breakdown not available in the reverse engineering setting. Tian {{cite:bc102d3b9114bb1f6f4ec4c1c5e758ad7ace9f21}} showed how shapes can be interpreted as programs represented using an RNN and voxel based \"neural renderer\".",
"Most relevant to us is DeepCAD {{cite:721ca3a53fac3cf2cff06c9153c0ec0a2e4c628f}}, which unconditionally generated entire 3D CAD models using a transformer based method. While DeepCAD demonstrates excellent results when auto-encoding in the “program space\", it leaves the guidance of the generated shape using point clouds for “future applications\", which is what we address in this work."
] | [] |
Expedition: A System for the Unsupervised Learning of a Hierarchy of
Concepts | We present a system for bottom-up cumulative learning of myriad conceptscorresponding to meaningful character strings, and their part-related andprediction edges. The learning is self-supervised in that the conceptsdiscovered are used as predictors as well as targets of prediction. We devisean objective for segmenting with the learned concepts, derived from comparingto a baseline prediction system, that promotes making and using largerconcepts, which in turn allows for predicting larger spans of text, and wedescribe a simple technique to promote exploration, i.e. trying out newlygenerated concepts in the segmentation process. We motivate and explain alayering of the concepts, to help separate the (conditional) distributionslearnt among concepts. The layering of the concepts roughly corresponds to apart-whole concept hierarchy. With rudimentary segmentation and learningalgorithms, the system is promising in that it acquires many concepts (tens ofthousands in our small-scale experiments), and it learns to segment text well:when fed with English text with spaces removed, starting at the characterlevel, much of what is learned respects word or phrase boundaries, and overtime the average number of "bad" splits within segmentations, i.e. splitsinside words, decreases as larger concepts are discovered and the system learnswhen to use them during segmentation. We report on promising experiments whenthe input text is converted to binary and the system begins with only twoconcepts, "0" and "1". The system is transparent, in the sense that it is easyto tell what the concepts learned correspond to, and which ones are active in asegmentation, or how the system "sees" its input. We expect this framework tobe extensible and we discuss the current limitations and a number of directionsfor enhancing the learning and inference capabilities. | [
"This work is a continuation of our research on prediction games{{cite:95b4ca63064f8f6ed24ec468b4e8a8cec9b17eb4}}, {{cite:48461039dc7ba33e68945ac532cba7c4279cb81e}}, {{cite:141dcd0f862daa143cf8965db6e44633c9e83c3d}}, sharing the goal of learning ahierarchy of concepts in a cumulative unsupervised manner. Themotivations and philosophy behind the approach and relations to a fewgeneral learning tasks, e.g. distribution or density learning, arediscussed in that work {{cite:95b4ca63064f8f6ed24ec468b4e8a8cec9b17eb4}}. Previously the focus was primarilyon prediction and composition {{cite:48461039dc7ba33e68945ac532cba7c4279cb81e}}, and the need for a moresophisticated segmentation was later identified but left to futurework {{cite:141dcd0f862daa143cf8965db6e44633c9e83c3d}}. Segmentation there was a simple leftto right and greedy process: a largest concept (sequence of words ortokens) was extracted and the resulting segmentation (or tokenization)was used for self training of prediction weights. To handle morecomplex concept structures, it was deemed that a fairly sophisticatedsegmentation process would be needed {{cite:141dcd0f862daa143cf8965db6e44633c9e83c3d}}. A moreelaborate segmentation or inference is also likely needed when noise anduncertainty is increased, e.g. when portions of input are corrupted.Furthermore, an appropriate smooth objective would also probably berequired to guide the inference. For this work, therefore, we neededto both devise a segmentation algorithm and develop an effectiveobjective to guide the search. The objective had to promote usinglarger and possibly recently created (new) concepts, but balancedagainst other desiderata, such as the fit with other concepts deemedpresent in an episode. The reliance of the segmentation objective onprobabilities and the desire to get better probablities motivated theuse of multiple levels for concepts.",
"In addition to composition for building higher level concepts(conjunctions), we have posited that discovery of groupings(disjunctions) in concept structure to be important and useful too,and the two concept creation operations, various forms of disjunctionsand conjunctions, together would create a hierarchy of larger and/ormore abstract concepts {{cite:95b4ca63064f8f6ed24ec468b4e8a8cec9b17eb4}}. This work provides empiricalevidence that creating a composition hierarchy is a practicalpossibility. However, learning concepts that also involve grouping orabstractions of some kinds, remains a challenge and a major openproblem. We are proposing a kind of structure learning, each concepta separate structure, thus we seek to learn many structures, withoutexplicit supervision, and with much sharing of substructure. Thecurrent concept structures are simple. Currently, updates are notperformed with respect to a concept's internal structure: updates are(mostly) limited to simple (scalar) fields and a concept'sassociations, i.e. the prediction weights. The system learns immutableconcepts.In this implementation, a concept may update itsparts to contain more than one part-pair, which is a very limitedstructure update. In general structure learning is verychallenging (such as learning various subclasses of finite-statemachines, grammar induction, etc.){{cite:858ca817ef69888326bd87923e3ca8ecf58fa25b}}, {{cite:5caa9576675584f4c7d6cffa2bdf0a9056b8e000}}. It is an intriguingquestion how much learning of additional sophistication in structurecan be effectively supported in the approach presented.",
"The self-supervised learning in prediction games requires efficientlypredicting and learning prediction weights for a large and a growing set ofconcepts in an online manner, and our work on sparse index learning(or \"recall systems\") has focused on designing such algorithms andupdate techniques{{cite:ae8812b80ecc9e846c3ed7f17765ca43e3dd636d}}, {{cite:9a686dd5be1cd213fc70e511964aed16dc8b2b0c}}, {{cite:25c9a90ffbec874745d1470c73e1493af5f2e825}}, {{cite:ec985861d2f9fc0c51d962fb889a52dbf0ca747c}}, {{cite:e7a6d4509aefd9f265760b8de88c8ddbc4f5b370}}. Inparticular, in this work we have focused on simple EMA updates andsimple ways of deriving probabilities from the predictions.",
"The working of the Expedition system has similarities to n-gram modelsfor language modeling {{cite:eb4347bbf066c942c36851933f48c13bbaf8dec8}}, {{cite:d5653a72727fe4a7a8a95a9217deb6d4e2305277}}, in that explicit n-gramsare stored for predicting. A major difference is that we aim to learna hierarchical vocabulary, while the vocabulary is fixed and flat inn-gram models, and there is no segmentation task that is integratedwith learning in such systems, to the best of our knowledge. Theprediction in Expedition is also more flexible than common approachesbased on n-grams. Early work on segmenting text include{{cite:00a9a2fe1451e3fede50196edfca5377fc35c148}}, which used character entropy.",
"In computer vision, semantic segmentation refers to the task ofassigning classes to pixels, but the classes are often given{{cite:829a313e398a556789bc8704e534968c02dfc91e}}, {{cite:d11d7a6268cca85a7e7efb836997789658db345b}}. In unsupervised (image) segmentation,pixels are grouped together to form parts of possible objects orobject classes, but no labels are induced. There is much recent workin vision in particular that involve various types of self-supervisedlearning and proxy or 'pretext' tasks, utilizing various cues such asvisual depth, for feature learning{{cite:bf23b28b736833b6e37736dfccaf24ed0c28600c}}, {{cite:968f840782df9717348917fb2966e7800a7cd491}}, {{cite:754b14eddf2623b1807f7d944e6fb6c9756546d3}},e.g. using convolution neural networks to build a vocabulary of visualfeatures.We have referred to the representations learned as\"concepts\", vs. \"features\" (elevated their status), as they are usedboth as predictors as well as targets of prediction, and theycontain considerable internal structure, such as the parts andpart-of links, in addition to several scalar fields. In the audiodomain, coincidences have recently been used, in part for clusteringto find useful representations {{cite:d034f2a42b459d9e22ceabd052f93eb8957c5dde}}.",
"Neural networks are universal function approximators{{cite:2cd1f32257dd3e6b0c0e8ab4397c2488f1648b9c}}, and with advances of the past decades(diverse architectures, development and advancement ofbackpropagation), and ample data and computation,have become highly powerful for extracting diverse regularities.Following the success of neural networks in the vision and speechdomains {{cite:b942ab08d519cea95f4a4387fc0d38080df4bebd}}, {{cite:dfc2257b3894ccbe5e39b51372204336b334e3c9}}, large language models viadeep neural networks, using a number of techniques such as embeddings,prediction, and attention, have had substantial recent successes invarious diverse NLP and related problems{{cite:8934d8438d7f92a21f4bc6257ab7da1ef8857626}}, {{cite:267fd52fa63e49aadeda7421913a103a700d9497}}, {{cite:7b89944f9b3183f7f0c985773437f7488c20b985}}, {{cite:52e1104704ed1eee8469fe0191d55aaf5cef31de}}, {{cite:4c0e7bdb1f75f1691043cd8acea9dbc5cfb3580f}}. In much of existing work intext, the networks begin with an existing vocabulary and theembeddings of that vocabulary as input, and it is remarkable that muchpowerful learning is achieved without the need to nor the complexityof segmenting. The meanings of words and patterns can become highlydistributed, providing advantages in making connections among similarpatterns, but also potentially losing some structure andinterpretability. There is also work on character n-gram models basedon neural networks {{cite:836a33861ecad7e0df17c8e30c35f6762ff28592}}. Sparse mixture of experts(MOEs), that attempt to activate a small portion of the neural networkon a per example basis (conditional learning or gating), trained viabackpropagation, have had considerable success in further scaling andspeeding up of neural networks training and inference{{cite:6d6b4bddab85af4910f195c20c2a883e8e0f5a9a}}.",
"Concepts are on one hand foundational to human cognition{{cite:2ca8ea45e5a9c57f722ab1b225e608c48768a3b5}}, {{cite:a9e1fe95b9dbd19ab827db9f4c54f491072491ba}}, {{cite:d48dcf6ddd15743e0dc0ec48ced69ca06b461f7e}}, and are on the other hand\"maddeningly complex\"{{cite:2ca8ea45e5a9c57f722ab1b225e608c48768a3b5}}. Concepts are inter-related in diverse ways(part-whole, taxonomic, spatio-temporal, domain-specific, ...), or putanother way, concepts seem to enjoy rich \"content\" (or attributes, interms of other concepts). The nature of concepts and how they areacquired and adapted over time, along with their rich relations andflexible use, remains largely a mystery. There is ongoing debateamong researchers on the various aspects of conceptual development,for instance, whether concepts develop in a general to specificmanner, or vice versa, the relation between language development andconceptual change, and whether there are fundamental differencesbetween say perceptual or more concrete concepts (such as visualobjects, e.g. a dog) vs. seemingly more abstract concepts (e.g. animal){{cite:d48dcf6ddd15743e0dc0ec48ced69ca06b461f7e}}, {{cite:a9e1fe95b9dbd19ab827db9f4c54f491072491ba}}. Researchers have gone as far astheorizing that children acquire knowledge in a manner very similar tothe progress in science, forming models (theories) that are adapted orrevised with new experience and evidence (the so-called \"theorytheory\" {{cite:295d32870d15643874b99f84b6ea64862fb0b5ec}}). Prediction has been proposed as a primarydriver of much of human intelligence {{cite:0132ab02c28e26b73a82e40fe9cffd0a089e02fe}}. Our workalso took inspiration from the neuroidal model of the neocortex and inparticular random access tasks, i.e. tasks that may involve pairsor multiplicities of concepts (stored items), such as associating twoarbitrary concepts, from a large space of acquired concepts,after observing them in one or a few episodes{{cite:7b83887f62ef553318f27d324f7e4763d7d077ee}}.Similar to that work, we have also assumedconcepts are more \"programmable\", than the nodes in the commonneural network models. Considering the importance and utility ofconcepts for solving advanced information processing tasks, orsymbolic computation under uncertainty, and the complexity ofconceptual phenomena, a diversity of algorithms or (sub)systems,working together, is likely required {{cite:aeed099d3b5ba3dae11ccc52ecb85444d36166e8}}. It is amajor open question whether existing neural network techniques basedon backpropagation, which have now substantially advanced many machinelearning applications, can be extended (e.g. perhaps in a posthocmanner) to support concepts or provide a basis for reachinghuman-level cognition. The extensive research work on theinterpretability of the models learned {{cite:c9a8a3d7216c3846de08af79b543d6c14404f11f}}and the related issues of model robustness and brittleness(adversarial attacks){{cite:9cbc539275e7ed6907b60ca5a7845586cdcc7087}}, {{cite:602ebbe201066e81a71efdbecc9c659db32e761b}}, may also belinked to the major question of whether neural networks can learnexplicit discernible concepts with some robust internal structure."
] | [] |
D3PG: Dirichlet DDPG for Task Partitioning and Offloading with
Constrained Hybrid Action Space in Mobile Edge Computing | Mobile Edge Computing (MEC) has been regarded as a promising paradigm toreduce service latency for data processing in the Internet of Things, byprovisioning computing resources at the network edge. In this work, we jointlyoptimize the task partitioning and computational power allocation forcomputation offloading in a dynamic environment with multiple IoT devices andmultiple edge servers. We formulate the problem as a Markov decision processwith constrained hybrid action space, which cannot be well handled by existingdeep reinforcement learning (DRL) algorithms. Therefore, we develop a novelDeep Reinforcement Learning called Dirichlet Deep Deterministic Policy Gradient(D3PG), which is built on Deep Deterministic Policy Gradient (DDPG) to solvethe problem. The developed model can learn to solve multi-objectiveoptimization, including maximizing the number of tasks processed beforeexpiration and minimizing the energy cost and service latency.} Moreimportantly, D3PG can effectively deal with constrained distribution-continuoushybrid action space, where the distribution variables are for the taskpartitioning and offloading, while the continuous variables are forcomputational frequency control. Moreover, the D3PG can address many similarissues in MEC and general reinforcement learning problems. Extensive simulationresults show that the proposed D3PG outperforms the state-of-art methods. | [
"In the literature, many methods have been proposed for task scheduling and offloading. A joint computation offloading and system resource allocation for MEC has been formulated as a mixed-integer non-linear programming format in {{cite:6c591a7fd5e80aa4a24058474ffee9f3e76fa882}}. Then the authors transformed the non-linear programming to linear programming to reduce the complexity and tackle the challenges. Liang et al. {{cite:8f67023119ed4d7853a03d955c4fc3f01ffa6826}}, adopt linear-fractional programming (LFP), a generalization of linear programming (LP), and a greedy algorithm to optimize the offloading rate and energy consumption. The tasks can also be divided into sub-task processes and processed on local devices or edge servers simultaneously {{cite:8c601a770d5627a6d57378ba978e250065f56641}}, {{cite:edd5a7f560eb6badb2fdab0d640759f0abfe2167}} based on optimization. Similarly, Gao et al. {{cite:e754fe219ec446baeafa9f9cff0cc9676030b687}} propose a method to find the optimal ratio of task partition into two sub-tasks for edge servers and local IoTs. He et al. {{cite:cc6128076fd8e4cc9e41b83f4c9d02047d0e30e5}} propose an optimization method to partition deep learning inference tasks and offload the partitions to edge servers to find optimal-delay for computing resource allocation. Those standard optimization methods are relatively straightforward to develop and have addressed many optimization problems in task offloading. However, MEC network environments are far complex to describe with mathematical forms, and it is considerably challenging to extend conventional optimization to high-dimensional observations.",
"Machine learning and deep learning {{cite:dccbad188baf82ca80269e750c97c02e0cea5056}} models can learn from historical labeled data and predict future computational offloading so that the system can make plausible decisions for MEC. Shen et al. {{cite:1ab163618cdbab929a5de8e7d61be99676d1ed1d}} surveyed machine learning methods for resource slice and planning for the next-generation network. They have also summarized various machine learning and deep learning methods adopted in computational offloading and content offloading for MEC. Lyu et al. {{cite:d88d43e9f408a193986c8f35565c59ededca2566}} use stochastic gradient descent, a popular machine learning training method, to learn and partition data to offload spatially distributed edge servers; the authors argue that the proposed method can make optimal decisions for data partitioning with respect to time delay. Yang et al. {{cite:ed2d996aaeb28dfb6daf61f61e800906ce7966c8}} proposed a statistical machine learning method to minimize the energy consumption for edge inference {{cite:c8789483e8fbd5c38f1aab87ca98d4f0beb2b1a2}}, {{cite:e97de2eeb4fa23786eb96d52084a80b10970fd3b}}, where deep learning inference tasks are processed on MEC. Ale et al. {{cite:7ee11c525b45a1ed92d942dd13e8f45e641d28ad}} introduced a deep recurrent neural network to capture and predict the user requests so that they can make decisions for the content offloading and allocate resources based on the prediction. Summarily, the computation offloading problems can be formulated as a supervised classification problem and minimize the cost using deep learning {{cite:8983160b2e1e4f7797147cb591eb3c1ecfbaa2e1}}. The deep learning models can also adopt bandwidth allocation optimization and maximize the system utility {{cite:2638e3d593119fdf3aad44ece945d6e352005faf}}. Machine learning and deep learning methods can learn and predict the offloading decisions optimized with respect to exploiting resources and reducing the cost for the MEC. However, machine learning and deep learning require training data with labels, which require enormous effort to collect and label data. Moreover, it is challenging for humans to label data from optimization problems because it is difficult to take optimal actions in such a complex system.",
"To mitigate the above issues, reinforcement learning {{cite:2e2f2863829b3ea5e185cf80002410e24e8966e8}} and Deep Reinforcement Learning {{cite:bf4c650a4ef52b34bd810b6178546cae077383ac}}, {{cite:2bc0bfa1203181128f43db9f47d4eef4ab64c4e3}} methods are extensively adopted for resource planning and optimization in MEC. In the reinforcement learning framework, we do not need to provide labeled data to train the models; instead, the learning agents interact with the environment (i.e., the MEC networks) to learn and find the optimal policies with respect to the objective function (e.g., minimize energy consumption). A Q-learning (a typical reinforcement learning algorithm) based method has been proposed {{cite:30e956a1fcc6669a6158e8fb2b9ae79f271cc939}} to make decisions for task offloading; the agents learn to decide whether the current tasks are offloaded to edge servers or processed on local devices to minimize the delay time. Similarly, deep Q-learning has been adopted to decide task offloading and select targeted edge servers for smart vehicles {{cite:a75f9cc9f20b30ee9828e8452bfe5bdac21645c9}}. Cheng et al. {{cite:077fb830dbf867b7b225723ce2e36fb99b5526b6}} adopted the DRL method to minimize the time delay for computation offloading to Unmanned Aerial Vehicle (UAV) based edge servers and the DRL model outperforms brute-force and the greedy algorithm. Similarly, Baek et al. {{cite:1825cf40ff01e421b505b101aaa1d1cd416a6d52}} proposed a Deep Q-Network (DQN) {{cite:bf4c650a4ef52b34bd810b6178546cae077383ac}} variant by replacing the convolutional neural network with a recurrent neural network to control and select task offloading to edge servers; the offloading actions generated by policies to exploit the limited MEC resources and maximize number of tasks been processed before expiring. Another DQN variant model {{cite:ec98c393a21543f95bce1b48260161dd36c3e4fc}} is proposed to resource allocation optimization by incorporating Bayesian learning and Long Short-Term Memory (LSTM) {{cite:c316ef1b7816b0ca7c47accf62ed2a1158406984}}. In {{cite:41f17411419c3c7c8c244a77d0f4bec6bf82b3c0}}, the authors have adopted DRL models to optimize task partitioning and scheduling for vehicular networks that allow two edge servers to process a task collaboratively. Yu et al. {{cite:0b839cae0d7cd4c1ce9d14da597d2fba5a0dbac8}} proposed a DRL model to optimize the task partitioning and offloading; the model makes decisions based on the profiles of sub-tasks and chooses the local devices or edge servers to process the sequentially depended sub-tasks.",
"Although reinforcement learning and DRL methods are adopted to address many task offloading and resource allocation problems in MEC, the currently existing methods can only deal with relatively small action spaces and are inflexible to partition large tasks. For example, the most of the DRL models for MEC can only take binary action that decides a task to offload to the MEC server or process on local devices. Another type of DRL model can find an optimal proportional (percentage) task to offload to edge servers to process. In addition, the majority of reinforcement learning models deal with discrete action {{cite:bf4c650a4ef52b34bd810b6178546cae077383ac}} or continuous action space {{cite:10910dbc2d97b36557b59f67e4bbdf2f7dfd3d3e}}. Several works try to handle hybrid action with approximation or relaxation of continuous action space {{cite:1507ed01b77196571c64c4bd1d9c8b7018ebc441}}. Hausknecht et al. {{cite:23b307372a9794977eb20f478e7f1662f7c7fdf2}} relaxed the action space to have to support the hybrid action. Masson et al. {{cite:a5e0addfb043376d50ef036eff971dddc01a9229}} handled discrete action with Q-learning and policy search for continuous action. Similarly, Khamassi et al. {{cite:02f64141a6b150104a38a394f163b34c9dc12dd9}} use Q-learning and policy gradient to achieve the same results. Those methods assume on-policy and handle discrete and continuous actions separately. Xiong et al. {{cite:911570c478b8b1523db86d558b409d55081f2e33}} and Fu et al. {{cite:e19b2b5a55c1d010e98fae939b738387e1f89732}} use a hierarchical structure {{cite:25d8953b0e8aea485911b2356aa114181ee052f8}} to deal with the discrete actions and generate the continuous action based on the discrete actions. Neunert et al. {{cite:208686718814a85cadb0399e6b9e58485b95e451}} provided mixed policy to handle the discrete-continuous action space. However, none of those methods can handle constrained action space well in edge computing."
] | [] |
Balancing Fairness and Robustness via Partial Invariance | The Invariant Risk Minimization (IRM) framework aims to learn invariantfeatures from a set of environments for solving the out-of-distribution (OOD)generalization problem. The underlying assumption is that the causal componentsof the data generating distributions remain constant across the environments oralternately, the data "overlaps" across environments to find meaningfulinvariant features. Consequently, when the "overlap" assumption does not hold,the set of truly invariant features may not be sufficient for optimalprediction performance. Such cases arise naturally in networked settings andhierarchical data-generating models, wherein the IRM performance becomessuboptimal. To mitigate this failure case, we argue for a partial invarianceframework. The key idea is to introduce flexibility into the IRM framework bypartitioning the environments based on hierarchical differences, whileenforcing invariance locally within the partitions. We motivate this frameworkin classification settings with causal distribution shifts across environments.Our results show the capability of the partial invariant risk minimization toalleviate the trade-off between fairness and risk in certain settings. | [
"Many recent approaches aim to learn deep invariant feature representations: some focus on domain adaptation by finding a representation whose distribution is invariant across source and targetdistributions {{cite:cc76e6cf3ab9da1611767d0087aa1ce80533df4f}}, {{cite:b89d6815134c140dd3e3b88de507c8fdfac6314f}}, while others focus on conditional domain-invariance {{cite:86864469e188a5d5e28f68ccac5ce0234a80bd1e}}, {{cite:6ec0e758c8cc6cf00616bc64e5b04f4909c8d440}}. However, there is evidence that domain adaption approaches are insufficient when the test distribution is unknown and may lie outside the convex hull of training distributions {{cite:8407e5f349b6b1a273ad8ab804ed4ebc1287df0f}}, {{cite:9b1919d95814c109bc8c735939757cb2a50e1c68}}, {{cite:4300d83e6cbff5e98345fa6ec1452b1cd2a94742}}.",
"The focus of this work is the Invariant Risk Minimization (IRM) framework, proposed in {{cite:fe3ad62dc9ed6a086cb72534cd85db44598bd61f}}. The framework relates to domain generalization where access to the test distribution is not assumed. It is rooted in the theory of causality {{cite:09aca2e404ebb8eb7d48c9990d08ca67bdd9386a}} and links to invariance, which is useful for generalization {{cite:e56865558951d85e7f83303bf0262fe5e53c328f}}, {{cite:fa92c842103810d5bbb7b9503f0f5431f62e8f69}}. In {{cite:e9e9a6b9a06d1871b9523c7e040612d4a27ca0fc}}, the authors reformulate IRM via a game-theoretic approach, wherein the invariant representation corresponds to the Nash equilibrium of a game. While the IRM framework assumes only the invariance of the conditional expectation of the label given the representation, some follow-up works rely on stronger invariance assumptions {{cite:68b6564adddfaa3c783e1366e423799750ba3404}}, {{cite:e99f4eccb1d6502069703b67788ef67e239279ba}}.",
"Several follow-up works comment on the performance of the invariant risk framework, and compare against standard Empirical Risk Minimization (ERM). It has been noted that carefully tuned ERM outperforms state-of-the-art domain generalization approaches, including IRM, across multiple benchmarks {{cite:4647fbe487761f9b55b2ed09ba9a3b91f1f9b6a0}}. The failure of IRM may stem from the gap between the proposed framework and its practical “linear” version (IRMv1), which fails to capture natural invariances {{cite:8ca45aa37154df64e5ec4fe63b682b4f212dce0b}}. The authors of {{cite:2b87f9075fc6bcf9192bd211fed6cb1e22f3ab00}} demonstrate that a near-optimal solution to the IRMv1 objective, which nearly matches IRM on training environments, does no better than ERM on environments that differ significantly from the training distribution. The authors of {{cite:758b19586c134904a1ce1e1f23b8f030056c0e69}} argue that IRM has an advantage over ERM only when the support of the different environment distributions have a significant “overlap”. However, unlike previous work which aim to minimize OOD risk, our primary focus is to understand the trade-off between risk minimization and fairness across environments in the setting when such a significant “overlap” may not exist. Consequently, we analyze how IRM behaves in these settings and explore alternatives via the notion of partial invariance. The notion of fairness considered in this paper relates to group sufficiency {{cite:60da019dfb7511b776ecf7fa7c863587c1f76e3a}}, which is intricately linked with the IRM objective. In fact, several works build upon the IRM framework by explicitly minimizing a quantity resembling this notion of fairness for achieving invariance {{cite:4f99257d883586f252daa2b6d2b0b04ef313484f}}, {{cite:a9445f405e9191a608dcb86465374973e6bc460a}}."
] | [] |
Privacy preserving n-party scalar product protocol | Privacy-preserving machine learning enables the training of models ondecentralized datasets without the need to reveal the data, both on horizontaland vertically partitioned data. However, it relies on specialized techniquesand algorithms to perform the necessary computations. The privacy preservingscalar product protocol, which enables the dot product of vectors withoutrevealing them, is one popular example for its versatility. Unfortunately, thesolutions currently proposed in the literature focus mainly on two-partyscenarios, even though scenarios with a higher number of data parties arebecoming more relevant. For example when performing analyses that requirecounting the number of samples which fulfill certain criteria defined acrossvarious sites, such as calculating the information gain at a node in a decisiontree. In this paper we propose a generalization of the protocol for anarbitrary number of parties, based on an existing two-party method. Ourproposed solution relies on a recursive resolution of smaller scalar products.After describing our proposed method, we discuss potential scalability issues.Finally, we describe the privacy guarantees and identify any concerns, as wellas comparing the proposed method to the original solution in this aspect. | [
"Federated learning is a field that has recently grown in prominence due to increasing awareness of data privacy issues and data ownership as well as the rising need to combine data originating from different sources {{cite:512c9a3546214eff9e77bd544d5fef81ecadeb65}}. It is a thriving research field that promises to make it possible to apply machine learning algorithms (or any other data analysis) on multiple decentralized datasets in a collaborative manner {{cite:030794c9e6ae42390cc4dee1d32f71a6ee96fd0a}}. This applies to both horizontally and vertically split data. Horizontally partitioned data is the situation where different organizations collect the same information from different individuals (e.g. the same clinical data collected in multiple hospitals). Vertically partitioned data is when different organizations collect different information about the same individuals (e.g. insurance claims and hospital records).",
"In order to apply machine learning algorithms on decentralized data, various techniques have been proposed to run the necessary analyses in a privacy-preserving manner. The techniques for vertically partitioned data are generally referred to with the umbrella term of secure multiparty computation (SMPC) {{cite:30f488b56a2a1173013c37040f1be4286b908e1f}}. SMPC is a research field that focuses on developing methods to calculate functions on decentralized data without revealing the data to other parties.",
"Examples of the various proposed techniques are methods to build different types of machine learning models, such as Bayesian networks {{cite:7ec47edcd8ffbb5649749aeca889f9df63c34c28}}, neural networks {{cite:53c02fbabd5a46afd2ad2e23a4e87472fa4696ee}}, or random forests {{cite:1a7013a3675ba579981956cc4896220b1cf7f52d}}. These techniques may rely on concepts such as secret sharing {{cite:042da8e39df62d7259f62368c2b2a00cc6034c8a}} and homomorphic encryption {{cite:5a9382e6f22e5aaf07a68c8dba24dd29e540daca}}. Both secret sharing and homomorphic encryption work at their core by transforming the original values {{formula:851adf2e-5cc7-4bea-8a9b-663f6eedddc5}} and {{formula:470a7ba4-9c81-438e-8eb9-759866c632c0}} , owned by different parties, into transformed values {{formula:5d694bcd-5952-4906-828e-7122a6bd0de8}} and {{formula:7c4c8996-e4dd-4cf9-97ef-b92142351dae}} such that {{formula:58edd02f-9e3b-455b-babd-065c9d9f2d8b}} , thus making it possible to calculate the result of function {{formula:f66cf5f6-8fa8-480d-bcef-81a72f40c0f3}} by calculating a different function {{formula:28c1e47b-ecae-4b5c-8ea0-f839eff4b375}} without ever needing to reveal {{formula:5b8deaa0-8173-4abc-a55b-afd8beaa36c1}} or {{formula:4d9019ba-e070-4484-99b2-43f6d40224d2}} to different parties. In the case of homomorphic encryption, this is achieved by using encryption schemes that are ‘homomorphic’ with respect to specific functions, allowing the user to do calculate these functions using encrypted data. In the case of secret sharing, the core concept relies on obfuscating the raw data with a secret share (e.g., a random number), and then applying calculations to the obfuscated data in such a way that the secret shares will cancel out in the end.",
"Other techniques focus on specific calculations that can be used as building blocks for machine learning algorithms, such as the scalar product (or dot product) of vectors. The scalar product is an integral part of various machine learning algorithms, such as neural network training {{cite:43693809e07260b506394638e64a4cbe1eb72c0a}}. Therefore, secure scalar product protocols have been widely studied in federated learning {{cite:88e1256033c23e5281a1aae6e29cde2582a90a1e}}. In addition, it can also be used in combination with clever data representations to calculate various statistical measures in a privacy preserving manner, such as the information gain of an attribute, as well as to classify an individual using a decision tree in a federated setting {{cite:88e1256033c23e5281a1aae6e29cde2582a90a1e}}. More generally speaking the scalar product protocol can be employed to determine the size of a subset of the population that fulfills a set of criteria in a privacy preserving manner, even if the relevant attributes are spread across multiple data owners. Hence any calculation that can be represented as a set-inclusion problem between two or more data owners can be solved using this method.",
"Because of its importance, multiple scalar product variants have been proposed. Du and Atallah proposed several methods for the scalar product {{cite:616890c1f4958b08a80362a15a4e77f12d336617}}, {{cite:2f9f75439946233b52b902e78e9f6ede238ebbf7}}. Vaidya and Clifton {{cite:ba0ce83846b8a2c7acb423c75ce764c5c80eac09}} proposed a new method to alleviate the scalability issues of existing methods and used this method to determine globally valid association rules. Du and Zhan {{cite:88e1256033c23e5281a1aae6e29cde2582a90a1e}} proposed yet another alternative, with better time complexity than the method proposed by Vaidya and Clifton{{cite:ba0ce83846b8a2c7acb423c75ce764c5c80eac09}}, and better communication cost than the methods proposed by Du and Atallah{{cite:616890c1f4958b08a80362a15a4e77f12d336617}}, {{cite:2f9f75439946233b52b902e78e9f6ede238ebbf7}}. Du and Zhan {{cite:88e1256033c23e5281a1aae6e29cde2582a90a1e}} then used it to train a decision tree in a federated setting. Goethals et al. {{cite:c65b781d4a3c26d08d378f25eedb210a9f4c6894}} discovered certain privacy flaws in some of the earlier mentioned protocols, and suggested an alternative with improved privacy guarantees.",
"However, all these solutions focus on two party scenarios. Translating them to scenarios involving more than two parties is not straightforward, when at all possible. This is a significant drawback since often three, or even more parties can be involved.",
"In this paper, we look at one of the method proposed by {{cite:88e1256033c23e5281a1aae6e29cde2582a90a1e}} and determine if and how it can be scaled to an arbitrary number of parties. This has applications for the various calculations which can (partially) be transformed into a scalar product problem mentioned before, such as calculating information gain or anything else that can be represented as a set-inclusion problem."
] | [] |
Watermarking Images in Self-Supervised Latent Spaces | We revisit watermarking techniques based on pre-trained deep networks, in thelight of self-supervised approaches. We present a way to embed both marks andbinary messages into their latent spaces, leveraging data augmentation atmarking time. Our method can operate at any resolution and creates watermarksrobust to a broad range of transformations (rotations, crops, JPEG, contrast,etc). It significantly outperforms the previous zero-bit methods, and itsperformance on multi-bit watermarking is on par with state-of-the-artencoder-decoder architectures trained end-to-end for watermarking. The code isavailable at github.com/facebookresearch/ssl_watermarking | [
"Image watermarking {{cite:12f20e88a2ecbada51634357cee8c496f2eafded}}approaches are often classified by their embedding space: few use the spatial domain{{cite:0ecb35df4654492bbc54027c85a5c9d27d6ea534}}, {{cite:4d0573958afdcbfa5349ae4a52b68da60685cfa0}}, most of them follow the TEmIt principlewith a well known transformation like DFT {{cite:802dbfa167d11c115229417dfb9f202cccbd77a1}}, DCT {{cite:b4550dd8ac73a52ba3b09ce910baab1377360d65}} or DWT {{cite:e2ed9a2bb70d8ee266f9003dd72cb02752f793f9}}.In zero-bit watermarking {{cite:9d36733dce6791eb7300e7e9a0ac52a094e267b1}}, the embedder only hides a mark and the detector checks for its presence in the content.For instance, the received content is deemed watermarked if its descriptor lies in a hypercone of the feature space.This detection strategy is near optimal {{cite:96564b0abab989eb11337ccda5e4a745fd88dad6}}, {{cite:02ca1fca0905f1ed5b02081c087561711857dbfa}}.In multi-bit watermarking, the embedder encodes a binary word into a signal that is hidden in the image. At the other end, the decoder retrieves the hidden message bit by bit.This is the case for most deep-learning-based methods presented below.",
"Deep-learning-based watermarkinghas emerged as a viable alternative to traditional methods. The networks are often built as encoder-decoder architectures {{cite:4c21310d0c2383a65a2cb6cee9fdf08e7d72f6e5}}, {{cite:c5733dc46481f3936b411d7cc068e5a20436bc4b}}, where an encoder embeds a message in the image and a decoder tries to extract it. For instance, HiDDeN {{cite:4c21310d0c2383a65a2cb6cee9fdf08e7d72f6e5}} jointly trains encoder and decoder networks with noise layers that simulate image perturbations. It uses an adversarial discriminator to improve visual quality, and was extended to arbitrary image resolutions and message lengths by Lee et al. {{cite:834f21da8d020b4d0fa1fde3c81371ced1395213}}.Distortion Agnostic {{cite:00aaf4e97a6959dd97c3ab530fb97889081a2415}} adds adversarial training that brings robustness to unknown transformations and {{cite:30f4b0e668b3f68d3dc4f76e03cfa559fd469fe1}}, {{cite:e85f938417473a32f5f69eb6826956cd25af9a13}} embed the mark with an attention filter further improving imperceptibility.ReDMark {{cite:6f5ed53d969d573725166c67690bd446e729d528}} adds a circular convolutional layer that diffuses the watermark signal all over the image.Finally, ROMark {{cite:e4aac5cc8cdb489e547c99f85ac9be7c0776dd71}} uses robust optimization with worst-case attack as if an adversary were trying to remove the mark.For more details, we refer to the review {{cite:c6f29e256bb47eece4db4bde963f6b2e1f96d2bb}}.This type of methods has its drawbacks: e.g. it is hard to control the embedding distortion and it is made for decoding and does not translate so well to detection.",
"Vukotić et al. {{cite:88f34ef90b3638aff2468025619b007978bd68da}} mark images with a neural network pre-trained on supervised classification instead of an encoder-decoder architecture.The network plays the role of the transform in a TEmIt approach.Since it has no explicit inverse, a gradient descent to the image pixels \"pushes\" the image feature vector into a hypercone.The follow-up work {{cite:8b6562b4e33ac9e40ac721277f13c48615a77ea8}} increases the inherent robustness of the network by applying increasingly harder data augmentation at pre-training.It offers a guarantee on the false positive rate without requiring to train a network explicitly for watermarking, but no multi-bit version was proposed."
] | [] |
Transcribing Natural Languages for The Deaf via Neural Editing Programs | This work studies the task of glossification, of which the aim is to emtranscribe natural spoken language sentences for the Deaf (hard-of-hearing)community to ordered sign language glosses. Previous sequence-to-sequencelanguage models trained with paired sentence-gloss data often fail to capturethe rich connections between the two distinct languages, leading tounsatisfactory transcriptions. We observe that despite different grammars,glosses effectively simplify sentences for the ease of deaf communication,while sharing a large portion of vocabulary with sentences. This has motivatedus to implement glossification by executing a collection of editing actions,e.g. word addition, deletion, and copying, called editing programs, on theirnatural spoken language counterparts. Specifically, we design a new neuralagent that learns to synthesize and execute editing programs, conditioned onsentence contexts and partial editing results. The agent is trained to imitateminimal editing programs, while exploring more widely the program space viapolicy gradients to optimize sequence-wise transcription quality. Results showthat our approach outperforms previous glossification models by a large margin. | [
"Sign Language Recognition, Translation and Production. Most research works in automated sign language interpretation {{cite:3d83d3333a7f18c516178455a60ea6d95731c23f}} aim at recognizing {{cite:b063f6ee9dad9919681d72e9174f1319d5ddc604}}, {{cite:821d17bb185af725b725ff5077409044188ce6db}}, {{cite:67ae7d6860670121ba932c5bdc64ab7c5e48f1fa}}, {{cite:5635be645b017c37d131247fd4faa5e65b4cc5f2}} and translating {{cite:891b29aa8898b56f45d620a004b3a957f8403493}}, {{cite:e58fc21915b43ee189f2c92757639b14ae40489f}}, {{cite:35721e9804e55563e8c01069a1af109ea700fb88}} visual sign gestures into sentences.However, to facilitate two-way deaf-hearing communication, it is also necessary to translate spoken sentences to sign gestures in videos or animations for the deaf {{cite:0e13dce903d01d8f9d641c2ee2ae9743c966f58d}}.In this regard, the recent work {{cite:e9da800b7b3e451f54b10aeb76d5a65457edee3b}} first translates sentences to glosses, which are later used to produce continuous sign language videos {{cite:f5f0a8f75ec268c546bdd70009368c27716b8fee}}.Our work follows this pipeline while for the first time, formalizing glossification as a standalone sign language interpretation task.Different from {{cite:e9da800b7b3e451f54b10aeb76d5a65457edee3b}} that adapts a neural machine translation approach, we propose to use editing labels to bridge the gap between two linguistic sources, achieving superior glossification results.",
"Neural Program Synthesis. Program synthesis techniques aim to generate programs that satisfy given specifications {{cite:b3ab137bb379ec0220a8b3ee4a0ee89675f31d36}}, {{cite:21c76f9427b1c6342881d76738b2c176cc36eeff}}, either in natural languages or as a set of example inputs and desired outputs.The advantage of neural programs is their flexibility in modeling compositional structures in textual and visual data, thus are widely applied in various domains, including string manipulation {{cite:a139131997a53151ab0e3e85c4b918e431a0a873}}, sentence simplification {{cite:b224bfab0100d82c16615ba4e9f39d4612ab6901}}, semantic parsing {{cite:b9385b4e7b8c273f32676b7b736f093d0644966b}} and shape generation {{cite:3e77a5b9f2b2c5fb746c74fe992b24979f81b669}}, {{cite:bd8f42561cba0b42882e1601d52037d1419aaecc}}.Inspired by these works, we use editing programs to exploit syntactic connections between sentences and glosses.Different from previous models that directly predict glosses {{cite:e9da800b7b3e451f54b10aeb76d5a65457edee3b}}, we instead obtain glosses as the result of executing editing programs on their sentence counterparts.In this way, our model better utilizes relations between sentences and glosses by explicitly copying or removing words.In addition, editing programs define each glossification step as an action, thereby, they are easier to interpret than results from black-box sequence-to-sequence machine translation models.{{table:ffad5057-7e05-41a0-a0d3-eeded15e4e51}}",
"Learning Symbolic Operations on Text Sequences. There are also prior approaches relying on predicting symbolic operations for other sequence learning problems.In particular, the dependency parsing models proposed in {{cite:26def6a94adf60c6d493366f3f1b82754916f53d}}, {{cite:90b3ef9e0d84191564b342e7795ea493e4b7b6a8}} predict transitions between initial and terminal configurations, which are then used to derive a target dependency parse tree.The models in {{cite:cd6163af1a5494e387fbd929d8890032829a14b6}}, {{cite:b224bfab0100d82c16615ba4e9f39d4612ab6901}} predict simplification operations to transform complex sentences into simple ones.Inspired by these works, we design our program synthesis model that takes natural language sentences as input and learns to perform editing operations on them to obtain glosses."
] | [] |
Autonomous Reinforcement Learning: Formalism and Benchmarking | Reinforcement learning (RL) provides a naturalistic framing for learningthrough trial and error, which is appealing both because of its simplicity andeffectiveness and because of its resemblance to how humans and animals acquireskills through experience. However, real-world embodied learning, such as thatperformed by humans and animals, is situated in a continual, non-episodicworld, whereas common benchmark tasks in RL are episodic, with the environmentresetting between trials to provide the agent with multiple attempts. Thisdiscrepancy presents a major challenge when attempting to take RL algorithmsdeveloped for episodic simulated environments and run them on real-worldplatforms, such as robots. In this paper, we aim to address this discrepancy bylaying out a framework for Autonomous Reinforcement Learning (ARL):reinforcement learning where the agent not only learns through its ownexperience, but also contends with lack of human supervision to reset betweentrials. We introduce a simulated benchmark EARL around this framework,containing a set of diverse and challenging simulated tasks reflective of thehurdles introduced to learning when only a minimal reliance on extrinsicintervention can be assumed. We show that standard approaches to episodic RLand existing approaches struggle as interventions are minimized, underscoringthe need for developing new algorithms for reinforcement learning with agreater focus on autonomy. | [
"Prior work has proposed a number of benchmarks for reinforcement learning, which are often either explicitly episodic {{cite:ffb4075fbcb9e36775ede65ca209836541835420}}, {{cite:d131441b2587a6c6e2688d3a6123edf5330b2daa}}, {{cite:edcaeed2c38a386b7f62535f3554fa28daa23793}}, or consist of games that are implicitly episodic after the player dies or completes the game {{cite:498de327346808a4d970107479d0efe5f967509d}}, {{cite:45066b97dcb6352d2372bffaee3be85521037b6a}}.In addition, RL benchmarks have been proposed in the episodic setting for studying a number of orthogonal questions, such multi-task learning {{cite:498de327346808a4d970107479d0efe5f967509d}}, {{cite:bcfc339ba0daf55af540abcce961d7a6259e28f6}}, sequential task learning {{cite:9e54b14bc5329aa374d7f06ba746681aa0d3225b}}, generalization {{cite:c9a0ae72a2a389a659f36d8201858752376e0665}}, and multi-agent learning {{cite:61cef5a51fe01377d23449cc2c873a6bf5c583a0}}, {{cite:09645f175e0b4783384f86c1a5067acee0bed176}}.These benchmarks differ from our own in that we propose to study the challenge of autonomy. Among recent benchmarks, the closest to our own is Jelly Bean World {{cite:2f2598edda6fd1d4fe29ef55c87bf21861efbf9f}}, which consists of a set of procedural generated gridworld tasks. While this benchmark also considers the non-episodic setting, our work is inspired by the challenge of autonomous learning in robotics, and hence considers an array of manipulation and locomotion tasks. In addition, our work aims to establish a conceptual framework for evaluating prior algorithms in light of the requirement for persistent autonomy.",
"Enabling embodied agents to learn continually with minimal interventions is a motivation shared by several subtopics of reinforcement learning research. The setting we study in our work shares conceptual similarities with prior work in continual and lifelong learning {{cite:3f48194fc8366b70938206665565e11961280de5}}, {{cite:710a167372ea3032c43056dec9e5e6e6363bf35c}}, {{cite:5a7bb490d79831cfdbb71e85aa57a3c513536f92}}, {{cite:2a9d22bd24a45346999ba57d73ffd615c2908a94}}.In context of reinforcement learning, this work has studied the problem of episodic learning in sequential MDPs {{cite:ff61ab16bd94c28724e0533b4b803eff7c7fe186}}, {{cite:12006fde92182478a05fba26d9732e5793553483}}, {{cite:d181173e3074f09714f17edf078283e4c921e0aa}}, {{cite:3a7d6d070beccb8688348b81676c7aa5e29e835b}}, {{cite:6a6cc5d97ad619c24685415949512e8a2d5b4a31}}, {{cite:1edfd0b602cd55b3de86674d9183bead02417bec}}, where the main objective is forward/backward transfer or learning in non-stationary dynamics {{cite:513dbfa558e4f45d34e9a92f38bc688f51da7307}}, {{cite:fed9593ccd52c5fa75bb6e4f8d9a57a36cedf38a}}, {{cite:1f9bc81ef09099e7ec8c6fdf166d0f00b967957d}}. In contrast, the emphasis of our work is learning in non-episodic settings while literature in continual RL assumes an episodic setting. As we will discuss, learning autonomously without access to oracle resets is a hard problem even when the task-distribution and dynamics are stationary.In a similar vein, unsupervised RL {{cite:28f123aeb3c31fc3627b0d78868ccf08aa432189}}, {{cite:a998b1e56d8f16998a45690617888af56d180b51}}, {{cite:f7a01143212665c67068002f712f222616f3fdc2}}, {{cite:60b8225e3ef2f2a0ee8d4d911b8dc19b7f2d7170}}, {{cite:803da7ec7295ccf987f5ff4135f0b5310f8d97fa}} also enables skill acquisition in the absence of rewards, reducing human intervention required for designing reward functions. These works are complimentary to our proposal and form interesting future work.",
"Reset-free RL has been studied by previous works with a focus on safety {{cite:da05b60ae1e974f8dfec6351cd16c8ac7c6be8b6}}, automated and unattended learning in the real world {{cite:e5d48612102abd55fec92414e625358b9e408420}}, {{cite:7a04f08701d842987030c3eef91593ecc2d83d7f}}, {{cite:66e2ed3c4e771a768bbf25adc32d0cb0b6ba26b5}}, skill discovery {{cite:4ff9f24eb52e05d12f091c6ee0f714effe54bca6}}, {{cite:1f9bc81ef09099e7ec8c6fdf166d0f00b967957d}}, and providing a curriculum {{cite:8b3bd1d5bd2a7ac89906296c0d644a369419880d}}.Strategies to learn reset-free behavior include directly learning a backward reset controller {{cite:da05b60ae1e974f8dfec6351cd16c8ac7c6be8b6}}, learning a set of auxillary tasks that can serve as an approximate reset {{cite:bd97db1e93582307596b14cb3f6aa647a251b356}}, {{cite:66e2ed3c4e771a768bbf25adc32d0cb0b6ba26b5}},or using a novelty seeking reset controller {{cite:7a04f08701d842987030c3eef91593ecc2d83d7f}}. Complementary to this literature, we aim to develop a set of benchmarks and a framework that allows for this class of algorithms to be studied in a unified way. Instead of proposing new algorithms, our work is focused on developing a set of unified tasks that emphasize and allow us to study algorithms through the lens of autonomy."
] | [] |
Align and Prompt: Video-and-Language Pre-training with Entity Prompts | Video-and-language pre-training has shown promising improvements on variousdownstream tasks. Most previous methods capture cross-modal interactions with atransformer-based multimodal encoder, not fully addressing the misalignmentbetween unimodal video and text features. Besides, learning fine-grainedvisual-language alignment usually requires off-the-shelf object detectors toprovide object information, which is bottlenecked by the detector's limitedvocabulary and expensive computation cost. We propose Align and Prompt: an efficient and effective video-and-languagepre-training framework with better cross-modal alignment. First, we introduce avideo-text contrastive (VTC) loss to align unimodal video-text features at theinstance level, which eases the modeling of cross-modal interactions. Then, wepropose a new visually-grounded pre-training task, prompting entity modeling(PEM), which aims to learn fine-grained region-entity alignment. To achievethis, we first introduce an entity prompter module, which is trained with VTCto produce the similarity between a video crop and text prompts instantiatedwith entity names. The PEM task then asks the model to predict the entitypseudo-labels (i.e~normalized similarity scores) for randomly-selected videocrops. The resulting pre-trained model achieves state-of-the-art performance onboth text-video retrieval and videoQA, outperforming prior work by asubstantial margin. Our code and pre-trained models are available athttps://github.com/salesforce/ALPRO. | [
"Dense versus Sparse Video Representation. Consecutive frames in videos usually contain visually similar information.Such redundancy opens up a research question on how to learn effective video-and-language representations without excessive computation overhead.Most prior methods on text-video retrieval {{cite:f50416fd2a429b55b14960159626bdf73f7c4d7f}}, {{cite:2f916c0419963a8bbb2bc3be9fd2149a469a7089}}, {{cite:809afb604237e08ea534513752c18a052012eaa7}}, {{cite:172399f776a0a3baf38df4a5fca909f700549eb6}}, {{cite:d1369dd9338191b97e13faf0a1b12f471ed2964e}} and videoQA {{cite:b1d9c0993ab7db3d683219917a1a1c4a73f9bc0d}}, {{cite:6e02eb08bd53fd7081e9762aacbd4f4d12634239}}, {{cite:7be40c91405e8b425620146696d9dbe0149bd108}}, {{cite:8141df159053fd697f4ae9992015ae88b8f0ecb6}} employ pre-trained visual backbones and extract video features for each frame densely yet offline. However, since visual backbones are usually pre-trained on image {{cite:d47029df70581dd7596a41a2ecf1370ce5236fe5}} and/or video datasets {{cite:d24ed8d6d971af0196d52f00d7ca49c1bca2c752}} without access to text, these features are less effective for video-and-language tasks.Besides, video feature extractors in these approaches are not finetuned on target task data, preventing features to easily adapt to different domains.In contrast, recent methods ClipBERT {{cite:227f4f28a091597b9f1400851e29ecdffad59fb7}} and FiT {{cite:7d861bc3cff12e59647dbb3fd9b35fb643b822ad}} demonstrate more effective results by end-to-end finetuning the visual backbone with only a few sparsely sampled frames.However, ClipBERT is pre-trained with image-text data thus is less effective in aggregating information across frames, while FiT is a retrieval-specific architecture that does not naturally generalize to videoQA task.In this regard, our AlPro is the first sparse pre-training architecture that tackles both tasks while in the meantime, demonstrating the benefit of pre-training on video-text pairs.",
"Video-and-language Pre-training. Apart from the canonical pre-training tasks, such as masked language modeling (MLM) {{cite:1ea4c3e40bb3a8e61abcacceab50b39ede2b2595}}, {{cite:31d0a30f76240f377f2732b7c3f1f6324bf94c9e}}, {{cite:32bae67e23331e0d2d690e9451c1d32019f68ddc}}, {{cite:e85778ed9b882b6e44f12cb2251111a8506956ac}}, {{cite:52c5d85316878ba00ca62d9dae5bbc3593b878e5}}, {{cite:227f4f28a091597b9f1400851e29ecdffad59fb7}} and video-text matching (VTM) {{cite:52c5d85316878ba00ca62d9dae5bbc3593b878e5}}, {{cite:e85778ed9b882b6e44f12cb2251111a8506956ac}},several methods {{cite:be0ac4d7d76b4f40f050fd50d74b8b39cb2e977c}}, {{cite:e85778ed9b882b6e44f12cb2251111a8506956ac}}, {{cite:4dcea1098a6683765609927571d1a9c2727b451f}} apply contrastive learning on offline extracted visual features.Without adapting the visual backbone, their ability to align cross-modal features remain limited.AlPro jointly learns the unimodal and multimodal encoders, thus mitigating the disconnection in-between.In order to design effective visually-grounded pre-training tasks, VideoBERT {{cite:31d0a30f76240f377f2732b7c3f1f6324bf94c9e}} predicts centroids of vector quantizations of video features.Such unsupervised quantization is noisy per se while neglecting textual cues, which curtails its capabilities in learning cross-modal interactions.ActBERT {{cite:32bae67e23331e0d2d690e9451c1d32019f68ddc}} uses detectors to acquire object information. In addition to their computational inefficiency, detectors pre-trained on images usually have limited categories and compromised detection results on videos.In contrast, our proposed prompting entity modeling task is detector-free.By exploiting the instance-level video-text alignment, we can generate reliable entity pseudo-labels with a large vocabulary,leading to more efficient and effective learning of region-entity alignment.",
"Zero-shot Visual Recognition with Prompts. There have been longstanding efforts to exploit text descriptions for learning visual recognition models.These include early efforts {{cite:774f57dfc35b0b13520e50e7336cfb4c2db5a379}}, {{cite:3ba137e498a8be3103d30c811c72a47953db2469}}, {{cite:baced56ad4a3e314052fecb8c9ee71346ef2bc0c}} that use text to learn attributes of images; {{cite:7c2410666917958aa0c5ff28e831d67ca2b57972}}, {{cite:661decfc3e45d87319c9e6e94a73a3d53ee1dd7c}}, {{cite:75032ab90c061e7139f4057caf457887862e505e}} that map images to the pretrained text embedding space, and visual n-gram {{cite:bca8feeaba577251fea1c2fcfcc7c89e9e63443c}} that predicts text n-grams given image inputs.More recently, CLIP {{cite:e426482a56bdb37091e8e8230b76a5bae1ec34bf}} instantiates prompt templates with label text of visual categories.It then predicts the category by computing the similarity between each image-prompt pairs.This inspires us to the design of entity prompter.Since the entity prompter is trained with the entire video-text corpus, during pre-training, it can provide additional entity information unavailable in the text description for each individual video-text pair,thus leading to better entity-informed video representations."
] | [] |
Linguistic and Gender Variation in Speech Emotion Recognition using
Spectral Features | This work explores the effect of gender and linguistic-based vocal variationson the accuracy of emotive expression classification. Emotive expressions areconsidered from the perspective of spectral features in speech (Mel-frequencyCepstral Coefficient, Melspectrogram, Spectral Contrast). Emotions areconsidered from the perspective of Basic Emotion Theory. A convolutional neuralnetwork is utilised to classify emotive expressions in emotive audio datasetsin English, German, and Italian. Vocal variations for spectral featuresassessed by (i) a comparative analysis identifying suitable spectral features,(ii) the classification performance for mono, multi and cross-lingual emotivedata and (iii) an empirical evaluation of a machine learning model to assessthe effects of gender and linguistic variation on classification accuracy. Theresults showed that spectral features provide a potential avenue for increasingemotive expression classification. Additionally, the accuracy of emotiveexpression classification was high within mono and cross-lingual emotive data,but poor in multi-lingual data. Similarly, there were differences inclassification accuracy between gender populations. These results demonstratethe importance of accounting for population differences to enable accuratespeech emotion recognition. | [
"Previous research has examined the differences between gender populations in acoustic features and their effect on SER performance. Several acoustic features have been described as “gender-dependent”. Gender dependent in this case means that gender significantly influences the expression of that feature {{cite:0be887a79d741567e5bd957d9303b86d8abf7521}}. Pitch is an example of a gender-dependent feature. Females score higher, on average, than males on pitch {{cite:43ebda0c54ffd5a417996a7b99caa393f9076578}}. This information, however, is not sufficient to enable generalisable SER classification, because there is considerable overlap between males and females on this feature {{cite:43ebda0c54ffd5a417996a7b99caa393f9076578}}. This is also the case for several other acoustic features described as “gender-dependent”.",
"Similarly, previous research has examined the impact of language in developing accurate SER models. Human observers can accurately classify emotions in their native language, and non-native languages {{cite:da13277bc18e18032ff4a19d265a4455fd850267}}. This result has been replicated across various cultures, including indigenous tribes, demonstrating the generalisability of traditional SER {{cite:4f6fa876c244c677181e26f26b26e20b75818e73}} {{cite:01edc1cebc70f3dab740cd2c23cc92333aff9e0a}}. Similarly, automated SER models achieve a high level of accuracy in mono-lingual settings, where the models are trained and evaluated in the same language. The performance of these models, however, significantly drop when they are tested across languages {{cite:8af49f34145860e938ae8a010b7f303f5217f38d}} {{cite:913940486fd3328f9746abbc0deb02fb2be2b575}}.",
"If ML models can differentiate between acoustic features that indicate emotional expressions for each gender and linguistic population reliably, then one should expect generalisability of SER performance between both populations {{cite:cde50f7068cd94875554f8bebfd15328f6f49b2d}}. To achieve this, spectral features that reliably indicate emotional experiences are first required.",
"A survey {{cite:5edf913e887af9bf82f1628eb0f643d7823f30c0}} analysed features across 17 distinctive datasets. These datasets comprise 9 languages, both male and female speakers, professional and non-professional acting sessions, recorded call centre conversations and speech recorded under simulated and naturally occurring stress situations.",
"A pool of acoustic features and SER were analysed in {{cite:5edf913e887af9bf82f1628eb0f643d7823f30c0}} and {{cite:8c9ffebf89db44322f26d742cddb03863d2ef5b4}}. Which can be divided into several categories (Continuous, Qualitative, Spectral and Non-linear Teager Energy Operated (TEO). The results of both reviews indicated that Continuous and Spectral features were related to emotional expressions in speech. A weaker relationship was found between emotional expressions and Qualitative and TEO features.",
"To automatically classify relevant spectral features, a ML methodology is proposed in {{cite:d20088fc86ae6e11215352d7c1c1045f707a1662}}. The methodology uses a convolutional neural network (CNN) and SER analysis. The approach consists of using spectral features such as Mel-frequency Cepstral Coefficients (MFCCs), Mel-scaled spectrogram, Chromagram, Spectral contrast and Tonnetz representation extracted from emotive audio belonging to three distinct datasets. This work, however, focused on a combined gender classification of emotive expression, inhibits the analysis of vocal variations between gender. This study extends this work to by analysing the effect of gender and linguistic populations on the emotive expression classification of spectral features."
] | [] |
Mitigating inefficient task mappings with an Adaptive Resource-Moldable
Scheduler (ARMS) | Efficient runtime task scheduling on complex memory hierarchy becomesincreasingly important as modern and future High-Performance Computing (HPC)systems are progressively composed of multisocket and multi-chiplet nodes withnonuniform memory access latencies. Existing locality-aware scheduling schemeseither require control of the data placement policy for memory-bound tasks ormaximize locality for all classes of computations, resulting in a loss ofpotential performance. While such approaches are viable, an adaptive schedulingstrategy is preferred to enhance locality and resource sharing efficiency usinga portable programming scheme. In this paper, we propose the AdaptiveResource-Moldable Scheduler (ARMS) that dynamically maps a task at runtime to apartition spanning one or more threads, based on the task and DAG requirements.The scheduler builds an online platform-independent model for the local andnon-local scheduling costs for each tuple consisting of task type (function)and task topology (task location within DAG). We evaluate ARMS usingtask-parallel versions of SparseLU, 2D Stencil, FMM, and MatMul as examples.Compared to previous approaches, ARMS achieves up to 3.5x performance gain overstate-of-the-art locality-aware scheduling schemes. | [
"There has been extensive literature addressing data locality concerns in HPC applications and systems. On the theoretical level, models that predict locality for parallel programs by means of reuse distance analysis have been discussed in {{cite:092f4605d83a225628f7cab35af1650f4fd3421f}}, {{cite:674c00ba6bb2f997ffcd3b92a7f1f9c6bc54d0ab}}, {{cite:eb8f9f4d204610d0dfaefa7108720bff06bd3fc6}}. Such models aid the programmers in understanding the memory access patterns of their codes. Additionally, data-centric programming models (e.g. Legion {{cite:e8412e8a3aaa132ade6cabe7156f112e2fbc34f0}}) and language extensions (e.g. Hierarchical Place Trees (HPT) {{cite:7dab24bb8cc0faa23fbfe9ac8e6354b55f968b56}}) require the domain expert programmer to analyze the data access regions and express the program using their interfaces, and to make explicit assignments to execution places.Despite their effectiveness, programmer level techniques (e.g. loop tiling or hierarchical cache optimizations {{cite:d1bdc08c40ac8c9c9db80508e620a7931f99374b}}) are known to have limited portability across platforms. Hence, high-level abstractions in data locality have received a growing attention from the HPC research community {{cite:9dc16519709c0db79c328b4a25a81592949c48ef}}, {{cite:65eac5e396c6bffa8a353c415fc643277988ed54}}. These abstractions are meant to adopt existing features from parallel execution frameworks, such as the affinity partitioner and task arenas from Intel TBB (Thread Building Blocks), places and processor binding from OpenMP, topology and communication reducing in MPI (Message Passing Interface) {{cite:5480fed3f8c1c9956207ec465c046819270552ca}}, {{cite:32b5aebfd46a94f7c74c0a15d703d1490e92f2da}}.",
"As discussed, many task-based parallel libraries embrace random work-stealing scheduling mechanisms, which are provably known to be effective in maximizing dynamic load-balancing. However, they are bound to suffer from scalability issues with memory-bound tasks. To address these, techniques have been proposed to achieve locality-aware work-stealing behavior such as the Almost Deterministic Work Stealing (ADWS) {{cite:701991dbcd6a9ba728c022ebe4df8887eafc544f}}. This is done via deterministic mapping of tasks to resources that is applied based on the programmer's annotation of workload size. Also, the paper discusses an approach to achieve hierarchical localized work stealing. Another technique is highlighted in Scalable Locality-aware Adaptive Work-stealing Scheduler (SLAW), which also allocates the resources based on user annotations of locality (a place index) {{cite:0083424efc96e45214c0234c7f5675f8bf489fff}}. Last but not least, XKaapi {{cite:85dd1584c25e60fc8e2398b032c9bb8d9438351b}} is a locality-aware work stealing scheduler that supports heterogeneous data flow programming targeting both CPUs and GPUs.",
"While these are effective techniques in reducing the side effects of random work-stealing, none of them consider the difficulty of obtaining a static classification of applications into memory and compute intensive classes. We think that ARMS is not completely orthogonal to the literature above, as it still maximizes locality when deemed necessary by the online model. However, ARMS is not yet-another locality-aware scheduler, but it rather adopts a dynamic strategy and opts for locality maximizing by understanding the behavior of a task at runtime. We believe that this strategy is only more realistic considering the evolving complexity of applications and hardware."
] | [] |
Challenge Dataset of Cognates and False Friend Pairs from Indian
Languages | Cognates are present in multiple variants of the same text across differentlanguages (e.g., "hund" in German and "hound" in English language mean "dog").They pose a challenge to various Natural Language Processing (NLP) applicationssuch as Machine Translation, Cross-lingual Sense Disambiguation, ComputationalPhylogenetics, and Information Retrieval. A possible solution to address thischallenge is to identify cognates across language pairs. In this paper, wedescribe the creation of two cognate datasets for twelve Indian languages,namely Sanskrit, Hindi, Assamese, Oriya, Kannada, Gujarati, Tamil, Telugu,Punjabi, Bengali, Marathi, and Malayalam. We digitize the cognate data from anIndian language cognate dictionary and utilize linked Indian language Wordnetsto generate cognate sets. Additionally, we use the Wordnet data to create aFalse Friends' dataset for eleven language pairs. We also evaluate the efficacyof our dataset using previously available baseline cognate detectionapproaches. We also perform a manual evaluation with the help of lexicographersand release the curated gold-standard dataset with this paper. | [
"wu2018creating release cognate sets for Romance language family and provide a methodology to complete the cognate chain for related languages. Our work releases similar data for Indian languages. Such a cognate set data has not been released previously for Indian languages, to the best of our knowledge. Additionally, we release lists of false friends' for language pairs. These cognates can be used to challenge the previously established cognate detection approaches further. kanojia2019cognate perform cognate detection for some Indian languages, but a prominent part of their work includes manual verification and segratation of their output into cognates and non-cognates. Identification of cognates for improving IR has already been explored for Indian languages {{cite:1deb2236a46f5b1f532a3d1338226a12d7fd0356}}. Orthographic/String similarity-based methods are often used as baseline methods for cognate detection, and the most commonly used method amongst them is the Edit distance-based similarity measure {{cite:95941fd5f0b2a1870339811adcdfaa6c2869857d}}.",
"Research in automatic cognate detection using various aspects involves computation of similarity by decomposing phonetically transcribed words {{cite:90d9446658e66c9ecc8aa9b00266f9bd4b606575}}, acoustic models {{cite:93de472724922794aa9ecf5a96c13d9f29df8863}}, clustering based on semantic equivalence {{cite:02d470cbdac106a01281ebc215984b8131164b16}}, and aligned segments of transcribed phonemes {{cite:0048542ae19b4c6efc27951f595ecd61e72b1d84}}. rama2016siamese employs a Siamese convolutional neural network to learn the phonetic features jointly with language relatedness for cognate identification, which was achieved through phoneme encodings. jager2017using use SVM for phonetic alignment and perform cognate detection for various language families. Various works on orthographic cognate detection usually take alignment of substrings within classifiers like SVM {{cite:eb901e42dbd4d03f7603abd025a05b223c48df6c}}, {{cite:ce98d331665c1b9488a409a83b9f5d4eeffab8ae}} or HMM {{cite:2f50ff801bc38b956542ae8a3a518dcc90b18ea0}}. ciobanu2014automatic employ dynamic programming based methods for sequence alignment. Among cognate sets, common overlap set measures like set intersection, Jaccard {{cite:82c9e41f522c2ca7ade741b60e1db53aef916f36}} or XDice {{cite:1d743a42dbb17bf5f82633d90e114e82cc282616}} could be used to measure similarities and validate the members of the set.{{table:5a46dc45-fa15-4667-ada4-2ca90b1eb554}}{{table:c75d86a1-b821-4db7-92fc-7daccdee47b7}}{{table:4081e83c-9510-429b-b01d-4afcf3a0c4d2}}"
] | [] |
Interpretable Data-Based Explanations for Fairness Debugging | A wide variety of fairness metrics and eXplainable Artificial Intelligence(XAI) approaches have been proposed in the literature to identify bias inmachine learning models that are used in critical real-life contexts. However,merely reporting on a model's bias, or generating explanations using existingXAI techniques is insufficient to locate and eventually mitigate sources ofbias. We introduce Gopher, a system that produces compact, interpretable andcausal explanations for bias or unexpected model behavior by identifyingcoherent subsets of the training data that are root-causes for this behavior.Specifically, we introduce the concept of causal responsibility that quantifiesthe extent to which intervening on training data by removing or updatingsubsets of it can resolve the bias. Building on this concept, we develop anefficient approach for generating the top-k patterns that explain model biasthat utilizes techniques from the machine learning (ML) community toapproximate causal responsibility and uses pruning rules to manage the largesearch space for patterns. Our experimental evaluation demonstrates theeffectiveness of Gopher in generating interpretable explanations foridentifying and debugging sources of bias. | [
"Our work relates to the following lines of research.",
"Feature-based explanations.XAI research focuses on explaining ML models in terms of patterns and dependencies between input features and their outcomes. Its methods are based on feature attribution, such as the Shapley value, to quantify the responsibility of input features for ML model predictions {{cite:ee7177945392f728e356cde344356f6ffe6585d7}}, {{cite:25ad538d90f98683c26f840cbef8aa2af2853590}}, {{cite:95fadb6a7a6165d9628dc1546c46e7a7f3f84b79}}, {{cite:1f564776e762259d8d491672dc6464b0f0ba15c1}}, {{cite:dff39be3c612b4ce01a3701576ffc140cd1d172a}}, {{cite:0362616942c22a1382c7c2b7f1ce211504242257}}, {{cite:a1f65d16920884ce06a4919e69743bbfd82b3ca4}}, {{cite:c385c60d7be88f7525bcbeadae1074fbc1a9b4c8}}. Methods based on surrogate explainability approximate ML models using a simple, interpretable model (such as linear regression) {{cite:95fadb6a7a6165d9628dc1546c46e7a7f3f84b79}}, {{cite:c00f6ba9e8e2a04e4815bf624bee448fa54b44c7}}, {{cite:ca248686f857630b6ddc47c64dad7f408304eae6}}. Contrastive and causal XAI methods explain ML model predictions in terms of minimal interventions or perturbations on input features that change the prediction {{cite:1e8642f7c663289cf64d2da4b176f83c23be33a4}}, {{cite:6800a341ffa240248f487ccb94394a3ac1a5fcd2}}, {{cite:ad5b157cbee89927b967c32cb6bc09793fb81068}}, {{cite:8cb50f3b804b14d478310cb186124fc7cc6eb794}}, {{cite:04e94f5c00b6987a864352bfaab8454a5a58f103}}, {{cite:45976afaed1934a9f69e2d9386a304629326537a}}, {{cite:15e91915dd9a479d1ed45049f11a2de22a05f7ff}}, {{cite:fd25932a2cb67c342f4006937d47e49175754fc3}}, {{cite:5719424af49632a47b430ad5a85d8fabef8128d7}}. Logic-based methods use tools from logic-based diagnosis that operate on logical representations ofML algorithms {{cite:04e73115122fc34657eccc987006cee08ed94855}}, {{cite:a1de554f821af910b92c39c4fbf3e4ec86c78568}}, {{cite:0779617b2899d660456f29f4a7cce652ed7e6124}} to compute minimal sets of features that aresufficient and necessary for ML model predictions. These approaches fall short in generating diagnostic explanations that help users trace an ML algorithm's unexpected or discriminatory behavior back to its training data.",
"Explanations based on training data.These XAI methods attribute ML model prediction to specific parts of its training data {{cite:77e0dfdccca24af4ba0a0b27bb149d497f1196ef}}. All current approaches rank individual training data points based on their influence on ML model predictions {{cite:a15e26fa26fa713ad71c282983fdb86bf8bc06e8}}, {{cite:c95ac5621e4562a448847ee1613dc4a648ef91c5}}, {{cite:c80087032f1022dee4c460c232f3743c4663d562}}, {{cite:baf695991d2b243c07a64381b667d9559135fd74}}, {{cite:bd87328d3f19be88c776ce4c6c18ace2d8a12e4e}}, {{cite:9261659459283148adea877047c8610494e5fe47}}, typically using influence functions {{cite:769b277708c7b7e4eb6b46a8ea3d96464d93ffaf}}. Influence functions are a classic technique from robust statistics that measure how optimal model parameters depend on training data instances. Based on first-order influence functions, {{cite:5dd8daac112a633d0581fe73a22b2a22c50e9fb4}} introduced an approach for identifying training data points that are responsible for user constraints specified by an SQL query.There are other methods than influence functions, however, e.g, {{cite:9261659459283148adea877047c8610494e5fe47}} develops an approach for identifying and ranking training data points based on their influence on predictions of neural networks, and {{cite:3b0ef119b93e95ec1aa1e87e3d6796b049608d58}} develops an approach for incremental computation of the influence of removing subset of training data points Furthermore, other recent work argues for the use of data Shapley values to quantify the contribution of individual data instances {{cite:c80087032f1022dee4c460c232f3743c4663d562}}, {{cite:baf695991d2b243c07a64381b667d9559135fd74}}, {{cite:bd87328d3f19be88c776ce4c6c18ace2d8a12e4e}}; these approaches are computationally expensive because each data instance requires the model to be retrained.Unlike prior methods, our method generates: (1) explanations for fairness of an ML model, (2) interpretable explanations based are first-order predicates that pinpoint a subset of training data responsible for model bias, and (3) update-based explanations that reveal data-errors in certain attributes of a training data subset.It is worth mentioning here thatreducing the bias in the training data can be done not only by removing subsets of training data but also by adding certain appropriate training samples. Searching for in-distribution data points that explain away bias upon insertion into training data is a challenging problem that we plan to explore in the future. Please note, however, that our update-based explanations can act as a combination of removing existing data points and adding new ones.",
"Debugging ML models.Much work examines the debugging of ML models for fairness and bias. fair-DAGs {{cite:3743937f5119ef8d88d213e9439e17d27be3795b}} and mlinspect {{cite:3bdcf0781a2eb4acedc83e0e8247fde8baa1d909}} regard class imbalance as a cause of discrimination and track the distribution of sensitive attributes along ML pipelines to address bias. These approaches cannot highlight problematic parts of training data, i.e., those responsible for biased outputs. MLDebugger {{cite:f200033b52970c00f93ed60b19548d31c4f0f200}} applies a method that automatically identifies one or more minimal causes of unsatisfactory performance in ML pipelines using provenance of the previous runs. {{cite:be91c590321209f1423c6585117860c0a1414d6b}}, {{cite:9de6806b5dff9c19ad560cc8faf069616b838c5f}} develop methods to identify regions of attribute space not adequately covered by training data. {{cite:b47f51b0c798c3a8623469106ed49d917d524c27}} develops an approach for debugging training data bymaking the smallest set of changes to training data labels so the ML model trained on the updated data can correctly predict labelsof a trusted set of items in test data; since the focus of their approach is on updating training labels, it is not applicable to our setting. Another related line of work concerns finding large data slices in which the model performs poorly {{cite:5c3d9a4a9782a7bede0bb330aa35cd226f562d67}}, {{cite:83515690c5091c50ec5c04f6d6f689d0295a7e3c}}, {{cite:e381b31467d2852a736a5da6c20bc43df641cd70}}; the slices are discovered based on their association with model error and do not capture the causal effect of interventions. These techniques are also not directly applicable in the context of fairness, where, unlike model error, bias due to a subset is not additive. Moreover, none of the preceding interventions update data instances.",
"Adversarial ML.Our work shares similarities with adversarial machine learning literature, which explores ways in which ML can be compromised by adversarial attacks. The most relevant classes of adversarial attacks are based on data poisoning {{cite:804bc6e315d8e20a34a706459268a7b6f21f09d5}}, {{cite:eac88a157bdf2d5b50114b23fc625f63fa7d8373}}, where polluting training data compromises model fairness or predictions {{cite:f6b481161a5ea3ff82e5244a2113b0820803aca1}}, {{cite:ecbbea95d642fc07f210ff0d964a24d42bea3e03}}, {{cite:f197f3ae69a5826b672cd601bad6c3fb047d6986}}. Data poisoning attacks on fairness aim to inject a minimum set of synthetic poisonous data points in training data that compromise the fairness of a model trained on the contaminated data.This is in reverse to the goal of our research. Note that many data poisoning attacks methods rely on influence functions.",
"Machine unlearning.Our research also relates to the nascent field of machine unlearning {{cite:740bf39bb7d234c522d7a99d89ec6ef91666c71c}}, {{cite:b5da9d1e8b9256de604881ccc16f5191fd9d8777}}, {{cite:029fdf0de44304f06881e75f09690921ed385e2d}}, {{cite:b9e6f894357d0fa0c7cb9dffb33753b3ad994cd7}}, which addresses the “right to be forgotten\" provisioned in recent legislation, such as the General Data Protection Regulation (GDPR) in the European Union {{cite:e938ba2f61fe91e129cdc3260fe3a73b3c6c10ed}} and the California Consumer Privacy Act in the United States {{cite:dab170d4bd1037efcb74e2f04826d4313986d535}}. Current methods typically designed for particular classes of ML models, e.g., {{cite:029fdf0de44304f06881e75f09690921ed385e2d}}, support efficient unlearning requests for decision trees. However, the techniques in this sub-field could be integrated into our framework for efficient computation of causal responsibility. We defer this investigation to future study.",
"Bias detection and mitigation.Prior work explores ways to detect and mitigate bias in ML models. These approaches can be categorized into pre-, post- and in-processing (see {{cite:9b7ab8798431cf89427e971339edf8cb642dd6b3}} for recent surveys).Our goal is to pre-process data to reduce bias which is different from post-processing that deals with model output to handle bias, and in-processing that is aimed at building fair ML models. The idea is to remove bias and discrimination signals from training data by repairing or pre-processing it (see, e.g., {{cite:83cdec692cdb70516c4e3efd7f68616fe00fb603}}, {{cite:d1c6e4c6612c1141a00ced8475fe8d576aa261a4}}, {{cite:81df004a4043881220201af05b5f08380f76c8ac}}. Pre-processing methods in fairness are independent of the downstream ML model and usually are not interpretable; hence, they are insufficient in generating explanations that reveal the potential source of bias. While bias mitigation is not a direct focus of our research, the approach we propose could be useful for developing bias mitigation algorithms that are interpretable and take into account the downstream ML model; hence, they would incur minimal information loss and generalize better. We defer this research to future work."
] | [] |
scenoRITA: Generating Less-Redundant, Safety-Critical and Motion
Sickness-Inducing Scenarios for Autonomous Vehicles | There is tremendous global enthusiasm for research, development, anddeployment of autonomous vehicles (AVs), e.g., self-driving taxis and trucksfrom Waymo and Baidu. The current practice for testing AVs uses virtualtests-where AVs are tested in software simulations-since they offer a moreefficient and safer alternative compared to field operational tests.Specifically, search-based approaches are used to find particularly criticalsituations. These approaches provide an opportunity to automatically generatetests; however, systematically creating valid and effective tests for AVsoftware remains a major challenge. To address this challenge, we introducescenoRITA, a test generation approach for AVs that uses evolutionary algorithmswith (1) a novel gene representation that allows obstacles to be fully mutable,hence, resulting in more reported violations, (2) 5 test oracles to determineboth safety and motion sickness-inducing violations, and (3) a novel techniqueto identify and eliminate duplicate tests. Our extensive evaluation shows thatscenoRITA can produce effective driving scenarios that expose an ego car tosafety critical situations. scenoRITA generated tests that resulted in a totalof 1,026 unique violations, increasing the number of reported violations by23.47% and 24.21% compared to random test generation and state-of-the-artpartially-mutable test generation, respectively. | [
"A wide array of studies focus on applying traditional testing techniques to AVs including adaptive stress testing {{cite:c9766788040e32354725804f503e53fe39d268a2}}, where noise is injected into the input sensors of an AV to cause accidents; fitness function templates for testing automated and autonomous driving systems with heuristic search ; and search-based optimization {{cite:928f7b46897a6992488cdaad6b0dfddc38336805}}. These studies provide limited insights into the testing of real-world AVs, since they do not evaluate their techniques on open-source, production-grade AV software.",
"Other related work reproduce tests from real crashes such as {{cite:8c2543591b9a55c070df1e0f100f041e84ce6a2d}}, where car crashes are recreated by replaying the sensory data collected during physical-world crashes. Similarly, AC3R {{cite:9b530c50ef206f762eac3cc368edfb520b18dc79}} generates driving simulations which reproduces car crashes from police reports using natural language processing (NLP). However, AC3R requires manual collection of police reports and inherit the accuracy limitations of the underlying NLP used to extract information from police reports.",
"Another line of work, combines procedural content generation and search-based testing to automatically create challenging virtual scenarios for AV software {{cite:4069b9c52025fa599dd4ba1334e54f1752b6ef4f}}. However, AsFault, does not take into account the behaviour of other obstacles when testing for safety violations in AVs. Calò et al. {{cite:359d98486a7ae782caa99a0717e1c99fe3b7d35c}} proposed two search-based approaches for finding avoidable collisions. They define comfort and speed as weights to rank short-term paths. However, they do not formalize comfort and speed in the fitness function, nor do they evaluate them. Furthermore, their work is applied to closed-source anonymous software, with no information about the maps used or scenarios’ configurations.",
"AV-Fuzzer {{cite:12d3bd283425805ebd0816411f6a0af44e297b90}} perturbs driving maneuvers of traffic participants to create situations in which an AV can run into safety violations. However, AV-Fuzzer requires the user to manually define a driving environment by specifying limited maneuvers (i.e., only acceleration/deceleration, lane following, and lane changing) of obstacles. scenoRITA automatically generates the maneuvers of obstacles and even supports a wider variety of maneuvers: lane follow, lane change, left turn, right turn, U-turn, and lane merge.",
"A series of work focuses on the vision and machine-learning aspects of AV software {{cite:6d9c6efaa94e0d22fdf11644c14f4e3f931a78ce}}, {{cite:9d090df6c6817bf6f491cfafa882cb191681fd07}}, {{cite:1b8628aa0e58ef141fa1a2562ea9a41703bf6ce5}}, {{cite:03292a1c5904197a7c4f20a9f3a30d63443db6c8}}, {{cite:70e7b325f15e591dd954973e4c320c2d82f0d3d0}}, {{cite:22c43de23662b162e6fe490bef17c7128aa46964}}, {{cite:861d1c22b6a99ef9a12bc34b1bc320bd1efe575a}}. Rather than focus on these aspects, scenoRITA targets the planning component of AV software. Previous work has shown that the most bug-ridden component of production-grade, open-source AV software systems is the planning component as opposed to the components responsible for or utilizing vision or machine-learning capabilities of AVs {{cite:fdeabbaa778c359e76921c9634f4cba8bb952d00}}.",
"Unlike previous work, the gene representation in scenoRITA allows the states and properties of obstacles in a scenario to be fully mutable. Previous techniques either do not specify their gene representations or only allow obstacles to be partially mutable, avoiding having to deal with generating valid obstacles trajectories. The aforementioned previous work uses highly limited numbers of oracle types and fitness functions—mainly consisting of one test oracle or fitness function per work (either collision detection or lane keeping) and ignoring motion sickness-inducing violations. scenoRITA, however, utilizes 3 safety oracles (i.e., collision detection, speeding detection, and unsafe lane change), 2 motion sickness-oriented oracles (i.e., fast acceleration and hard braking detection), and includes associated fitness functions. The lack of sufficient test oracles and fitness functions in previous work reduces the safety assurance and ignores vehicle-induced motion sickness—which many people (e.g., one-third of Americans) suffer from in the world {{cite:91ac800c0447de338f336710a3a8908a6d447d64}}, {{cite:17c78f4f5f5d9f3f8ef5dcfbd49117af6cd81e83}}, {{cite:6ab4d4901623e338ad9facebd4c31e58eb51089d}}."
] | [] |
Towards More Effective PRM-based Crowd Counting via A Multi-resolution
Fusion and Attention Network | The paper focuses on improving the recent plug-and-play patch rescalingmodule (PRM) based approaches for crowd counting. In order to make full use ofthe PRM potential and obtain more reliable and accurate results for challengingimages with crowd-variation, large perspective, extreme occlusions, andcluttered background regions, we propose a new PRM based multi-resolution andmulti-task crowd counting network by exploiting the PRM module with moreeffectiveness and potency. The proposed model consists of three deep-layeredbranches with each branch generating feature maps of different resolutions.These branches perform a feature-level fusion across each other to build thevital collective knowledge to be used for the final crowd estimate.Additionally, early-stage feature maps undergo visual attention to strengthenthe later-stage channels understanding of the foreground regions. Theintegration of these deep branches with the PRM module and the early-attendedblocks proves to be more effective than the original PRM based schemes throughextensive numerical and visual evaluations on four benchmark datasets. Theproposed approach yields a significant improvement by a margin of 12.6% interms of the RMSE evaluation criterion. It also outperforms state-of-the-artmethods in cross-dataset evaluations. | [
"Crowd counting has many different key challenges including image resolution and crowd-density fluctuation in and across different images, extreme occlusions, far-reaching crowd perspective within an image, cluttered no-crowd regions in the images. Many classical schemes were proposed that generally aimed to solve this task as either a regression or a detection problem. Regression-based counting {{cite:0cbebcd7663a43efbf4f679c67f90348629919c9}}, {{cite:7bb6f8fd82c0115e9956cbfc274c932ce2bf31de}}, {{cite:934d69d747c9b967aa12ca4dbaa5ca097c05d5f0}} learns the local crowd features mapping function for the given image to regress the crowd count. But they fail to reasonably comprehend the key crowd counting challenges. Detection-dependent {{cite:a5d381ae600fe18c9289ec00ce76da090665f749}}, {{cite:32143675478ea583d8c1c419ec47f88606f0bc39}}, {{cite:a5b4dda370d5b644a1eab9bd5390a8930fef7bf1}}, {{cite:45eb715a7f56b91b31497a71a85d976697cbc29a}} schemes detect hand-crafted features to estimate the crowd count. These methods become impractical and ineffective in the case of high-dense crowd images because of the very small area per head or person.",
"Recently, Convolutional Neural Networks (CNNs) based approaches are the backbone of almost all crowd counting methods. Generally, we can divide them into three classes: Detection-dependent, Regression-dependent, and density-map estimation {{cite:cfbb5f15daf02b78f679d13f5e51b7c9e497160c}}. Well established object detectors (like YOLO et al. {{cite:92ac5a3a0b45c0791115dca577cdc4662321e7fc}}, {{cite:37d0cfc62157533c3f3365c7635c5e2d213b83cd}}) are being employed by the detection-dependent {{cite:792277daaf98f3b5222298dd871f1ce01b8d184e}}, {{cite:d593b139dd16ca1a80fe2bec95b5b6e395b01f2b}} methods to identify the people in the given image, and then sum all the detections to give the final crowd count. However, they fail for the images with high-density crowd due to smaller area per person or head. Shami et al. {{cite:792277daaf98f3b5222298dd871f1ce01b8d184e}} performed the weighted average of the CNN-based detections to produce the final people count. Li et al. {{cite:d593b139dd16ca1a80fe2bec95b5b6e395b01f2b}} employed the CNN based adaptive head detection method that used the context-based information for the crowd estimation.",
"Regression-dependent {{cite:00ae20ff1b1d22d758b2f01ee4e53a7c093b07b3}}, {{cite:cfbb5f15daf02b78f679d13f5e51b7c9e497160c}}, {{cite:84314f25c2559f5018bc5ef3f54ad462809ab158}} methods map the input image directly to its crowd estimate. These methods also do not generalize well without any specialized and supportive mechanism. Wang et. al {{cite:00ae20ff1b1d22d758b2f01ee4e53a7c093b07b3}} used the AlexNet {{cite:671a04fcee60a742fc006871b858035e1362f8c1}} based scheme to directly regress and estimate the people number. Density-map estimation {{cite:75241d26c54f671e510ee124ff8f39001dc8fe0f}}, {{cite:518d411a60a5e5d510f3786a78e9c8a96586d904}}, {{cite:f670e9f440066e99de404ad78e49a2e62ce8c4e6}}, {{cite:e23bd2dd8485199481e5c81ce14bf8551509aaf1}}, {{cite:088930fbcbe89f4b7bee5fc1d45b7406381c38e6}}, {{cite:4e829d4f1e60e829b45a6f5d3bf220a2bf854979}} methods estimate the crowd-density value per pixel. The pixel-wise density values get added to produce the image's final crowd number. These methods first require the conversion of the ground-truth crowd dot-maps to their respective density-maps using a point-spread function (e.g. multivariate Gaussian distribution kernel) for the model training. However, the spread or kernel size remains a major issue that is vital for their training process quality. MCNN {{cite:75241d26c54f671e510ee124ff8f39001dc8fe0f}}, a density-map estimation method, uses three columns or branches, each with different filter sizes to account for the respective crowd-scale. These branches are merged together at the end to generate the final crowd density-map.",
"Sam et al. {{cite:f670e9f440066e99de404ad78e49a2e62ce8c4e6}} designed the Switch-CNN, another density-map method, that contains a (CNN-based) switch classifier that labels each input image or patch and routes it accordingly to one of three available specialized crowd regressors. Sindagi et al. {{cite:518d411a60a5e5d510f3786a78e9c8a96586d904}} devised a cascaded network that computes the crowd by first classifying the input 10-way based on its crowd-density level. Ranjan et al. {{cite:e23bd2dd8485199481e5c81ce14bf8551509aaf1}} designed a 2-branch network where the lower-resolution branch or column helps the higher-resolution for better crowd calculation. Li et al. {{cite:c779099c2af5bb4f4526a565e490667c3a50c745}} presented the CSRNet to obtain the context-based information through the dilation based deep layers for effective crowd estimation. Idrees et al. {{cite:2afeeba85586706d5389607756757b9a7617dee0}} designed the density-map estimation network that performs better for the crowd counting task by simultaneously solving the crowd localization as well as for the crowd-density estimation. Xu et al. {{cite:4e829d4f1e60e829b45a6f5d3bf220a2bf854979}} designed the density-map based method that used the patch-based density-maps, and grouped them into several levels based on the crowd-density, followed by the online learning mechanism for the automatic normalization process with a multi-polar center loss. Recently, Sajid et al. {{cite:cfbb5f15daf02b78f679d13f5e51b7c9e497160c}}, {{cite:84314f25c2559f5018bc5ef3f54ad462809ab158}} proposed a simple plug-and-play patch rescaling module (PRM) that rescales the input image or patch based on the respective crowd-density label.",
"Although recent works have produced very promising results, none of them make full use of their capabilities nor generalize well to hugely varying crowd-density and other major challenges. In this work, we focus on addressing the above crowd estimation challenges and issues.{{table:a4ab90cd-5b46-4149-a20f-2c647fe3e4fc}}"
] | [] |
Light Field Neural Rendering | Classical light field rendering for novel view synthesis can accuratelyreproduce view-dependent effects such as reflection, refraction, andtranslucency, but requires a dense view sampling of the scene. Methods based ongeometric reconstruction need only sparse views, but cannot accurately modelnon-Lambertian effects. We introduce a model that combines the strengths andmitigates the limitations of these two directions. By operating on afour-dimensional representation of the light field, our model learns torepresent view-dependent effects accurately. By enforcing geometric constraintsduring training and inference, the scene geometry is implicitly learned from asparse set of views. Concretely, we introduce a two-stage transformer-basedmodel that first aggregates features along epipolar lines, then aggregatesfeatures along reference views to produce the color of a target ray. Our modeloutperforms the state-of-the-art on multiple forward-facing and 360{\deg}datasets, with larger margins on scenes with severe view-dependent variations. | [
"Light field rendering.{{cite:4fce872e9ef26da434fd212af7742464466ca4b0}} defined the 4D light field as a function that specifies the radiance of any given ray in free space.They forwent geometric reasoning to directly synthesize novel views from input samples. Lumigraph rendering {{cite:8c1774a71fd88a187b97addf74fd7c479f6cb4b5}} exploits proxy geometry to counter aliasing effects that stem from irregularity or view under-sampling. Recent works {{cite:231834fe97bf474465c71859033afa81e2ec4e37}}, {{cite:9bba233ac504c7c6a6a4ceee3e22bb1a01420d19}}, {{cite:7c6636dfeb868974c3d4fea2a161297f58187bb8}}, {{cite:0739f32aad302d00d6775c95d3bbb56073348898}}, {{cite:efa0cf954cb44d36c3f24d904627e361816583ea}} have explored learning based methods to light field rendering. These methods, however, either require dense input sampling {{cite:231834fe97bf474465c71859033afa81e2ec4e37}}, have limited range of motion {{cite:9bba233ac504c7c6a6a4ceee3e22bb1a01420d19}} or are limited to simple scenes {{cite:7c6636dfeb868974c3d4fea2a161297f58187bb8}}.In this work, we focus on novel view synthesis forcomplex scenes with challenging non-Lambertian effects from a sparse set of viewpoints.",
"Neural scene representation.Representing shape and appearance of scenes using neural networks has recently gained immense popularity. Explicit representation-based methods use differentiable rendering to learn 3D representation such as point clouds {{cite:bd51bbe7a3fa6dcea2f9d36a32a5e86c24108016}}, {{cite:f1869ee4ffe9aa1d07f18c661c40352dacc9daa5}}, {{cite:3a423330b3d35a32384f8d4db5ad7071ef315b9b}}, meshes {{cite:c0c99e03c3e1f9e83d54b7b37bac57835bc9f00e}} or voxels {{cite:fdc906e8e1e2842ab3833120b4fcd43e86762d4c}}, {{cite:84cb1fdf4fa6cd5963dcbb74dde7f19dc5256b4c}} for the scene. Implicit representation-based methods represent scenes using continuous coordinate-based functions such as signed distance fields {{cite:ac491e1a647a364df578dd3dfa14536755217da2}}, {{cite:b4a88b15366e9d90cb00977c6224e63a40670f86}}, {{cite:8d0ece41a09caf45fb1b75d4633bdbb3e1200b4f}}, {{cite:264728974d13daa92f652415591a3df6449c20ed}}, {{cite:ee8b2e6aa97c98d9d5e40347722e70a1a2ba6da2}} or occupancy fields {{cite:fe5c64da18c9b09925d515540e9d71ce08ba84e6}}, {{cite:8be9cf497561f19aad6fcf24e9c9338e11aea570}}. Scene Representation Networks {{cite:b2d76454fcb1b2eed1f4a78c17215954d15d3fd7}} use a differentiable ray marching algorithm along with a continuous function that maps coordinates to features. NeRF {{cite:0092476f9d33a71109035ef2c2d2cdb10dcf4ef0}} achieves photo-realistic rendering by learning a function that maps points along a ray to color and opacity followed by volumetric rendering. NeX {{cite:1fc87a86f41966c14a7109a0dc536e80fd046338}} is a multiplane image-based scene representation that addresses NeRF's difficulty to model large view dependent effects. However, NeX is still challenged in scenarios such as interference patterns caused by reflection or refraction through liquid. In this work, we introduce a model that can faithfully render novel views in the presence of complex view-based effects in scenarios where other methods fail (, fig:introfigure). For a comprehensive survey of recent advances in neural rendering please refer to {{cite:f35bfa90ae81c6093642c0f66b1d8335e1b2ef41}}.",
"Image-based rendering.Image-based rendering (IBR) methods {{cite:4ed86f90f120c2ae579ac470a62aa54e451873f4}}, {{cite:4a829ae0eaa1de46d2d64d0184def3514740b4ba}}, {{cite:0397a6260f1c4a7856a778134532cfb430fb8087}}, {{cite:c567d56215a7a61a5e7aba4cdb8e28c7c301d50a}} are built on the notions that novel views can be render by “borrowing” pixel values from a given set of input images. Global geometry-based methods such as {{cite:644e728f83c24a0b5339f360e994d4d4fe787060}}, {{cite:f9453de8281c784cd17e7244263dc94ade366874}}, {{cite:520ad29e5fc27ba912b0f3980815e58b54cd405a}}, {{cite:38c3b1ad23d837fdd57f7316b05b4ef32fd69e7c}} rely on dense reconstruction from input views to obtain a global mesh for the scene. These meshes are used for projecting target rays onto nearby view for feature and color extraction. Other methods such as {{cite:473a017c0add4e51d7200ce1566fb5e0faa0956f}}, {{cite:c567d56215a7a61a5e7aba4cdb8e28c7c301d50a}} infer depth maps using multi-view stereo methods to compute warping transforms from the given input to target viewpoints. {{cite:87b5b8ca2f6fc7e404478799996420e99071bab4}} combine IBR with GAN-based image synthesis to learn view-dependent effects. To overcome difficulties caused by error in depth estimation, {{cite:f9248b91168a121308284fbf85b13c4935409d61}} estimate a depth uncertainty distribution to refine images.",
"Recently, {{cite:0397a6260f1c4a7856a778134532cfb430fb8087}} introduced IBRNet, a NeRF-based model that incorporates features from nearby views for rendering. Their architecture predicts colors for each point on the ray as weighted average of colors from neighboring views. The densities for each point are predicted by aggregating information from all other points using a single attention layer. Similarly for category-specific reconstruction, NerFormer {{cite:2130bddd730091689f90bff1faf0f38eca5e7b3a}} proposed to replace the MLP in NeRF-WCE {{cite:2615341d50d008969089caddf978a5aa4774bbab}} with a transformer model to allow for spatial reasoning.Our work crucially differs from these methods at their core: the rendering framework.Our method employs a light field representation, forgoing the need for volumetric rendering. Furthermore, we introduce a transformer-based model that first reasons about correspondences to aggregate features along each epipolar line, then reasons about occlusion and lighting effects to aggregate features from multiple views to produce the final color."
] | [] |
Distill and De-bias: Mitigating Bias in Face Verification using
Knowledge Distillation | Face recognition networks generally demonstrate bias with respect tosensitive attributes like gender, skintone etc. For gender and skintone, weobserve that the regions of the face that a network attends to vary by thecategory of an attribute. This might contribute to bias. Building on thisintuition, we propose a novel distillation-based approach called Distill andDe-bias (D&D) to enforce a network to attend to similar face regions,irrespective of the attribute category. In D&D, we train a teacher network onimages from one category of an attribute; e.g. light skintone. Then distillinginformation from the teacher, we train a student network on images of theremaining category; e.g., dark skintone. A feature-level distillation lossconstrains the student network to generate teacher-like representations. Thisallows the student network to attend to similar face regions for all attributecategories and enables it to reduce bias. We also propose a second distillationstep on top of D&D, called D&D++. Here, we distill the `un-biasedness' of theD&D network into a new student network, the D&D++ network, while training thisnew network on all attribute categories; e.g., both light and dark skintones.This helps us train a network that is less biased for an attribute, whileobtaining higher face verification performance than D&D. We show that D&D++outperforms existing baselines in reducing gender and skintone bias on theIJB-C dataset, while obtaining higher face verification performance thanexisting adversarial de-biasing methods. We evaluate the effectiveness of ourproposed methods on two state-of-the-art face recognition networks: ArcFace andCrystalface. | [
"Bias in face recognition: Empirical studies {{cite:53657c55f92d2efad4927ba35526bd90e885d7e0}}, {{cite:61688764fcea7639d726b4e14039474c89c9d9f4}}, {{cite:b1d4447c0c4dbf8264faac2d2367144d8d9aa000}}, {{cite:aea92abab3a81f7481f74c00a587de42a0b8007e}}, {{cite:3581ab320a05ef6493770182293bc6d4713fa155}} have shown that many publicly available systems performing face recognition or face analysis demonstrate bias towards sensitive attributes such as race and gender. With respect to gender bias, {{cite:292676e35e23bb6fa2cfa0be277d4025fedd0a74}}, {{cite:31734f65ca0b714d667ce74fedb5c5c548078774}} show that the performance of face recognition on females is lower than that of males. Use of cosmetics by females {{cite:56abcfee631917d0a14887a9a355a35f9a15a7cd}}, {{cite:f05857aec2f6b895cd62706c22a8e127c2c1f94a}} and gendered hairstyles {{cite:ed04e3acfc0975824be0bb86785241a776524a89}} has been assumed to play a major role in the resulting gender bias. However, {{cite:2380dec7effeb09f829c10591d70d1b4aec89127}} shows that cosmetics only play a minor role in the gender gap. {{cite:0c008554e4dac95002c75f7e2f23f356dbebbfda}}, {{cite:1d1107829975146ce137cc6fe88e6e355ea120d1}}, {{cite:69bf889a3d9f0cb73c9de876837c3c7a581c0cbf}}, {{cite:b2b30bb00ea3d3969b49f2d8b2fa4bcf1386e599}} explore the issue of racial bias in face recognition, and propose strategies to mitigate the same. {{cite:31734f65ca0b714d667ce74fedb5c5c548078774}} shows that face verification systems perform better on lighter skintones than darker skintones. {{cite:15b57ec327c85b43e35a58521a428e643b52dcca}}, {{cite:7dd30dd1ce82f5991225afa882b5af2412548166}} show that face recognition networks implicitly encode information about sensitive attributes during training, which might lead to bias w.r.t these attributes.",
"Building fairer datasets: It has been speculated that the imbalanced training datasets might lead to bias in face recognition. However, {{cite:292676e35e23bb6fa2cfa0be277d4025fedd0a74}} shows that the gender bias is not mitigated when equal number of male and female identities are used for training. {{cite:50d9d6473726219b730f6a7dc20f6e52c5db45cf}} introduces a dataset to measure the robustness of AI models to a diverse set of genders and skintones. Similarly, {{cite:e900d9bdd26993bee313ff31da7ee2922e5ebe5c}} presents an evaluation dataset that is balanced in terms of race and gender and provides the verification protocol for the same.{{table:b9363db7-8e00-42be-899e-ea0e619bca00}}",
"Adversarial de-biasing: Several researchers have proposed adversarial strategies to reduce the encoding of sensitive attributes (to reduce the bias with respect to these attributes), while performing a face-based target task. We provide a brief summary of these works in Table REF . {{cite:b806cb860619e843d42aad408933084419e0d061}}, {{cite:25f891a78437de5400b9cc50468131a13ecb779c}}, {{cite:f136fa2976f96969f00f24499937197596ff52ba}} have reported that the face verification performance of the adversarially-debiased systems in the target task decreases due to the removal of sensitive attributes such as gender.",
"Knowledge distillation (KD): KD {{cite:71d170345b82051b1dd636deda2c183a3b7435d0}} has been primarily applied to continual learning tasks {{cite:132ea9d3523b99dd648c4c3d34dacd80552afea2}}, {{cite:88f480e283b1d63db13cf7dad9f5b56664f76215}}. In a KD framework, a student network (initialized using a pre-trained teacher network) is trained to learn new classes or tasks that are not recognizable by the teacher network, and mimic the output score distribution of the teacher model for preventing forgetting of the classes/tasks that the teacher network was trained on. Several works {{cite:ba27acf1c6876f7f57d023ed92a12901d5468e9c}}, {{cite:46760f91310dabeecbaed7edd9c3508853bbbc56}}, {{cite:412606b2487ca71a63abd5001311ec0e3811d914}} have shown that directly matching feature activations of teacher and student networks from their intermediate layers is also an effective way to distill the teacher's knowledge into the student. Inspired by this, we employ feature-level KD in our work to enforce a student trained on a specific attribute category (e.g. males) to mimic a teacher that is trained on a different attribute category (e.g. females)."
] | [] |
Privacy Leakage over Dependent Attributes in One-Sided Differential
Privacy | Providing a provable privacy guarantees while maintaining the utility of datais a challenging task in many real-world applications. Recently, a newframework called One-Sided Differential Privacy (OSDP) was introduced thatextends existing differential privacy approaches. OSDP increases the utility ofthe data by taking advantage of the fact that not all records are sensitive.However, the previous work assumed that all records are statisticallyindependent from each other. Motivated by occupancy data in building managementsystems, this paper extends the existing one-sided differential privacyframework. In this paper, we quantify the overall privacy leakage when theadversary is given dependency information between the records. In addition, weshow how an optimization problem can be constructed that efficiently trades offbetween the utility and privacy. | [
"Analysis of potential privacy leakage and mitigation strategies have been active areas of research {{cite:a4bfd83dd6eb85fc897b2feb69f8de02753cb541}}, {{cite:32568f6818a9f47cd7b5715f443bca30cf02a1fc}}. Access control mechanisms for private data have been studied in {{cite:ca127b7813389683a79cba74f89ee71242897eaa}}, {{cite:6feb24e8c99aeec136d1faee9f23a5babeca5b1e}}. In {{cite:ca127b7813389683a79cba74f89ee71242897eaa}}, the existing role-based access control model is extended to incorporate complex privacy policies including purposes and obligations while detecting potential conflicts in privacy policies.Similarly in {{cite:6feb24e8c99aeec136d1faee9f23a5babeca5b1e}}, a privacy-preserving access control mechanisms for a relational database has been proposed by associating purpose information for each data element.",
"Differential privacy {{cite:cce5f13c2c87042602d7e71bf6b786f40c334949}} has gained attraction from the research community as a promising framework that provides provable privacy guarantees. In differential privacy, the query response is randomized in such a way that any query response would be statistically similar in the presence or absence of presence of any individual in the database. In addition to the rigorous theoretical privacy guarantees, differential privacy allows the user to tune the privacy parameter {{formula:0aa61ab6-93f7-48b1-94f1-a49d316ab83b}} to trade-off between utility and privacy.",
"Differential privacy has been applied to building applications in {{cite:781e1b90683e5fa755af3d1ca02ac97a705f2018}}, {{cite:d7d441cbab1304f55222ff97cc45c6b85e779f92}}, {{cite:375c8af44ede064b8b903c0d5467d4678c4a103b}} in the context of streaming occupancy data. In {{cite:375c8af44ede064b8b903c0d5467d4678c4a103b}}, utility of the occupancy data has been studied when varying values of {{formula:ab311003-3973-4bc3-8cdb-bc54bdc37b30}} have been applied to the data. While {{cite:375c8af44ede064b8b903c0d5467d4678c4a103b}} finds that information at the aggregated level are mainly preserved under differential privacy, preserving the utility of data at the individual level remains a challenge.",
"Recently, OSDP {{cite:e55c6ac472a2ff781612f2c89a551479a91c3dab}} has been proposed to overcome this challenge. However, the privacy guarantee when the records in the database are statistically dependent remains an open problem as noted in {{cite:e55c6ac472a2ff781612f2c89a551479a91c3dab}}."
] | [] |
Rank4Class: A Ranking Formulation for Multiclass Classification | Multiclass classification (MCC) is a fundamental machine learning problemwhich aims to classify each instance into one of a predefined set of classes.Given an instance, a classification model computes a score for each class, allof which are then used to sort the classes. The performance of a classificationmodel is usually measured by Top-K Accuracy/Error (e.g., K=1 or 5). In thispaper, we do not aim to propose new neural representation learning models asmost recent works do, but to show that it is easy to boost MCC performance witha novel formulation through the lens of ranking. In particular, by viewing MCCas to rank classes for an instance, we first argue that ranking metrics, suchas Normalized Discounted Cumulative Gain (NDCG), can be more informative thanexisting Top-K metrics. We further demonstrate that the dominant neural MCCarchitecture can be formulated as a neural ranking framework with a specificset of design choices. Based on such generalization, we show that it isstraightforward and intuitive to leverage techniques from the rich informationretrieval literature to improve the MCC performance out of the box. Extensiveempirical results on both text and image classification tasks with diversedatasets and backbone models (e.g., BERT and ResNet for text and imageclassification) show the value of our proposed framework. | [
"Modern deep neural networks for MCC converge to the same recipe: given an input instance, a neural encoder is learned to output scores for a set of classes. The vast research literature focus on developing more effective encoders in diverse domains, such as computer vision {{cite:57c98b3a960e46c80c58ebbb2e5d782204e931e2}}, {{cite:0f1587868ff5a72a7e1961fbd28e66ce0445300d}}, {{cite:ec4de64e69e9d01c4090b708d291b10ea3cf1f33}}, {{cite:09914f4b8d36b21375c8591b1c5e77170705368a}}, {{cite:37f15d646efd44312bd9ef779e6be4f0c0009543}}, natural language processing {{cite:509829522005ff1b7503d8315df017f823a59698}}, {{cite:1792b9906f0f101e06e40b9228fdcaaeb85f7ec8}}, {{cite:b1cf95b36724f4c1cd34a986d87343aa5629b3f2}}, and automatic speech recognition {{cite:331e5927ccebabdedfd62d4583c99195f4bbb723}}, among others.The softmax cross entropy loss is the dominant loss function discussed in these papers. Only very recently, {{cite:29c8a0c2bd4003f009a72cc10a19c36763409c1d}} study the mean squared error as another loss for MCC. Our work is orthogonal to the extensive research on neural encoders in that we provide a new formulation from the LTR perspective. Such a perspective inspires more diverse loss functions and model architectures that can model interactions between inputs and classes more effectively.",
"Learning to rank (LTR) is a long-established interdisciplinary research area at the intersection of machine learning and information retrieval {{cite:341f117cf9a939779a9e92a4af0f70f6dcef0234}}. Neural rankers are dominating in ranking virtually all modalities recently, including text ranking {{cite:e279015c94ecc3e19c7f3a3cbe02a7a6a713a05f}}, image retrieval {{cite:720aff45c50adc4b2322f430b3a72a11a5a1b74f}}, and tabular data ranking {{cite:a64f9ae39642eb0ba8cc8db034ec394c105a49cd}}. Many LTR papers focus on more effective loss functions {{cite:bd3ade8bfa428001d36856aafd06235ce0e22528}}, {{cite:a1a75357163c3bcfc074d43a6696fe3c5e5e890d}} to rank items with respect to a query. One of the focuses of this paper is to introduce new techniques stemmed from LTR to solve MCC problems.",
"Multi-label classification (MLC) differs from MCC in that there are more than one labels for each instance. MLC is generally treated as a different problem from classical MCC: the number of labels assigned to an instance could be arbitrary and one research focus is to decide the threshold to cutoff the prediction list. Its recently popular sub-field, extreme multi-label classification {{cite:baaa36d055a0ebd999b73762b4d235c9f3ccb464}}, also has a different and specific setting (e.g., the number of labels is huge) and research focus (e.g., efficiency to rank in the huge label space). LTR techniques have been proposed for MLC in the past {{cite:9ea50cfac75ac19b001ca4ea2504564edcd1a44c}}. While the perspective is similar to ours, the work is not in the deep learning setting and thus does not have the equivalence view of MCC in LTR. Our work makes the connection and also proposes extensions for MCC in the deep learning setting."
] | [] |
Reproducible and Portable Big Data Analytics in the Cloud | Cloud computing has become a major approach to help reproduce computationalexperiments because it supports on-demand hardware and software resourceprovisioning. Yet there are still two main difficulties in reproducing big dataapplications in the cloud. The first is how to automate end-to-end execution ofanalytics including environment provisioning, analytics pipeline description,pipeline execution, and resource termination. The second is that an applicationdeveloped for one cloud is difficult to be reproduced in another cloud, a.k.a.vendor lock-in problem. To tackle these problems, we leverage serverlesscomputing and containerization techniques for automated scalable execution andreproducibility, and utilize the adapter design pattern to enable applicationportability and reproducibility across different clouds. We propose and developan open-source toolkit that supports 1) fully automated end-to-end executionand reproduction via a single command, 2) automated data and configurationstorage for each execution, 3) flexible client modes based on user preferences,4) execution history query, and 5) simple reproduction of existing executionsin the same environment or a different environment. We did extensiveexperiments on both AWS and Azure using four big data analytics applicationsthat run on virtual CPU/GPU clusters. The experiments show our toolkit canachieve good execution performance, scalability, and efficient reproducibilityfor cloud-based big data analytics. | [
"There have been many studies on cloud-based reproducibility. Some of them {{cite:f1d11f96cc4afefbcdc4d90b6a147b140301bce7}}, {{cite:6604b8550729e52059435f93b2e6acb242d2dbd7}}, {{cite:7c24e6c41a659460cf0622b221099bf808662e2c}}, {{cite:b8f366db4bd30ca2a48cdacc658a6826a902817f}}, {{cite:e5094b7f96eb4be6d664677598bbd3bd908f44f4}} only study its conceptual frameworks. In this section, we only discuss those having actual systems. As shown in Table REF , we categorize capabilities of such systems into four types and group related work based on their capabilities.",
"Among the related studies in Table REF , nearly all related approaches achieve the software environment provision for reproducibility. However, approaches in group 1 mainly use the archived or containerized software environment, which limits the scope of applicability and lacks support for maintaining hardware configurations within cloud. Additionally, they monitor execution status by system commands or cloud APIs based periodical pulling which is less efficient than asynchronous event-based execution in our work. For example, ReproZip {{cite:9d3a2a72317a38e62ac4473f060f75f19efc888a}} tracks system commands and zips collected information along with all the used system files together for reproducibility. CARE {{cite:44ddfc2d1f799b4301c7851e2130c04e128a69f0}} reproduces a job execution by monitoring and archiving all the material required to re-execute operations. For related work in group 2, their proposed approaches encapsulate the code dependencies and software in virtual machine images or graphs, and enable history retrieval for reproduction. For instance, WSSE {{cite:7c24e6c41a659460cf0622b221099bf808662e2c}} proposes to generate digital data and source code snapshots to be reproduced and distributed within a cloud-computing provider. AMOS {{cite:6604b8550729e52059435f93b2e6acb242d2dbd7}} uses a VM containing a set of tools previously installed to implement a mechanism that initializes and configures VMs on demand. However, this reproduction is more like a history repetition, which is designed for verification and validation of history execution. They also provide configurable environment variables for automatic resource deployment in a single cloud, but do not support cross-cloud reproducibility. Instead, our proposed RPAC uses a data abstraction for information needed for reproducibility and transforms resource configurations used in one cloud into those in another cloud.",
"For the related studies in Table REF , group 3's capabilities are closest to ours. These approaches rely on annotated information provided by a user to assign workflow, and software/hardware environment. For example, Apt {{cite:16aa6a1aae5c7879f423964c78c2af2cba05bbe4}} uses user-provided profiles, which consists of a cluster with a control system to instantiate encapsulated experiment environments on that cluster, for repeating historical research. From this information, they deduce the required execution resources in cloud and then re-provision or configure them through their own APIs. In comparison, we use a serverless-based pipeline and follow cloud function APIs provided by cloud providers so the execution/reproduction process can be managed by the cloud without communications with toolkit. Also, their fully created annotations, even in cross-cloud reproduction, rely heavily on the users instead of execution history. Our work abstracts information required by users from information in execution history, users only need to provide minimal information to reproduce. Our toolkit will transform user-provided information into a piece of executable information. The event-based automation also achieves a more efficient execution. For cross-cloud reproducibility, we further use the adapter pattern model to achieve the configuration mapping without taking all inputs from the user."
] | [] |
Improving the performance of bagging ensembles for data streams through
mini-batching | Often, machine learning applications have to cope with dynamic environmentswhere data are collected in the form of continuous data streams withpotentially infinite length and transient behavior. Compared to traditional(batch) data mining, stream processing algorithms have additional requirementsregarding computational resources and adaptability to data evolution. They mustprocess instances incrementally because the data's continuous flow prohibitsstoring data for multiple passes. Ensemble learning achieved remarkablepredictive performance in this scenario. Implemented as a set of (several)individual classifiers, ensembles are naturally amendable for task parallelism.However, the incremental learning and dynamic data structures used to capturethe concept drift increase the cache misses and hinder the benefit ofparallelism. This paper proposes a mini-batching strategy that can improvememory access locality and performance of several ensemble algorithms forstream mining in multi-core environments. With the aid of a formal framework,we demonstrate that mini-batching can significantly decrease the reuse distance(and the number of cache misses). Experiments on six different state-of-the-artensemble algorithms applying four benchmark datasets with variedcharacteristics show speedups of up to 5X on 8-core processors. These benefitscome at the expense of a small reduction in predictive performance. | [
"Research on the parallelization of machine learning methods dates to the early 90s when most papers focused on batch ML methods that require the whole dataset in the main memory to train the model. In early data mining, the batch learning approach is applied by processing the whole training data (one or multiple times) to output the decision models. Then, the decision models can be applied to new production data. Such procedure is usually referred to by batch learning, batch-mode algorithms {{cite:b4191d50bace28ef06578b4887ccf0fd25bebbb5}}, static data mining, and others in the literature. From now on, we will use batch learning to refer to non-stream learning methods.",
"Over the years, with the increase in computational power, the focus shifted from single classifiers{{cite:8ad9a075870f86e24831d45e0e8f92e349447f74}}to ensembles.In the context of ensemble learners, many studies have been made on MapReduce (MR) frameworks{{cite:6fdf2dd86baa4f9c2a950d8c23434fc918f89bf4}}, {{cite:09cdb2c7134f1ee0189946fcc192e470523763f6}}, {{cite:23499c8d7ac81e5d180ca23ad0f7119a80ed6a8d}}, which are not suitable for applications with requirements of low response times, even being capable of processing huge amounts of data with high scalability {{cite:0d1d529d77342d5fc6580e20b3f9d6ca50195317}}. Another investigation approach is using GPUs to process ensembles {{cite:4d727006c11cdf5315af201ff1d21994a0c365c5}}, {{cite:f2afece13e3d9b25fb70f6a6ec3934aee6ca0fc8}}, {{cite:ffee9548054325eec3cae51ad4597916a3cd10dc}}, {{cite:1679620b82a711161547a65d1e8e48d13bd84b59}}. However, GPUs are better for data-parallel problems.",
"Among the works that explored multi-core parallelism,distributed or not, we can further subdivide it into batch{{cite:29778d4b443452002b69c47e44c2aa0a3b69f6e5}}, {{cite:0903cbe634048403ecf808c0a92e421c41440b9a}}, {{cite:8841062dcc7cea62ab3e2e7c593d943350edbb11}}, {{cite:4c10a786da619affec1aefd565b6aabf0e9cbdcf}}, {{cite:414e51692dc7e8ae13845b1cbf5df50ac15c2f86}}, {{cite:1679620b82a711161547a65d1e8e48d13bd84b59}}or data stream {{cite:222c095f7504270f04e3f00235a00b138f992648}}, {{cite:8ce58fadceff1ab20973cfcab4848723558cf739}}, {{cite:4f515296e004079bff8ebe22084e8fa9cc0c0c7e}}, {{cite:33cae0d32d9256bade2b23fc139ccdaa5c1b0a5c}} methods.Many works with various ensemble methods used the Message Passing Interface (MPI) standard, such as for ensembles of improved and faster Support Vector Machine (SVM) {{cite:0903cbe634048403ecf808c0a92e421c41440b9a}},bagging decision rule ensembles {{cite:222c095f7504270f04e3f00235a00b138f992648}} and regression ensembles {{cite:4f515296e004079bff8ebe22084e8fa9cc0c0c7e}}.The remaining literature presents a miscellaneous of tools and different scopes.In {{cite:29778d4b443452002b69c47e44c2aa0a3b69f6e5}}, multi-classifiers are implemented using OpenMP.In {{cite:8841062dcc7cea62ab3e2e7c593d943350edbb11}}, an ensemble of SVMs is implemented using joblib (a Python multiprocessing lib) and scikit-learn.In {{cite:1679620b82a711161547a65d1e8e48d13bd84b59}}, TensorFlow is used to build a scalable and extensible framework for ensembles parallelization.In {{cite:414e51692dc7e8ae13845b1cbf5df50ac15c2f86}}, an efficient Random Forest (RF) implementation that improves memory access due to better data representation on machines that combine both shared and distributed memory is proposed and implemented using FREERIDE (previous work from the authors).In {{cite:33cae0d32d9256bade2b23fc139ccdaa5c1b0a5c}}, an ensemble of J48 is parallelized for grid platforms using Java.In {{cite:8ce58fadceff1ab20973cfcab4848723558cf739}}, a low-latency Hoeffding Tree (HT) is implemented in C++ and used in RFs.In general, the related works mentioned so far differ from the present work in two main aspects: focusing on the implementation and performance aspects of specific ensemble methods or batch approaches (i.e., they do not focus on stream processing).",
"In more closely related work, Horak et al. {{cite:222c095f7504270f04e3f00235a00b138f992648}} study the impact of concurrency on memory access pattern and performance of ensembles. They proposed a two-stage bagging architecture that combines single-class recognizers with two-class discriminators to improve accuracy and allow parallel processing. They also addressed load balancing for the parallel classifier construction and used the algorithm SCALLOP as the base for the experiments to validate the architecture implemented in MPI.Martinovic et al. {{cite:4f515296e004079bff8ebe22084e8fa9cc0c0c7e}} enhanced a dynamic auto-tuning framework in a distributed fashion by using two strategies. They use a scalable infrastructure capable of leveraging the parallelism of the underlying platform for ensemble models to speed up the predictive capabilities while iteratively gathering production data. Results show that the approach implemented in MPI can learn the application knowledge by exploring a small fraction of the design space. Qian et al. {{cite:33cae0d32d9256bade2b23fc139ccdaa5c1b0a5c}} propose a novel ensemble for data stream classification. This solution maps different raw data to multiple grids, where the first-order geometric center is used to represent and classify data. This method performs data compression, which increases the accuracy and computational efficiency. It was implemented in Java and tested in both multi-core and grid environments.",
"Marrón et al. {{cite:8ce58fadceff1ab20973cfcab4848723558cf739}} propose an implementation of RF-based on vector SIMD instructions and changes the representation of Binary Hoeffding Trees (HT) to fit into the L1 cache. It was implemented in C++ and benchmarked against MOA and StreamDM using two real and eleven synthetic datasets. It is noteworthy that the authors compare the performance of a single tree and the ensemble using different hardware architectures.Actually, these last four works are more closely related to the present work as they approach the performance of ensembles in the context of data streams. However, they are different from the present work in the following aspects.The works in {{cite:222c095f7504270f04e3f00235a00b138f992648}}, {{cite:8ce58fadceff1ab20973cfcab4848723558cf739}} focus on specific algorithms, SCALLOP and Binary HT, respectively. The work in {{cite:4f515296e004079bff8ebe22084e8fa9cc0c0c7e}} leverages parallel processing to improve the algorithm's parameters, while {{cite:33cae0d32d9256bade2b23fc139ccdaa5c1b0a5c}} is focused on data compression.",
"The summary of related work regarding the parallelization of Machine Learning methods is shown in Table REF .{{table:b79aa0d6-eb14-41c6-8f03-457bf7654425}}"
] | [] |
Manifold embedding data-driven mechanics | This article introduces a new data-driven approach that leverages a manifoldembedding generated by the invertible neural network to improve the robustness,efficiency, and accuracy of the constitutive-law-free simulations with limiteddata. We achieve this by training a deep neural network to globally map datafrom the constitutive manifold onto a lower-dimensional Euclidean vector space.As such, we establish the relation between the norm of the mapped Euclideanvector space and the metric of the manifold and lead to a more physicallyconsistent notion of distance for the material data. This treatment in returnallows us to bypass the expensive combinatorial optimization, which maysignificantly speed up the model-free simulations when data are abundant and ofhigh dimensions. Meanwhile, the learning of embedding also improves therobustness of the algorithm when the data is sparse or distributed unevenly inthe parametric space. Numerical experiments are provided to demonstrate andmeasure the performance of the manifold embedding technique under differentcircumstances. Results obtained from the proposed method and those obtained viathe classical energy norms are compared. | [
"{{cite:08b13d2c8d712826b8130fa0fbccea1ffb65a80c}} propose a method based on the siamese architecture {{cite:e92fd2c8732f6b0073924c259b04f3c00691fe5a}} to approximately embed the Wasserstein metric into the Euclidean metric. This method finds the backward map from the Euclidean metric to the initial Wasserstein metric in an autoencoder fashion {{cite:96c3032b13424065ebcb0faffce54ca2ff520c44}} that is approximately invertible. A similar idea is used for point cloud embedding to speed-up Wasserstein distance calculations {{cite:a8093b9415ff2a4d6f2cfa5f0f9aec723b2d8782}}. {{cite:30b4ec06c6623a7e798063de2c5e749b36a5f333}} embed sub-samples of datasets with an arbitrary metric into the Euclidean metric to address mode collapse for generative adversarial networks (GANs) {{cite:d8e9a295f6533559ad7a0c91f9539ee10c4094be}}. {{cite:c10ff395d62b9b7f78b8fe52cdec0270eb4178db}} aim to parametrize Poincaré map by a deep autoencoder architecture that enforces invertibility as a soft constraint.",
"{{cite:1567c7ff0364e7a1920e3a7c05db00ceb2c2daed}}, {{cite:d97efb249fff36fbbf6d5b741dfc776e0dacf57e}}, {{cite:efd4f16ef133244af67ec2fbd3f5b7f46055bb18}} use deep autoencoder architecture to find a coordinate transformation that linearizes a nonlinear partial differential equation (PDE). {{cite:acfbb7e88aa24f03a36c083de1fcbf3b25335047}} use deep convolutional autoencoders to project a dynamical system onto a nonlinear manifold of fewer dimensions instead of the classical approach of finding a linear subspace. {{cite:ed077943e79226e3b6be17f0c1c47a94e79b61ea}} construct a reduced-order model based on the autoencoder architecture with shallow multilayer perceptrons (MLPs) to find a proper linear subspace for a nonlinear PDE that describes advection-dominated phenomena.To the best of our knowledge, the current work is the first paper that leverages the bijectivity of the invertible neural network to generate the desired embedding space for the data-driven paradigm."
] | [] |
Playing Against Fair Adversaries in Stochastic Games with Total Rewards | We investigate zero-sum turn-based two-player stochastic games in which theobjective of one player is to maximize the amount of rewards obtained during aplay, while the other aims at minimizing it. We focus on games in which theminimizer plays in a fair way. We believe that these kinds of games enjoyinteresting applications in software verification, where the maximizer playsthe role of a system intending to maximize the number of "milestones" achieved,and the minimizer represents the behavior of some uncooperative but yet fairenvironment. Normally, to study total reward properties, games are requested tobe stopping (i.e., they reach a terminal state with probability 1). We relaxthe property to request that the game is stopping only under a fair minimizingplayer. We prove that these games are determined, i.e., each state of the gamehas a value defined. Furthermore, we show that both players have memoryless anddeterministic optimal strategies, and the game value can be computed byapproximating the greatest-fixed point of a set of functional equations. Weimplemented our approach in a prototype tool, and evaluated it on anillustrating example and an Unmanned Aerial Vehicle case study. | [
"Stochastic games with payoff functions have been extensively investigated in the literature. In {{cite:741de45d7c19c9971023e666dcba85d6b514730e}}, several results are presented about transient games,a generalized version of stopping stochastic games with total reward payoff.In transient games, both players possess optimal (memoryless and deterministic) strategies.Most importantly, the games are determined and their value can be computed as the least fixed point of a set of equations.Most of these results are based on the fact that the {{formula:e0d89bea-7389-4721-b40d-1a52d0d45c6b}} functional (see Section ) for transient games has a unique fixed point.Notice that in this paper we have dealt with games that are stopping only under fairness assumptions. Thus, the corresponding functionalmay have several fixed points. Hence, the main results presented in {{cite:741de45d7c19c9971023e666dcba85d6b514730e}} do not apply to our setting.",
"{{cite:26c3cce100b6ecf8d11c2e2dcdd015821319c406}} and {{cite:1f83c450e7af98164831d8a3c4f2c6b36e79de57}} present logical frameworks for the verification and synthesis of systems. While {{cite:26c3cce100b6ecf8d11c2e2dcdd015821319c406}} provides a solution for a probabilistic branching temporal logic extended with expected total, discounted, and average reward objective functions, {{cite:1f83c450e7af98164831d8a3c4f2c6b36e79de57}} does the same in a similar extension of a probabilistic linear temporal logic. Both frameworks were implemented in the tool Prism {{cite:4e9bdbe05091aa91d0d5e942d858fff0fb422466}}. Although a vast class of properties can be expressed in these frameworks, none of them are presented under fair environments. In fact, these works are on stochastic multiplayer games in which each player is treated equally.",
"Stochastic shortest path games {{cite:6929ab8f14f0fabfdff9c5d4bdf51c7e16def493}} are two-player stochastic games with (negative or positive) rewards in which the minimizer's strategies are classified into proper and improper,proper strategies are those ensuring termination. As proven in {{cite:6929ab8f14f0fabfdff9c5d4bdf51c7e16def493}}, these games are determined, and both players posses memoryless optimal strategies. To prove these results, the authors assume that the expected game value for improper strategies is {{formula:919c876e-9a98-49a5-bdd8-91cbac197265}} , this ensures that the corresponding functional is a contraction and thus it has a unique fixpoint. In contrast, we restrict ourselves to non-negative rewards but we do not make any assumptions over unfair strategies, as mentioned above the corresponding functional for our games may have several fixpoints. Furthermore, we proved that the value of the game is given by the greatest fixpoint of {{formula:45714cc6-5841-407e-abae-91d0219b82ba}} . In recent years, several authors have investigated stochastic shortest path problems for MDPs (i.e, one-player games), where the assumption over improper strategies is relaxed (e.g., {{cite:fad4bc485c39f7cf99560486701081d1ad617c3f}}); to the best of our knowledge, these results have not be extended to two-player games.",
"In {{cite:ab7781a056b06fa6d79bd43b564cdf92f3b6942d}} the authors tackle the problem of synthesizing a controller that maximizes the probability of satisfying an LTL property. Fairness strategies are used to reduce this problem to the synthesis of a controller maximizing a PCTL property over a product game. However, this article does not address expected rewards and game determinacy under fairness assumptions.",
"Interestingly, in {{cite:3f24f5e84071deeeb6c082e77b0e05f5333868df}} the authors consider the problem of winning a (non-stochastic) two-player games with fairness assumptions over the environment. The objective of the system is to guarantee an {{formula:b3f61f2b-1feb-4fcf-8cb1-4eed78cf6b9a}} -regular property. The authors show that winning in these games is equivalent to almost-sure winning in a Markov decision process. It must be noted that this work only considers non-stochastic games. Furthermore, payoff functions are not considered therein."
] | [] |
A data-centric weak supervised learning for highway traffic incident
detection | Using the data from loop detector sensors for near-real-time detection oftraffic incidents in highways is crucial to averting major traffic congestion.While recent supervised machine learning methods offer solutions to incidentdetection by leveraging human-labeled incident data, the false alarm rate isoften too high to be used in practice. Specifically, the inconsistency in thehuman labeling of the incidents significantly affects the performance ofsupervised learning models. To that end, we focus on a data-centric approach toimprove the accuracy and reduce the false alarm rate of traffic incidentdetection on highways. We develop a weak supervised learning workflow togenerate high-quality training labels for the incident data without the groundtruth labels, and we use those generated labels in the supervised learningsetup for final detection. This approach comprises three stages. First, weintroduce a data preprocessing and curation pipeline that processes trafficsensor data to generate high-quality training data through leveraging labelingfunctions, which can be domain knowledge-related or simple heuristic rules.Second, we evaluate the training data generated by weak supervision using threesupervised learning models -- random forest, k-nearest neighbors, and a supportvector machine ensemble -- and long short-term memory classifiers. The resultsshow that the accuracy of all of the models improves significantly after usingthe training data generated by weak supervision. Third, we develop an onlinereal-time incident detection approach that leverages the model ensemble and theuncertainty quantification while detecting incidents. Overall, we show that ourproposed weak supervised learning workflow achieves a high incident detectionrate (0.90) and low false alarm rate (0.08). | [
"Automatic incident detection on traffic networks has been a research topic for decades. Starting in the 1970s, algorithms were developed to detect potential incidents on the traffic networks. These AID algorithms were based mostly on the information provided by the inductive loop detectors embedded in the freeway pavement, from which traffic occupancy and volume profiles could be recorded. Accordingly, with the usage of occupancy, the California algorithm {{cite:b070dd0a0fb91c5ef152c386973fdb9f70364c30}} and its variations were developed. With the installations of dual-loop detectors that can record traffic speed information, the detecting methods were expanded. In general, these algorithms can be categorized into pattern-based algorithms, catastrophe-based methods, and statistical methods {{cite:efcf8bf08a4c1f544624c72f0c8491d839927ff2}}. With abundant data available from the traffic systems nowadays, data-driven methods, such as ML model-based AID frameworks, have been receiving increased attention.",
"Traffic data, regardless of the type, can be complicated and noisy because of sophisticated road layouts, human recording of incidents, and faulty sensors. Using ML models to achieve detection from the traffic data has shown great advantages thanks to their ability to learn complex mappings from data. In this section we discuss prior work on using ML models to conduct AID.",
"Several studies focused on detecting whether there is an incident or congestion on a network level, where the detection model takes in the information of the entire traffic network of interest and decides whether an incident is happening within the network without specifying its location. {{cite:5824d7dfe9cce190c5525ea63eb0339295b6ffaa}} adopted an RNN combined with a restricted Boltzmann machine to detect network-wide congestion based on GPS data collected from taxis. {{cite:e718424961419e8405986aa4e623ed9d54e06dc8}} developed a CNN-based model to detect incidents in urban traffic networks, where the GIS information was adopted to match the measured traffic parameters and recorded incidents. {{cite:cd1caefe5b513f74b0cde5d28354e3652b07c35e}} developed a spatiotemporal graphical model called a spatiotemporal pattern network to predict traffic dynamics in large networks and detect incidents by calculating anomaly scores.",
"As opposed to network-level detection, sensor-level detection emphasizes the usage of local information. {{cite:4e9d95dab96fb87f8da05c6cc2b7434697c3d1fd}} developed an AID model where the real-time traffic data were compared with the traffic parameter estimation using a moving average, and then incidents were detected according to the difference. {{cite:1394665d25e3da3b418c0cada869de82c86e13eb}} used an SVM to classify incident or non-incident instances given traffic speed, flow, and occupancy information. {{cite:0151294e1659680b2cef6dd40b01713dba7c3f07}} utilized the traffic information on the nodes of interest and their neighbors to construct a spatiotemporal model using CNN and RNN to detect congestion. This work is easily scalable because only the information on a few neighbor nodes is needed. {{cite:895be9173a0f41712b6c8923478710b582070b0f}} incorporated vehicle-to-infrastructure communications with the traffic parameters from loop detectors to detect incidents based on distance and time of changing lanes. {{cite:7aa9f56bf10db27d1acff882798f131d39ca1fea}} used a stacked autoencoder to generate a high-level representation of traffic surveillance videos and adopted one-class SVM to detect potential traffic accidents. {{cite:16257a39fb6afec99b7bdea7af0b455b85113691}} proposed an ensemble of k-nearest neighbors and SVM to classify instances with speed, flow, and occupancy information into incident or non-incident. {{cite:8c4887c918f75ab952468f5731fd9ddfe7955c67}} used CNN and extreme gradient boosting to detect incidents in expressways with data from microwave detectors.",
"Network-level detection models rely on the specific traffic network they have been trained on, which cannot be used for a different setting. Thus, we aim to develop a sensor-level detection model so that it has the potential of being deployed at different locations. However, the senor-level detection methods described above required various types of data, were evaluated on synthetic data, or did not provide any model prediction uncertainty measure. Also, data quality, which can play an essential role in training effective models, was not mentioned in the previous studies. Our work addresses these issues. Specifically, we adopt a data-centric approach to utilize real-world loop detector speed measurements, curate preprocessed data to create high-quality datasets, develop RNN-based classifiers, and create a deep ensemble model to quantify the model predictive uncertainty."
] | [] |
Time-Aware Neighbor Sampling for Temporal Graph Networks | We present a new neighbor sampling method on temporal graphs. In a temporalgraph, predicting different nodes' time-varying properties can require thereceptive neighborhood of various temporal scales. In this work, we propose theTNS (Time-aware Neighbor Sampling) method: TNS learns from temporal informationto provide an adaptive receptive neighborhood for every node at any time.Learning how to sample neighbors is non-trivial, since the neighbor indices intime order are discrete and not differentiable. To address this challenge, wetransform neighbor indices from discrete values to continuous ones byinterpolating the neighbors' messages. TNS can be flexibly incorporated intopopular temporal graph networks to improve their effectiveness withoutincreasing their time complexity. TNS can be trained in an end-to-end manner.It needs no extra supervision and is automatically and implicitly guided tosample the neighbors that are most beneficial for prediction. Empirical resultson multiple standard datasets show that TNS yields significant gains on edgeprediction and node classification. | [
"There has been a burst of remarkable work for representation learning on static graphs {{cite:0b932208550542de4b6753f685e372c03960c4f3}}, {{cite:b7c13e02c19f79dd9f12b5a45e24faa03483adb4}}, {{cite:1327dbee3681d6595998aaf0ab0099e0c7f432c1}}, {{cite:2322533aff4193515da62a1679161607015effa3}}, {{cite:0424e74197afb327eca47b4b27770c5a64567d6a}}, {{cite:df23362e9fd66ce22470fbf8d3cd3b714bf1604a}},but the work on temporal graphs is much sparser.There exist two main classes of temporal graphs: discrete-time dynamic graphs (DTDG) {{cite:96b4d97fbf34e40b84bf385f5f4fd19375e98f18}} and continuous-time dynamic graphs (CTDG) {{cite:6424b3001d44839bc49a12dfb349db83b026958e}}.DTDG are sequences of static graph snapshots taken at intervals in time, while CTDG can be represented as timed lists of events.Representation learning on CTDG is more flexible, general and challenging, which is the focus of this paper.",
"Early models for temporal graph learning focus on DTDGs {{cite:13195e53ca16f10009c49fe94051dce52deb1946}}, {{cite:2dc3b7b95335950a176557344a2ef11166c40c24}}, which aggregategraph snapshots and then apply static methods {{cite:02131c2133dacbbea0a3ff68c502b8a7bff0e31f}}, {{cite:96b4d97fbf34e40b84bf385f5f4fd19375e98f18}}, {{cite:dca31f48506b11eed6e1088715036d3b946036d8}},or encode each snapshot to produce a series of embeddings {{cite:fdd119a9c8ace7e45fc7d4ffa43fabd1b1204028}}, {{cite:f6846eb8c14803d6f9e7e7bb8a1cc8af6e55e8ea}}.More recently, some work consider the CTDGs {{cite:5f8ffb07024d317331cba684b6befdbc4467d12a}}, {{cite:59a0dd493e4f8eb809774d37cd1fd28180f4af77}}.{{cite:dee8195f233d7bc33edccc1cbfb1adbcc79dc829}}, {{cite:34afa8280432f5ef4a6cbb6697ce5da6116da7bb}}, {{cite:6424b3001d44839bc49a12dfb349db83b026958e}} aggregates the messages from neighborsthrough a temporal aggregation module, which present superior performance.To prevent the number of neighbors from growing without limitation as time flows, {{cite:34afa8280432f5ef4a6cbb6697ce5da6116da7bb}} uniformly samples the neighbors to improve the efficiency and stabilization.{{cite:6424b3001d44839bc49a12dfb349db83b026958e}} improves the neighbor sampling further by incorporating the temporal information, which samples the most recent neighbors.Our work proposes new approaches for utilizing temporal information for neighbor sampling.First, our expanded neighbor sampling approachreduces information redundancyand extendsthe temporal scale ofreceptive neighborhoods by inserting spacebetween the sampled neighbors alongthe time axis.Second, beyond the unified sampling strategy,our TNS method provides the adaptive receptive neighborhood for every node at any time in a learnable way.Our methods improves the effectiveness of popular TGNs without increasing their time complexity."
] | [] |
Multimeasurement Generative Models | We formally map the problem of sampling from an unknown distribution with adensity in $\mathbb{R}^d$ to the problem of learning and sampling a smootherdensity in $\mathbb{R}^{Md}$ obtained by convolution with a fixed factorialkernel: the new density is referred to as M-density and the kernel asmultimeasurement noise model (MNM). The M-density in $\mathbb{R}^{Md}$ issmoother than the original density in $\mathbb{R}^d$, easier to learn andsample from, yet for large $M$ the two problems are mathematically equivalentsince clean data can be estimated exactly given a multimeasurement noisyobservation using the Bayes estimator. To formulate the problem, we derive theBayes estimator for Poisson and Gaussian MNMs in closed form in terms of theunnormalized M-density. This leads to a simple least-squares objective forlearning parametric energy and score functions. We present variousparametrization schemes of interest including one in which studying GaussianM-densities directly leads to multidenoising autoencoders--this is the firsttheoretical connection made between denoising autoencoders and empirical Bayesin the literature. Samples in $\mathbb{R}^d$ are obtained by walk-jump sampling(Saremi & Hyvarinen, 2019) via underdamped Langevin MCMC (walk) to sample fromM-density and the multimeasurement Bayes estimation (jump). We studypermutation invariant Gaussian M-densities on MNIST, CIFAR-10, and FFHQ-256datasets, and demonstrate the effectiveness of this framework for realizingfast-mixing stable Markov chains in high dimensions. | [
"Despite being a fundamental framework for denoising, empirical Bayes has been surprisingly absent in the DAE/DSM literature. {{cite:0b2b8a8261cb4b6eaa51dd706eacac25aae3f070}} first mentioned the connection between empirical Bayes and score matching, but this work is practically unknown in the literature; for much of its development, DSM was all tied to DAEs {{cite:ae5eedbfd1b091bb8754a9407e27edf038ad18b8}}, {{cite:1de7969c4a6e20cfc4f69ed81f14d9b9b1599efb}}; see {{cite:a8d0b1397ab80eb7b63729497daa510d9ac30f7b}} for an extensive survey. Abstracting DSM away from DAE is due to {{cite:9090afaf807421de1f4e8acb07d359df4e56c57d}} where they directly parametrized the energy function. In this work, we closed the circle, from empirical Bayes to DAEs. One can indeed use the MDAE objective and learn a generative model: a highly generalized DAE naturally arise from empirical Bayes. The important subtlety here is, as we remarked in sec:score-matching, we can only arrive at this connection starting from our generalization of empirical Bayes, not the other way around. Of course, this core multimeasurement aspect of our approach, without which we do not have a generative model, does not exist in the DAE literature.There are similarities between Eq. 5 in {{cite:1de7969c4a6e20cfc4f69ed81f14d9b9b1599efb}} and eq:g here. However, one is on barely noisy data ({{formula:684b1ed9-f3b9-422d-bf0d-a1a10dc76c70}} ) in {{formula:71d08f93-8472-4673-bcc1-0b6116e88dd9}} , the other on multimeasurements in {{formula:ad5ab0a6-6114-4cc9-a172-f063bc13b24c}} for any (high) noise level {{formula:28ce439a-3041-44e7-b6b2-ab9be381d3e0}} . There, they arrived at Eq. 5 (with some effort) starting from a DAE objective. Here, we start with eq:g (it defines the score function) and arrive at MDAE objective (eq:mdae), in one line of algebra.",
"To address the problem of choosing a noise level in DSM {{cite:9090afaf807421de1f4e8acb07d359df4e56c57d}}, {{cite:1c3e454d02a5921cb7ba95c8bdcb6ec6e37ced39}} studied it with multiple noise levels by summing up the losses using a weighing scheme. See {{cite:17412da3ae1090d004f5cefcdaae770d768c7364}}, {{cite:f0792060ff2d14cd9fbdf05f8cf988422dbc5175}}, {{cite:8698d367678d3b2d6db534ff9e678df03408a117}}, {{cite:b55dfa1dbe95a449bed2da45e4a1a8b8b38357a2}}, {{cite:daa8db184f1c906bbdb06c2a4af2e7ff4e8b3233}} in that direction. The learning objectives in these models are based on heuristics in how different noise levels are weighted. Denoising diffusion models {{cite:396a575df7a4878c361aa3d28a4b5b8c701204db}}, {{cite:ae02f92b4c5713b5093db0e83381cc6bda8278c4}}, {{cite:6d984f92cbf67992980b19e12c3718c99b63f701}} follow the same philosophy while being theoretically sound. Sampling in these models are based on annealing or reversing a diffusion process. Our philosophy here is fundamentally different. Even noise levels have very different meaning here, associated with the M noise channels of the factorial kernel. All noise levels (which we took to be equal in later parts, a meaningless choice in other methods) are encapsulated in a single energy/score function. Using Langevin MCMC we only sample highly noisy data {{formula:68b6fff0-7e80-43b4-b4af-c015154e174f}} . Clean data in {{formula:a7660b00-815f-4b01-84be-0cefb05d826a}} is generated via a single step by computing {{formula:11325f3d-7cae-4c09-9d13-5db5e88f1091}} ."
] | [] |
Beyond Single-Deletion Correcting Codes: Substitutions and
Transpositions | We consider the problem of designing low-redundancy codes in settings whereone must correct deletions in conjunction with substitutions or adjacenttranspositions; a combination of errors that is usually observed in DNA-baseddata storage. One of the most basic versions of this problem was settled morethan 50 years ago by Levenshtein, or one substitution, with nearly optimalredundancy. However, this approach fails to extend to many simple and naturalvariations of the binary single-edit error setting. In this work, we makeprogress on the code design problem above in three such variations: We construct linear-time encodable and decodable length-$n$ non-binary codescorrecting a single edit error with nearly optimal redundancy $\logn+O(\log\log n)$, providing an alternative simpler proof of a result by Cai,Chee, Gabrys, Kiah, and Nguyen (IEEE Trans. Inf. Theory 2021). This is achievedby employing what we call weighted VT sketches, a notion that may be ofindependent interest. We construct linear-time encodable and list-decodable binary codes withlist-size $2$ for one deletion and one substitution with redundancy $4\logn+O(\log\log n)$. This matches the existential bound up to an $O(\log\log n)$additive term. We show the existence of a binary code correcting one deletion or oneadjacent transposition with nearly optimal redundancy $\log n+O(\log\log n)$. | [
"Recently, there has been a flurry of works making progress in coding-theoretic questions analogous to the ones we consider here in other extensions of the binary single-edit error setting.",
"A line of work culminating in {{cite:93b4eb69444ef3acf2963a350a75a92b787855b2}}, {{cite:8aa80e4d37ddae0295465a1b033160febe6013f2}}, {{cite:f1ddc2adf6d0cf18918c1c4b6ff645869b45fa77}} has succeeded in constructing explicit low-redundancy codes correcting a constant number of worst-case deletions. Constructions focused on the two-deletion case have also been given, e.g., in {{cite:f1ddc2adf6d0cf18918c1c4b6ff645869b45fa77}}, {{cite:4a8cf3a79d09e160c543637156a4149e0e300e36}}, {{cite:8aa80e4d37ddae0295465a1b033160febe6013f2}}.Explicit binary codes correcting a sublinear number of edit errors with redundancy optimal up to a constant factor have also been constructed recently {{cite:17b569595d475aed3cc1036299a22ae16d98806a}}, {{cite:fe0eaff01e5a578dbc2f11cafdfb177a1466219b}}.Other works have considered the related setting where one wishes to correct a burst of deletions or insertions {{cite:1076d7651c6db2f922265fc432c9599848da76a9}}, {{cite:a6f84be7c8efd45c5367f7442ffc07e2fb3fca34}}, {{cite:e9cc046f4cf38f87efd3d06c460c288a5630129a}}.Following up on {{cite:95fafdebcf34f09b0f71f5b467f690cdcf98870e}}, codes correcting a combination of more than one deletion and one substitution were given in {{cite:3cd5619c8b29d27fab7ad1497ce58fda49e1c2c9}} with sub-optimal redundancy.",
"List-decodable codes in settings with indel errors have also been considered before.For example, Wachter-Zeh {{cite:ec686619334078cb1aa363fabb9345136477d60c}} and Guruswami, Haeupler, and Shahrasbi {{cite:2bbe76cdedd20c9071e55dfad0dc289558925d4f}} study list-decodability from a linear fraction of deletions and insertions.Most relevant to our result in Section REF , Guruswami and Håstad {{cite:8aa80e4d37ddae0295465a1b033160febe6013f2}} consideredconstructed an explicit list-size two code correcting two deletions with redundancy {{formula:7cd58c77-e401-47d0-ac2d-0521f298261b}} , thus beating the greedy existential bound in this setting.",
"The interplay between deletions and transpositions has also been considered before.Gabrys, Yaakobi, and Milenkovic {{cite:715563e1387dd87bd8a31229ecfd07f2d078e44a}} construct codes correcting a single deletion and many adjacent transpositions.In an incomparable regime, Schulman and Zuckerman {{cite:fd74f2671714c16079de4fb847907bbdb8db57a7}}, Cheng, Jin, Li, and Wu {{cite:2ab0737ff16efb4b85531a41a59eb62f9340b197}}, and Haeupler and Shahrasbi {{cite:602fd0846828962d07606a91a087e36e57f48a06}} construct explicit codes with good redundancy correcting a linear fraction of deletions and insertions and a nearly-linear fraction of transpositions."
] | [] |
Serverless data pipeline approaches for IoT data in fog and cloud
computing | With the increasing number of Internet of Things (IoT) devices, massiveamounts of raw data is being generated. The latency, cost, and other challengesin cloud-based IoT data processing have driven the adoption of Edge and Fogcomputing models, where some data processing tasks are moved closer to datasources. Properly dealing with the flow of such data requires building datapipelines, to control the complete life cycle of data streams from dataacquisition at the data source, edge and fog processing, to Cloud side storageand analytics. Data analytics tasks need to be executed dynamically atdifferent distances from the data sources and often on very heterogeneoushardware devices. This can be streamlined by the use of a Serverless (or FaaS)cloud computing model, where tasks are defined as virtual functions, which canbe migrated from edge to cloud (and vice versa) and executed in an event-drivenmanner on data streams. In this work, we investigate the benefits of buildingServerless data pipelines (SDP) for IoT data analytics and evaluate threedifferent approaches for designing SDPs: 1) Off-the-shelf data flow tool (DFT)based, 2) Object storage service (OSS) based and 3) MQTT based. Further, weapplied these strategies on three fog applications (Aeneas, PocketSphinx, andcustom Video processing application) and evaluated the performance by comparingtheir processing time (computation time, network communication and disk accesstime), and resource utilization. Results show that DFT is unsuitable forcompute-intensive applications such as video or image processing, whereas OSSis best suitable for this task. However, DFT is nicely fit forbandwidth-intensive applications due to the minimum use of network resources.On the other hand, MQTT-based SDP is observed with increase in CPU and Memoryusage as the number of...<truncted to fit character limit in Arxiv> | [
"A number of SDP architectures and solutions have been proposed in the field of IoT data management in edge, fog and cloud environments. This section briefly summarizes the recent work done in the context of SDP architectures and models.",
"The public cloud service providers such as AWS greengrass {{cite:49afe7f7ef69ba240b350564514c4fff33857deb}}, Google Cloud IoT {{cite:f4aa16281fa5486dd7e00784730cd636c127e86d}} and Microsoft- Azure IoT Edge {{cite:3444ab54261e93dae05c40bb5c8b3219d5834fe8}} have typical IoT data pipeline solutions for industrial, healthcare, smart city and other real time use cases. For example, consider AWS IoT Greengrass, where Lambda service will be executed at the edge layer for data acquisition and pre-processing. Later data is forwarded to the cloud by edge devices and then it passes through pipeline of activities for post processing and finally is delivered to the data sink.",
"blueValeria et al. proposed a solution of IoT data stream processing in distributed fog and edge computing environments with decentralized scaleable manner {{cite:db72238a04274befdb7688528b2bdbedfcfc6ac7}} and further extended to how data processing operators were placed in computing nodes considering the efficiency, application topology and resources configurations {{cite:f4b813a494c261ee48dba1e3b5f6ef46c546a4ba}}. These works provide hint that off-the-shelf data stream processing tools such as Apache Storm can be used for the task. However, these stream processing tools require huge computing clusters and in IoT deployments more often devices are heterogeneous with limited computing capacity. More often IoT workloads are event and time driven which motivates us to investigate the serverless based data processing pipelines. Further, SDPs easily been deployed at various levels in the IoT hierarchy (Edge, Fog and Cloud Infrastructure) with efficient granular scaling of the serverless functions.",
"Das et al. {{cite:b1cd5491dfc02bcf20b8f35463727307563d90f1}} proposed a model for efficient execution of user tasks as serverless functions in edge/cloud environments and designed a set of data pipelines using AWS Greengrass on edge devices along with Lambda capabilities. Our approach looks similar to this model, however it lacks fog based processing pipeline model.",
"Dehury et al. {{cite:894861e72c8e101084eb26b33bc134ba744ecc84}} designed a framework known as CCoDaMiC, which aims to ensure data accuracy, trustworthiness, and validation in SDP. This work directly relates to our proposed DFT based SDP approach. However, CCoDaMiC mainly focuses on data accuracy and trustworthiness and not on the performance of the applications. Lixiang et al. {{cite:0b342f52f71608dd3691639c75593206e9d0ebd4}} designed a framework for video processing using serverless lambda functions known as Sprocket. Authors demonstrated the efficiency of serverless functions for faster execution by constructing pipeline of activities for video handling. However, their primary focus is to reduce latency and cost by using the techniques of parallelism. Interestingly, this work motivated us to consider the complex video processing use case in our proposed research.",
"Several techniques and methods have been proposed illustrating the use of MQTT for data acquisition from different data sources via publish/subscribe model {{cite:f190f3673851ad6e058dd5b9e604c28938363975}}, {{cite:e87b95d351503e5f23cbf7b0afc2748bfc693c76}}, {{cite:5ea722b5b8bb6c6ee52a18b840ece5d88244a571}}. MQTT brokers can act as data carriers and can store data until subscribers consume it. This approach is well suited to store temporary or intermediate data between processing elements in SDP. In our work, one of the approaches uses MQTT together with serverless framework to construct data pipelines from data source to sink.",
"Thus to the best of our knowledge, none of the research works attempted to investigate and compare the different techniques in the construction of SDP with different approaches for intermediate data handling between serverless functions."
] | [] |
QU-BraTS: MICCAI BraTS 2020 Challenge on Quantifying Uncertainty in
Brain Tumor Segmentation - Analysis of Ranking Scores and Benchmarking
Results | Deep learning (DL) models have provided state-of-the-art performance invarious medical imaging benchmarking challenges, including the Brain TumorSegmentation (BraTS) challenges. However, the task of focal pathologymulti-compartment segmentation (e.g., tumor and lesion sub-regions) isparticularly challenging, and potential errors hinder translating DL modelsinto clinical workflows. Quantifying the reliability of DL model predictions inthe form of uncertainties could enable clinical review of the most uncertainregions, thereby building trust and paving the way toward clinical translation.Several uncertainty estimation methods have recently been introduced for DLmedical image segmentation tasks. Developing scores to evaluate and compare theperformance of uncertainty measures will assist the end-user in making moreinformed decisions. In this study, we explore and evaluate a score developedduring the BraTS 2019 and BraTS 2020 task on uncertainty quantification(QU-BraTS) and designed to assess and rank uncertainty estimates for braintumor multi-compartment segmentation. This score (1) rewards uncertaintyestimates that produce high confidence in correct assertions and those thatassign low confidence levels at incorrect assertions, and (2) penalizesuncertainty measures that lead to a higher percentage of under-confidentcorrect assertions. We further benchmark the segmentation uncertaintiesgenerated by 14 independent participating teams of QU-BraTS 2020, all of whichalso participated in the main BraTS segmentation task. Overall, our findingsconfirm the importance and complementary value that uncertainty estimatesprovide to segmentation algorithms, highlighting the need for uncertaintyquantification in medical image analyses. Finally, in favor of transparency andreproducibility, our evaluation code is made publicly available at:https://github.com/RagMeh11/QU-BraTS. | [
"In recent works {{cite:3dd16f52c82ecb581bae11dc50cbc0c0cfbfb0be}}, {{cite:dd32c730b7f538147a41815923dbeb1df70a5bb8}}, it has been shown that uncertainties associated with the outputs of a machine learning model are primarily divided into two sub-types: (i) Epistemic uncertainty, which captures the uncertainty associated with the model parameters, and (ii) Aleatoric uncertainty, which captures the uncertainty inherent in the data. The epistemic uncertainty captures our ignorance about which model generated our collected data. This uncertainty can be reduced to zero if the model is provided with an infinite amount of data, permitting the model parameters to learn the true distribution of the data generation model. The aleatoric uncertainty could result from measurement noise, for example, and therefore cannot be reduced even with the collection of more data. Both epistemic and aleatoric uncertainties play important roles in the field of medical image analysis. Epistemic uncertainty indicates where to trust the model output {{cite:2edc5983149087536f476427f8576ee70d9159a0}}, {{cite:4df48fd0011fdbd9c9e2c4469e96ba57317f670b}}, and aleatoric uncertainty reflects the prevalent noise in the data {{cite:c5781431bcf1b4e526c14a3ad7db158a1c2d8254}}.",
"Several recent papers {{cite:c551d2c9239e40c5aa413842e43c8a2461398f43}}, {{cite:2edc5983149087536f476427f8576ee70d9159a0}}, {{cite:6b3bc35d492a882fc30ded649e36b3d8383592c0}}, {{cite:8a769412584627d2f155014addb15c948179d9cd}} show cases where uncertainty estimates correlate with errors in a machine learning model. These results show promise that by estimating uncertainties, better adaptation of deep learning models in real-world scenarios is possible. However, in the medical image analysis field, to date, there is an unmet need to: (1) systemically quantify and compare how well different uncertainty estimates properly communicate the degree of confidence in the output, and (2) to rank the performance of competing estimates, given the objectives of the task and the requirements during clinical review.",
"The most popular metrics for measuring model confidence output are the expected calibration error (ECE) and the maximum calibration error (MCE) {{cite:dcd23fccec4724a1cf1ca5f20ac43de8164e43de}}, {{cite:a8d5c9359677d6d1973f0c3ee3d3a88289049dae}}. These metrics are useful for quantitatively measuring model calibration. However, these metrics are based on the softmax probabilities. Furthermore, a simple post-processing technique like temperature scaling {{cite:74a9e0bfc525b33d3cc61d16574ddde527eb57f9}} can make a deterministic and a probabilistic model equally calibrated. ECE and MCE metrics cannot differentiate between these temperature-calibrated models.",
"In another paper, {{cite:8501680d0dbd80dafa02cfb85f815051a7126c85}} evaluate the usefulness of the predictive uncertainty for decision making by evaluating the model output only in cases where the model’s confidence is above a user-specified threshold. Their main idea is that if the confidence estimates are well-calibrated on the data distribution seen during training, one can trust the model’s predictions when the reported confidence is high and result to a different solution when the model is not confident. They showed that, indeed, when the model is evaluated on its most confident prediction, the model accuracy is high compared to when the model is evaluated on all outputs. Though this is encouraging and allows for the comparison of different uncertainty generation methods, it does not consider how many model outputs were discarded at a certain threshold. Using this criterion, a model which has low accuracy but high uncertainty when evaluated on all predictions is rewarded. However, this model is undesirable in a practical scenario, and it leads to the rejection of most of its predictions to achieve high accuracy.",
"{{cite:a3bdfcfe369b499ff0d1818fc74cc0e6387dd80e}} designed a metric to quantify uncertainty for the task of semantic segmentation. They made the following assumption during the metric design: if a model is confident about its prediction, it should be accurate, which implies that if a model is inaccurate on output, it should be uncertain. With this in mind, they calculate the following two probabilities at different uncertainty thresholds: (i) p(accurate{{formula:470b99b3-4849-4eda-bcbb-1ced325cd5d9}} certain): the probability that the model is accurate on its output given that it is confident; (ii) p(uncertain{{formula:7575e3b6-8123-40db-8596-8b5c0b9b760b}} inaccurate): the probability that the model is uncertain about its output given that it has made a mistake in its prediction (i.e., is inaccurate). They used the metric to compare different BDL methods for the semantic segmentation task. Though this metric is indeed useful for semantic segmentation, where each pixel in an image is labelled as one class, it is not useful for the task of pathology segmentation where there is a high class-imbalance problem and the number of pixels (voxels) of interest (pathology) is low compared to the background-healthy class. For example, in the brain tumour segmentation task, 99.9% voxels belongs to background (healthy tissue) while only 0.1% belongs to pathology. Due to high class imbalance, p(accurate{{formula:3cdec192-d23d-4797-acc1-82883e103e59}} certain) would be dominated by healthy (background) voxels and majority of them can be accurately classified with high certainty.",
"{{cite:142b8e2cb23a37acd8889687f6bebab09c582c5a}} developed a metric, Probability-based Detection Quality (PDQ), to evaluate the uncertainty estimate for the task of object detection. The authors combine the class labelling measure (i.e., label quality) and the bounding box detection measure (i.e., spatial quality) into the metric. Here, spatial quality measures how well the detection describes where the object is within the image. Label quality measures how effectively a detection identifies the object class. These are averaged over all possible combinations of bounding boxes and labels generated using multiple samples. The authors also organized a challenge associated with this task at the Annual Conference on Computer Vision and Pattern Recognition (CVPR) 2019. The paper and its associated challenge {{cite:4b763581c0e0c0f86745628a2d722b9cae9e0d32}} illustrate the importance of developing uncertainty quantification metrics that are tailored to the task of interest.",
"{{cite:a0941b22f19ac82b23e74e13004f4dcb5a99ea48}} made the first step towards quantifying uncertainty for the brain tumor segmentation task. They compared various uncertainty generation methods such as MC-Dropout, Deep Ensemble {{cite:9c6a8be805ecdc477ebc84c6abf924d0f1e692af}}, {{cite:8501680d0dbd80dafa02cfb85f815051a7126c85}}, and others, using the standard metrics like ECE, MCE, and reliability diagrams. In addition, they proposed a new metric, Uncertainty-Error (U-E) overlap. The results showed that Deep Ensemble was able to produce more reliable uncertainty measures compared to other methods."
] | [] |
Online Grounding of Symbolic Planning Domains in Unknown Environments | If a robotic agent wants to exploit symbolic planning techniques to achievesome goal, it must be able to properly ground an abstract planning domain inthe environment in which it operates. However, if the environment is initiallyunknown by the agent, the agent needs to explore it and discover the salientaspects of the environment needed to reach its goals. Namely, the agent has todiscover: (i) the objects present in the environment, (ii) the properties ofthese objects and their relations, and finally (iii) how abstract actions canbe successfully executed. The paper proposes a framework that aims toaccomplish the aforementioned perspective for an agent that perceives theenvironment partially and subjectively, through real value sensors (e.g., GPS,and on-board camera) and can operate in the environment through low levelactuators (e.g., move forward of 20 cm). We evaluate the proposed architecturein photo-realistic simulated environments, where the sensors are RGB-D on-boardcamera, GPS and compass, and low level actions include movements,grasping/releasing objects, and manipulating objects. The agent is placed in anunknown environment and asked to find objects of a certain type, place anobject on top of another, close or open an object of a certain type. We compareour approach with the state of the art methods on object goal navigation basedon reinforcement learning, showing better performances. | [
"The problem of integrating symbolic action models with low level sensory data and actions has been addressed by different approaches. Most of them are based on RL techniques.{{cite:35afd829b8537070463f212496ba418b18140276}} ({{cite:35afd829b8537070463f212496ba418b18140276}}) propose a framework, called SDRL, which combinessymbolic planning on PDDL and Deep RL to learn policies that compilehigh level actions into low level operations. SDRL assumes that agrounded domain model is provided in input and neverupdated; Ogamus, instead, learns how to ground the domain model with newobjects discovered at run time. Moreover, SDRL assumes a perfectoracle that maps low level perceptions into symbolic states, whileOgamus deals with faulty mappings.NSRL {{cite:5d232fb49667f448963727ff1b518eeecfe89d74}} represents abstract domains in firstorder logic and uses RL to learn high level policies. NSRL generatesa compact representation of the learned policies as a set of rules via Inductive LogicProgramming. Similarlty to SDRL, NSRL assumes a given and fixedabstract domain instantiation and a perfect mapping from sensory data tosymbolic state.DPDL {{cite:3cd1beb726b330e14dfc674dd83c2e436f6f6ec8}} represents abstract domains inPDDL. It learns online both mappings from sensory data to symbolic statesand low level policies for high level actions. In Ogamus, instead, the mappingfrom perceptions to symbolic states is obtainedby combining a set of neural networks that are trained off-line.Moreover some of the highlevel actions are pre-compiled in low level operations (e.g., pick-upan object at a given position), while policies for moving actions are computed on-linevia path-planning. As the other methods mentioned above, DPDL assumes a given and fixedgrounded PDDL domain. Moreover, it focuses on performingmanipulation tasks in a single scene. Ogamus, instead, can work ondifferent scenes (we evaluate it on 35 different scenes). Finally, Ogamus uses egocentric and dynamic views, while DPDL works with an external fixed camera.",
"Differently from the above mentioned works, in the approach proposed by{{cite:805a5eac7c9b938963fd3f4341595a486cbf281c}} ({{cite:805a5eac7c9b938963fd3f4341595a486cbf281c}}), likein Ogamus, theagent instantiate the abstract domain online byaugmenting the set of constants every time the agent discovers new objects.The states of the instantiated abstract model isrepresented with a set of propositional atoms on the current set of constants.However, this approach is evaluatedonly with an extremely simple environment, while Ogamus is tested inaccurate simulated environments with complex objects and egocentric images. Furthermore the approach in{{cite:805a5eac7c9b938963fd3f4341595a486cbf281c}} does not take advantage of the power ofsymbolic planning techniques on PDDL domain descriptions, and itdoes not generalize over different tasks.",
"Some works in the literature deal with action schema learning from input traces, adifferent but related problem w.r.t. action schema instantiation.Some approaches to action schema learning do not consider the low-level high dimensional sensory data(see e.g., {{cite:56a4e5847a5549c09fcb18c586e6caeea2b8da59}}, {{cite:475883b9dccf6656e7a038f3f7b49fac8ee63c15}}, {{cite:4b8d68f356404db4659b274c3a470d0469e3da92}}), while others(e.g., {{cite:e6cabe436081455a7cec391a3ad3783a9af73577}}, {{cite:345ff9962e1fc8d43d1cdff2348ae55c5d92e8d1}}, {{cite:6dbd3e0a13fce66d904638a68d18a1c4f4d4e960}}) learn action models fromtraces composed of sequences of low-level data generated by actionexecutions. Themain difference with Ogamus is that they learnaction models, while Ogamus focuses on solving goals in a wide classof environments by instantiating a lifted action model given ininput. Moreover, in Ogamus, the dimensionality of the ground domain isnot fixed, but it is discovered by observing the environment. A furtherdifference is that the above approaches assume that the lowlevel actions coincide with the abstract actions and do not take intoaccount the problem of producing low-level policies that realize highlevel actions.",
"We experimentally evaluate Ogamus on the Object Goal Navigation taskthat has recently received much attention in the embodied AIcommunity{{cite:20b104f20f9de4a1b5db40e6150f140b750e1aec}}, {{cite:cb590cba10add6e03fabe936049d0f60d35211b7}}, {{cite:daf0c997fc6378b80717a9f511d046e2779ed324}}, {{cite:392e03fa6bfe84b676f847e1d2e040fe33971b43}}, {{cite:8f19c289764e639428f03257472dcf95f5f0854f}}, {{cite:5fcb150f643a145b2c35e7da938d2598af7e4b8a}}, {{cite:d1bc524649b5aefb1d7c9281b547b6018c6f1572}}, {{cite:ed8330d50e2a8a78086960709fac312e0832f675}}, {{cite:8d72f6a2d57c67c74331e271765aab31232468a2}}.We experimentally show that, for the RoboThor object goal navigation challenge,https://ai2thor.allenai.org/ Ogamus performs better thana method based on DD-PPO {{cite:314549a109560f21b772492a6f9f9b5b396da2b0}} which won the challenge using pure RLbased on low-level features, without exploiting a symbolic domain."
] | [] |
Continual Learning of a Mixed Sequence of Similar and Dissimilar Tasks | Existing research on continual learning of a sequence of tasks focused ondealing with catastrophic forgetting, where the tasks are assumed to bedissimilar and have little shared knowledge. Some work has also been done totransfer previously learned knowledge to the new task when the tasks aresimilar and have shared knowledge. To the best of our knowledge, no techniquehas been proposed to learn a sequence of mixed similar and dissimilar tasksthat can deal with forgetting and also transfer knowledge forward and backward.This paper proposes such a technique to learn both types of tasks in the samenetwork. For dissimilar tasks, the algorithm focuses on dealing withforgetting, and for similar tasks, the algorithm focuses on selectivelytransferring the knowledge learned from some similar previous tasks to improvethe new task learning. Additionally, the algorithm automatically detectswhether a new task is similar to any previous tasks. Empirical evaluation usingsequences of mixed tasks demonstrates the effectiveness of the proposed model. | [
"Work on continual learning (CL) started in 1990s (see a review in {{cite:498e70e52c313b46ff4858ba835a6a93610b9789}}). Most existing papers focus on dealing with catastrophic forgetting in neural networks. {{cite:498e70e52c313b46ff4858ba835a6a93610b9789}} proposed the technique LwF to deal with forgetting using knowledge distillation. {{cite:408feb818301a706e0f89bfcd7da390783f783c9}} proposed EWC to quantify the importance of network weights to previous tasks, and update weights that are not important for previous tasks. Similar methods are also used in {{cite:9f7be6be52c945a573131c03647f9a5b50eadb6a}}, {{cite:0fa3b836d861f357e5acf5db32013a4b6e2b828a}}, {{cite:7dd57fd8b58664d7cd7bb1a9284c60e5d688e762}}.Some methods memorize a small set of training examples in each task and use them in learning a new task to deal with forgetting (called replay) {{cite:0c539f83def01d4c4901d63d982866ed6ab30de0}}, {{cite:f27f94e76011e319502abd5f505ee5e07ff0cb46}}, {{cite:b6d9a695b01b0399c93c82f6cb579971405b3fff}}, {{cite:7b9e464ec47c4901ca1de179ef2c993a2729af44}}, {{cite:4d5335895541b6a21c0a054e2b9dd02a3e2ea804}}.Some works built generators for previous tasks so that learning is done using a mixed set of real data of the new task and generated data of previous tasks (called pseudo-replay) {{cite:d0a907bb0ab1f899f47e07b49f4be99b999bb4f1}}, {{cite:da0c68b05fd3a0249b911856cfa04ef41245ef80}}, {{cite:c056539f6aa2a2049e89da365c81a840b06600cb}}, {{cite:d0c4de52e7fbcb3ada6143f543ce5e014c377918}}.",
"Additionally, {{cite:b30441b0e60ba92e39ed3c70c8a1507009b9590f}} proposed to optimize loss on the new task with representations learned from old tasks. {{cite:d0c4de52e7fbcb3ada6143f543ce5e014c377918}} proposed PGMA, which deals with forgetting by adapting a shared model through parameter generation.{{cite:7946e1ffad30bba9a5a019e4b15f20da237e44fe}} tried to learn the new task by revising the weights in the orthogonal direction of the old task data. {{cite:e0daee94073192a9d30bfbd043e914915f0a699b}} combined three loss functions to encourage the model resulted from the new task to be similar to the previous model. Other related works include Phantom Sampling {{cite:ae0aa001a7029f1ad863db900e7876df302a4c56}}, Conceptor-Aided Backprop {{cite:ca8bb828ff9b95ab54a75a1b44ea401738718b44}}, Gating Networks {{cite:fe201314c9237c08fcc3572b9cd7222b7e20daf1}}, {{cite:e54d857d19fff78a016808a914a42bcbe860a473}}, PackNet {{cite:aff8db3314c0978144348fe527ecbb9edcd7bb68}}, Diffusion-based Neuromodulation {{cite:e0cbf482a5830518db008f32956c3a13e0561681}}, IMM {{cite:e5666980ee89f6841aa79a31b16c1a5e3f08cb60}}, Expandable Networks {{cite:efd1bf80b05ce0d905ed98d22f28fd400695deef}}, {{cite:28d0fe0cc8b0106102381583c2d19c832fe5a88c}}, RPSNet {{cite:1f1009bc758efe2ae6ca58cfce4cc3b065f3da55}}, reinforcement learning {{cite:be7106bece3bfb0aeca4b8e6701c7b3f412d2cc1}}, {{cite:4b69a03153778823f4997ce8b2e42a3582a9f096}}, and meta-learning {{cite:761fcac7260ff729857eaec529367e489a5d332a}}. See the surveys in {{cite:7f8f914ce4e99c30aacbfa1dd35c7cf319ad07a7}}, {{cite:f330e2c09d63ef41821611faefeba230276a46cf}}.",
"Most CL works focus on class continual learning (CCL). This paper focuses on task continual learning (TCL) {{cite:32092fd4aeed28a983a49b3f27d232af250d4d93}}, {{cite:e54d857d19fff78a016808a914a42bcbe860a473}}.For example, GEM {{cite:f27f94e76011e319502abd5f505ee5e07ff0cb46}} takes task id in addition to the training data of the specific task as input. A-GEM {{cite:b6d9a695b01b0399c93c82f6cb579971405b3fff}} improves GEM's efficiency. HAT {{cite:e54d857d19fff78a016808a914a42bcbe860a473}} takes the same inputs and use hard attention to learn binary masks to protect old models in the TCL setting.But unlike CAT, HAT does not have mechanisms for knowledge transfer.Note that the controller in iTAML {{cite:a977eb340c7e7028100e176152eb1ec1b900569f}} behaves similarly to HAT's annealing strategy in training. But the controller is for balancing between plasticity and stability, while HAT trains binary masks. UCL {{cite:16804cc0d4f0301d039aad1a9814218f4e0f217e}} is a latest work on TCL. However, none of these methods can deal with forgetting and perform knowledge transfer to improve the new task learning at the same time. Progressive Network {{cite:7c15c54f2ae70f2f7c2d8c59a130757607fb5fe4}} tries to perform forward knowledge transfer. It first builds one model for each task and then connects them. However, it cannot do backward transfer and its whole network size grows quadratically in the number of tasks, which makes it difficult to handle a large number of tasks. It also does not deal with a mixed sequence of tasks.",
"Earlier work on lifelong learning has focused on forward knowledge transfer to help learn the new task better {{cite:6c60f7d1d3bddc88a870bb4315baf7b887e93a39}}, {{cite:45deae8cbcc917bed4d6e622d4d51fd66375b880}}, {{cite:04f199cddadbef4f11f6fe785e1d7e8b2b8ce1cb}}, {{cite:9f83a6032541c93a1039418ea1b824eb4826c091}}, {{cite:1862ed2de4c7ebf0884c28b593187449e3f86641}}. {{cite:20319bbbe0a590033c82d1a0c43fcdbe607b2531}} and {{cite:bf36565b5761184d7747c12bc323846845de7bd7}} also did backward transfer. However, these lifelong learning works mainly use the traditional learning methods such as regression {{cite:45deae8cbcc917bed4d6e622d4d51fd66375b880}}, naive Bayes {{cite:9f83a6032541c93a1039418ea1b824eb4826c091}}, {{cite:bf36565b5761184d7747c12bc323846845de7bd7}}, and KNN {{cite:6c60f7d1d3bddc88a870bb4315baf7b887e93a39}} to build a model for each task independently and hence there is no forgetting problem. Although there are also works using neural networks {{cite:20319bbbe0a590033c82d1a0c43fcdbe607b2531}}, {{cite:863243a75bbca19825d4c9ad0dc3bccd73fee8c2}}, {{cite:6c60f7d1d3bddc88a870bb4315baf7b887e93a39}}, {{cite:04f199cddadbef4f11f6fe785e1d7e8b2b8ce1cb}}, all of them (including those based on traditional learning methods) work on tasks that are very similar and thus there is almost no forgetting. Earlier research on lifelong reinforcement learning worked on cross-domain lifelong reinforcement learning, where each domain has a sequence of similar tasks {{cite:484c2c777ce230311e326ff8e3eaa30a4d8657f6}}, {{cite:ef1f2389ba6c26c9913801cd88081a2e99396483}}. But like others, they don't deal with forgetting. To the best of our knowledge, no existing work has been done to learn a sequence of mixed similar and dissimilar tasks that deal with forgetting and improve learning at the same time."
] | [] |
Pre-Training Transformers for Domain Adaptation | The Visual Domain Adaptation Challenge 2021 called for unsupervised domainadaptation methods that could improve the performance of models by transferringthe knowledge obtained from source datasets to out-of-distribution targetdatasets. In this paper, we utilize BeiT [1] and demonstrate its capability ofcapturing key attributes from source datasets and apply it to target datasetsin a semi-supervised manner. Our method was able to outperform currentstate-of-the-art (SoTA) techniques and was able to achieve 1st place on theViSDA Domain Adaptation Challenge with ACC of 56.29% and AUROC of 69.79%. | [
"Following AlexNet{{cite:ca8e031352191c5108134c17ce628fd13c3f25a1}}, convolutional neural networks (CNNs) have become standard for image classification tasks. Various models based on CNNs{{cite:601d93e519e5b00993f84f359ff90c3db51a4edc}}{{cite:369621095c8b1690bfc24ad01684ade4a7b6a941}}{{cite:484a842bffa837fcb5f4431b8109bb0e93fc90ce}} have been introduced that achieve a significant increase in accuracy on various datasets{{cite:63c237371e0cb6070473c2e737d1bf022724229b}}. However, these models fail to perform in conditions where there is a big input distribution shift between training and testing dataset and/or has label set variance{{cite:0533d3cb1528eab7c321753fc6865ef155c43189}}. Transformers such as BiT{{cite:09b5db843e01771a0525ed434b4d58120b857f83}}, ViT{{cite:c15faf517aca00d62ca2b0ed0c142675f9615250}} have demonstrated significant improvement over CNNs in terms of accuracy at image classification tasks, however they require huge amounts of data to train. Meanwhile{{cite:0911d2fb53d8db8c72c7a195c44af07eae9a86ff}} utilizes the idea of using a single fixation of parse trees for image classification and attribute learning and{{cite:7a4b34fac94fb0700875ee789de75166a7c3a43b}} improves that by using dynamic routing and fixed vector representation for image classification, however both of the ideas don’t work significantly well on large datasets because of its nonlinearity which results an increase in training complexity.",
"Domain Adaptation refers to fitting a model that has been trained on a particular source dataset on an out-of-source novel target distribution, which is not part of the training set. Closed-Set Domain Adaptation{{cite:3f0942eedee2cf979484190ce0cae90b10b0284b}}{{cite:b58a8aad8db350b5dfdbc923d1969b67ee1644b4}} methods, where the source and the target domain completely share the class of their samples, work extensively well and have low input distribution shifts, however they fail to work in open-set environments because of unknown target input samples. Self-supervised learning methods could be used to solve these issues by either distillation{{cite:6b7cc21771e8dbd6eda17ac1cdd8881aecc780af}} or contrastive learning{{cite:fa0a0de12cc141247bafd9d61d9265c313148de2}}, however these methods have significant drawbacks{{cite:bc1173492fb122c0bb0602399692d7c5b4f5cf0e}}{{cite:2b1df3f438b39c39946d3ce040793bb08f735198}}."
] | [] |
DegreEmbed: incorporating entity embedding into logic rule learning for
knowledge graph reasoning | Knowledge graphs (KGs), as structured representations of real world facts,are intelligent databases incorporating human knowledge that can help machineimitate the way of human problem solving. However, due to the nature of rapiditeration as well as incompleteness of data, KGs are usually huge and there areinevitably missing facts in KGs. Link prediction for knowledge graphs is thetask aiming to complete missing facts by reasoning based on the existingknowledge. Two main streams of research are widely studied: one learnslow-dimensional embeddings for entities and relations that can capture latentpatterns, and the other gains good interpretability by mining logical rules.Unfortunately, previous studies rarely pay attention to heterogeneous KGs. Inthis paper, we propose DegreEmbed, a model that combines embedding-basedlearning and logic rule mining for inferring on KGs. Specifically, we study theproblem of predicting missing links in heterogeneous KGs that involve entitiesand relations of various types from the perspective of the degrees of nodes.Experimentally, we demonstrate that our DegreEmbed model outperforms thestate-of-the-art methods on real world datasets. Meanwhile, the rules mined byour model are of high quality and interpretability. | [
"Relational data mining. The problem of learning relational rules has been traditionallyaddressed in the field of inductive logic programming (ILP) {{cite:fd3a458281103858b0be884326090646347c6b76}}.These methods often view the task of completing a missing triple as a query {{formula:f35a45ec-0221-4212-ad26-4d0a94766b37}} wherethey learn a probability as a confidence score for each rule between query entities and answer entities.Among these studies, Path-Ranking Algorithm (PRA) {{cite:d9e8816ab8afca9bb05a988a9716f22b3eab9871}} investigated the frameworkof random walk inference, where they select a relational path under a set of constraints andperform maximum-likelihood classification. A RNN model was developed by {{cite:8d49d1a088e85399afcf9e4f00f8c12d501ed487}}to compose the semantics of relations for arbitrary-length reasoning. Chain-Reasoning proposed by{{cite:9d22be4d77b137a3eaad7a4c400d923429c85db1}}, enabling multi-hop reasoning through a neural attention mechanism,reveals logical rules across all relations and entities. Although ILP models are capable ofmining interpretable rules for human, these models typically take both positive and negativeexamples for training and suffer from a potentially large version space, which is a criticalshortage since most modern KGs are huge and contain only positive instances.",
"Neural logic programming. Extending the idea of TensorLog that tackles the problem ofrule-based logic reasoning through sparse matrix multiplication, Neural LP {{cite:0d3e2ba244a3496ede390506d54d7fb71b10fc91}}is the first end-to-end differentiable approach to simultaneously learn continuous parameters anddiscrete structure of rules. Some recent methods {{cite:5d31105ec01defc8810327012c72fc5016580af3}}, {{cite:33f8d1a87c5853bf65286ea61e2d5170fef46245}}, {{cite:faa1f6329a37e393fb91d108d4e8eb1bdf797f74}}have improved the framework done by Neural LP {{cite:0d3e2ba244a3496ede390506d54d7fb71b10fc91}} in different manners.DRUM {{cite:5d31105ec01defc8810327012c72fc5016580af3}} introduces tensor approximation for optimization and reformulateNeural LP to support rules of varied lengths. Neural-Num-LP {{cite:33f8d1a87c5853bf65286ea61e2d5170fef46245}}extends Neural LP to learn numerical relations like age and weight with dynamicprogramming and cumulative sum operation. NLIL {{cite:faa1f6329a37e393fb91d108d4e8eb1bdf797f74}} proposes a multi-hop reasoningframework for general ILP problem through a divide-and-conquer strategy as well asdecomposing the search space into three subspaces. However, the existing methods ignore the effectscaused by the entities while reasoning over a specific relational path, thus witness a moreobvious failure where heterogeneous entities and relations are involved in the KGs.",
"Representation learning. Capturing their semantic information by learning low-dimensionalembeddings of entities and relations, also known as knowledge graph embedding,is a vital research issue in KGR, and we term those models as embedding-based models.Newly proposed methods, including RotatE {{cite:042784fc87317431d315b05e3879623c0621d16a}}, ConvE {{cite:01b1332a94aef6580dd4f1e1766bf21d1d9ab9df}}and TuckER {{cite:154c0668ddc72a751776bdb3dd74d6f7c50bb974}}, predict missing links by learning embedding vectors fromvarious perspectives of the problem. Specifically, the work of RotatE {{cite:042784fc87317431d315b05e3879623c0621d16a}}focuses on inferring patterns such as symmetry and inversion, where they proposed a rotationalmodel that rotates the relation from the subject to the object in the complex space as{{formula:cb062de0-e8c8-41a6-ab0d-3cb59f215e57}} where the {{formula:5a4330d4-5e8f-4705-84a5-d9863705b8de}} denotes the element-wise Hadmard product.ConvE introduces a highly parameter efficient model, which uses 2D convolution over embeddings andmultiple layers of nonlinear features to express semantic information.TuckER, inspired by the Tucker decomposition {{cite:63bf6d6c897b8fd6e1c93a591e3f2615c09dada2}} that factorizes a tensorinto a core tensor along with a set of matrices, is a linear model for link prediction that hasgood expressive power. Unfortunately, the biggest problem is that these sort of methods canhardly be comprehended by human, but we relate to these methods for their ability to capturelatent information of entities and relations through embedding."
] | [] |
MoCaNet: Motion Retargeting in-the-wild via Canonicalization Networks | We present a novel framework that brings the 3D motion retargeting task fromcontrolled environments to in-the-wild scenarios. In particular, our method iscapable of retargeting body motion from a character in a 2D monocular video toa 3D character without using any motion capture system or 3D reconstructionprocedure. It is designed to leverage massive online videos for unsupervisedtraining, needless of 3D annotations or motion-body pairing information. Theproposed method is built upon two novel canonicalization operations, structurecanonicalization and view canonicalization. Trained with the canonicalizationoperations and the derived regularizations, our method learns to factorize askeleton sequence into three independent semantic subspaces, i.e., motion,structure, and view angle. The disentangled representation enables motionretargeting from 2D to 3D with high precision. Our method achieves superiorperformance on motion transfer benchmarks with large body variations andchallenging actions. Notably, the canonicalized skeleton sequence could serveas a disentangled and interpretable representation of human motion thatbenefits action analysis and motion retrieval. | [
"Motion Retargeting.Motion retargeting has been extensively studied in the area of computer vision and graphics. Classical motion retargeting methods {{cite:4266905d686aa8a1559257fc7f7c49fcb5df9c39}}, {{cite:8cd50aa14d99cf774d1f54db8e7d96e5601eb731}}, {{cite:815a2707a8f8d7f3a1be2867b0d4156ba90fd3fa}}, {{cite:918052a4ca959eeaeca173c174bb3392d2e808c4}} mainly rely on simplified assumptions and hand-crafted kinematic constraints. In recent years, deep-learning-based approaches show promising results on obtaining human pose and action, which not only increases the availability of motion data, but also inspires motion retargeting research with deep neural networks. {{cite:8ebb3843f673cf2d1c7aed8eb637eacd49c093ef}} proposed to capture high-level motion with Forward Kinematics (FK) layers in recurrent neural networks. {{cite:d140dbcee36acb6e26cc8d46cb659cb9fc9663d4}} further optimized the retargeting precision by disentangling pose and movement. Skeleton-Aware Networks {{cite:a177706d2bb4c3b487ff6d01129da08656b9630f}} could automatically adapt to different skeleton topologies. Nevertheless, the above methods all require high-precision 3D motion (e.g., quaternions), which takes expensive motion capture systems or complex optimization in practice. Otherwise, estimating 3D motion from monocular RGB video could be error-prone {{cite:96c111ee4f9e335bf06e9dd554a09b759602ba97}}, {{cite:af3e042748f906378c7adf0bec459db364716877}}, {{cite:eab70218e25ff1ff456091e20fd292ff7504ecb2}}.",
"Several works have also explored to retarget motion from 2D inputs that are more accessible to everyday use cases. {{cite:e758aecee96073c4ade1edae416317fefbe0293b}} proposed a two-branch framework with part confidence map as 2D pose representations. {{cite:b5bf64758ba3f4734a646cf0b84403d756ad3fd9}} designed a global pose normalization to handle different body structures. {{cite:0b21e2c6b25d32978706e14c64f7fd68c1147d25}} proposed a supervised learning approach to retarget motion in 2D. It requires the exact same motion performed by different characters at different view angles as training data. TransMoMo {{cite:3f8cae715cd3e1d22c9b87285f191b0ed9165802}} explored unsupervised motion retargeting via invariance properties. These methods all generate 2D label maps for further rendering via image-to-image translation methods. Therefore, they can not be readily applied to driving 3D characters. They only consider accuracy for 2D joint positions and assume static view angle. Meanwhile, our canonicalization-based approach enables applying regularization on 3D poses directly and training with time-varying explicit 3D view.",
"In our framework, 2D poses are used as input and our method produces 3D skeletons as output. Therefore, our approach can exploit user-friendly monocular RGB videos with the help of robust 2D pose estimation algorithms {{cite:facab2e852a8881b9d09980d32b91a9bcc4b3923}}, {{cite:e670824b4c1f5358467c592132326e42c97b15a7}}, {{cite:3e458e7544030e9f4782ec982a1f33a5083a7d90}}. In addition, the 3D skeleton output allows combination with state-of-the-art neural rendering {{cite:6c0e086993d79beb3cdf124582a22a1e8a59fc44}}, {{cite:efb0e39adb2b8e3f9a0efe9ef30e9d72c1fdd0bc}} and avatar creation {{cite:9e0dd80fbdfa0835d547f3ace5c8f4986c56b025}}, {{cite:84a42db087bcdf2338e8b7440edfbc76697c405a}}, {{cite:f12caf45c1a0fbc9ec59a367b2a525169bc6f8d3}}, {{cite:e1ea39d8d5c0524eb60fefdeff8c3235dd24413a}} techniques. Our flexible approach thus provides a viable alternative to the traditional 3D reconstruction+IK pipeline.",
"Representation Disentanglement.Disentangling latent variables from observations has been a fundamental topic in machine learning. Previous approaches {{cite:542791d085d560df29a2dd1191ba2def39b8291b}}, {{cite:649234123f1168b74cc8e5f9c17e317605bc5bb7}} use labeled data to separate class-dependent and class-independent features. A vast literature focuses on unsupervised learning of disentangled representations with generative models {{cite:dcd192e8a398a9ed73cfb379c886ff48fb8bd7eb}}, {{cite:9c52229f02494489b507db156ee71e5f8be7d7fd}}, {{cite:bbc1bab2eefcbd03f46a479392a7d67c4735da38}}, {{cite:3dba9fc082155743edf5b9dfa5bc55b32dd4091d}}, {{cite:86ec3287786acb6fabe241db8a0bb7381878c36c}}, {{cite:8e4f138df9617b653649f435c182a5adc5356866}}, {{cite:94b513300139cc3c5959b257d8a5e4074de761a1}}. A recent line of work explored view-angle-agnostic motion representations aimed at improving motion recognition performance {{cite:259feb3123c2de1a83c2f964bcb3625c6c1a2e64}}, {{cite:2742c706c72095ea982e4382232ae2161c738ede}}, {{cite:e48a2299fe354d0780d26f9449a4b204ad32b922}}. In this work, we achieve disentanglement of motion, body-structure and view-angle via unsupervised canonicalization.",
"3D Canonicalization.Several different methods have been proposed for automatically computing the canonical form from a 3D pose or mesh. {{cite:adb15fcf44ccd33e499993036f46619d9a80ba0c}}, {{cite:85bbb49a7ea13786c54bd18e80f00b43697b8d1b}}, {{cite:082dc25ec06bc488e06f570bc133eb4f3023520c}}, {{cite:32e7b0a29c3963596023350f7987220b45f09b12}} propose to learn pose-neutral shape of any non-rigid meshes for 3D shape retrieval. {{cite:1e1f8aaeb77db6a6ae33290b3433ac8746c8dc64}} further improves the method with multidimensional scaling. {{cite:d462d95e75c77439325d5c6dc31b60b751620282}} present a feature-preserved canonical form that directly deforms the 3D models. Body reconstruction methods {{cite:9e0dd80fbdfa0835d547f3ace5c8f4986c56b025}}, {{cite:f91853684efc365e3e8b1d1b4a27a0c1fa70e457}} sometimes involve the canonical pose space for handling pose-dependent deformations. In those papers, canonicalization refers to pose canonicalization, that is, “transforming to a standard pose”. In this work, we explore the canonicalization of different factors, namely structure and view angle, while keeping the motion unchanged. C3DPO {{cite:2526e4e939f42d673c3cf8abfebd6a1cceb4f1d8}} designed a rotation-invariant canonical shape for structure reconstruction. CanonPose {{cite:45422aa27474a7c60d0c3b36fe3241566f4093b3}} learns a canonical camera view and respective camera rotations, but it relies on multiple synchronized cameras for training. We design canonicalization operations for both structure and view angle, which are applied to the entire action sequence rather than a single pose. Based on the designed operations, our model can be trained without data pairing."
] | [] |
Elastic-Link for Binarized Neural Network | Recent work has shown that Binarized Neural Networks (BNNs) are able togreatly reduce computational costs and memory footprints, facilitating modeldeployment on resource-constrained devices. However, in comparison to theirfull-precision counterparts, BNNs suffer from severe accuracy degradation.Research aiming to reduce this accuracy gap has thus far largely focused onspecific network architectures with few or no 1x1 convolutional layers, forwhich standard binarization methods do not work well. Because 1x1 convolutionsare common in the design of modern architectures (e.g. GoogleNet, ResNet,DenseNet), it is crucial to develop a method to binarize them effectively forBNNs to be more widely adopted. In this work, we propose an "Elastic-Link" (EL)module to enrich information flow within a BNN by adaptively adding real-valuedinput features to the subsequent convolutional output features. The proposed ELmodule is easily implemented and can be used in conjunction with other methodsfor BNNs. We demonstrate that adding EL to BNNs produces a significantimprovement on the challenging large-scale ImageNet dataset. For example, weraise the top-1 accuracy of binarized ResNet26 from 57.9% to 64.0%. EL alsoaids convergence in the training of binarized MobileNet, for which a top-1accuracy of 56.4% is achieved. Finally, with the integration of ReActNet, ityields a new state-of-the-art result of 71.9% top-1 accuracy. | [
"Quantized Weights with Real-Valued Activations. Restricting weights to be either +1 or -1 allows the time-consuming multiply-accumulate (MAC) operations to be replaced with simple addition operations. Recent studies {{cite:adb1b73271e01193cf6a6d247043e8dc111f84bb}}, {{cite:19111403de3b71aa5431c5864342b7b55ccca017}} rely on this insight and employ the straight-through estimator (STE) {{cite:45337193b73586a0bbfe0eadac83fc8196b552f2}} to tackle non-differentiability in the back-propagation of gradients during training. Courbariaux et al. proposed BinaryConnect {{cite:adb1b73271e01193cf6a6d247043e8dc111f84bb}} which drastically decreases computational complexity and storage requirements while achieving good results on the CIFAR-10 and SVHN datasets, but lacks experiments on large-scale datasets like ImageNet {{cite:f1515eb416fc77534a25203ec3b044fe198a5687}}. In Binarized Weight Networks (BWN) {{cite:19111403de3b71aa5431c5864342b7b55ccca017}}, the full-precision weights are binarized during each forward and backward pass on the fly, updating only the full-precision weights. BWN achieves a notable accuracy increase, especially on large-scale classification tasks. Finally, TWN {{cite:23a0bb6ebf77d70bed1d1004a0ce44fdd1ce3c6e}} and TTQ {{cite:1b8e0d401e101847d3babad36811eb89f34616e3}} use ternary instead of binary weights {{{formula:cbf98899-7510-42fa-a1d3-5496e649ba53}} , 0, {{formula:53233ead-514e-4e46-9446-0107ad7de0dc}} } to pass more information.",
"Quantized Weights and Activations. Another approach which has recently gained popularity restricts both the weights and the activations to {-1, +1}. This allows the convolutional operations to be completely replaced by the efficient xnor and bitcount operations {{cite:3c8534704281b33e34aed489a1a6ed4f7362f7d4}}, {{cite:19111403de3b71aa5431c5864342b7b55ccca017}}, thus gaining extreme efficiency. XNOR-Net {{cite:19111403de3b71aa5431c5864342b7b55ccca017}} is one of the most representative works of this approach and achieved remarkable accuracy with various networks. Bi-Real net {{cite:2d6c6c4387bd9be5aa4e040779cecae9cd39a093}} introduced additional shortcuts to retain the information of the real-valued activations, so as to alleviate information loss from binarization. The paper additionally introduced a custom gradient approximation method with a sophisticated fine-tuning strategy to further increase accuracy. Recently, Bethge et al. {{cite:31880289f312bc6c3c1f45449960d5ae47697dd3}} demonstrated that the said fine-tuning strategy and gradient approximation methods are not necessary for training BNNs - even without these, the authors achieved superior accuracy training from scratch with the simple straight-through estimator. Another method proposed to counteract the information degradation phenomenon is the linear combination of multiple binary weights to approximate full-precision weights, as introduced by ABC-net {{cite:e3ead6ea7868667bb9c8f3d36bcba08913c90f21}}. Finally, TBN {{cite:ed626b460b35ed50bf8079ca309db04ed6f1203b}} takes ternary inputs{-1, 0, +1} and binary weights with scale factors {{{formula:dc9b1fa4-a5fc-4b7d-95ba-26ab40cbea5a}} , {{formula:9d6b3982-3719-4c90-a4f8-6a723426f2c9}} }, demonstrating the method on both image classification and object detection tasks."
] | [] |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.