Datasets:

naivete5656

/

strogen-arxiv

like 0

[ "Some of the earliest methods for the RC task were based on pattern extraction. They {{cite:f3df0c2eb3ae4f302f454f0f318b899460d1ff7f}}, {{cite:4a83277e4833228b565c65bd7251de0c4c708f29}} use syntactic patterns to construct rule for extracting relational facts from texts. To reduce the human effort in identifying relation facts, statistical relational relation extraction has been extensively explored in two directions, namely, feature-based methods {{cite:d4bd8efdcc6b4804308333ac07b72f36e1edf7d1}}, {{cite:ea3b0c4f4ef7619d38358b1df5f705879d41d76c}}, {{cite:4d043183c84f87ac36b57451904bf0c90c4f5169}} and kernel-based methods {{cite:c787b00a85e5255738dc2324f64bb540fe5bd9cc}}, {{cite:4f00ebbec21a0cd811678f93fc9ec8dcca6eff44}}, {{cite:cef92e73e0ed428ff4921a86b8a1c3ff85bc5f06}}, {{cite:1276fa769e3e934062d4b2cae412ee69361dacc3}}. Following the success of embedding models in other NLP challenges, {{cite:97c3f7fa866fb2303e96cc2c5b004d0d934b66e5}}, {{cite:a63cb6fa21489ec4fc40bf6a18138b9963edc18b}} have made efforts in utilising low-dimensional textual embedding for relational learning.", "Neural Relation Extraction models introduce neural networks for capturing relational facts within text, such as recursive neural networks {{cite:3d3f1d2d171eddf08e1f55eb65dcdb7e2cd0928d}}, {{cite:281048dbd265f51b3a820833bbc216e6e67763c1}}, convolutional neural networks (CNN) {{cite:5a0dff45ddf20e3ce3c17498a549be5a4c2415ce}}, {{cite:dfdf8df8c7aab1661ce8fe8d43cdc999dce8be3c}}, {{cite:c10682f4ceb85ad397c9fd5ed208f9baad6c1f03}}, recurrent neural networks (RNN) {{cite:5fd1377fa5c5e653d432b1f250ac1172efe25019}}, {{cite:ec1e8851643a777e2307e5398c1ece3f3114f239}}, and attention-based neural networks {{cite:30850740520cd7e76d7e3ac63172c313ee700f44}}, {{cite:e2ec3ce76ac36bdc0904eae487da7163dbeb3365}}. Recently, Transformer {{cite:ee1f0f32f6fcc96ed8917459ce663330cb5c4b97}} and pre-trained language models {{cite:f31ff32306b868adcb7815985646493c90e417e5}} have also been explored for relation extraction {{cite:6a9d80dc1d239da87cc4b6d27e9657bc286b712b}}, {{cite:7500478cb7ec9897313db92b3e73a9a6467a71dd}}, {{cite:071fc82aa90c6c6dc9c903a7ae8748abe782b2b2}}, {{cite:339b9d67036d114290e9a367f6130827d48c6a8a}}, {{cite:3cacf27645219190bcdedcda79144bc99f766c56}}, {{cite:b55da367ee3dcbe5436bc16679b19dbc05cf4049}} and have achieved new state-of-the-arts performance.", "Out of all the RC datasets {{cite:724fface81267646d51381c2636bbd40d3f0d790}}, {{cite:5840560daafbb3abf22b5bb8b4085198bea626b6}}, {{cite:726c20003635262b00cc1f0fc3f8b212e424cb80}}, {{cite:5fd1377fa5c5e653d432b1f250ac1172efe25019}}, {{cite:bf734f8a95b9f708ae8ddf9c88cc5d33d1d14af9}}, TACRED {{cite:5fd1377fa5c5e653d432b1f250ac1172efe25019}} is the largest and most widely used dataset. It contains more than 100 thousand sentences for 42 relation classes (Train: 68124 sentences, Dev: 22631 sentences, and Test: 15509 sentences). Sentences are collected from TAC knowledge base population 2009-2014 tasks {{cite:5fd1377fa5c5e653d432b1f250ac1172efe25019}}. Each sentence contains a pair of real-world entities representing subject and object entities and is annotated with either one of the 41 relations or no_relation. Sentences are manually annotated with relation labels using the Amazon Mechanical Turk crowdsourcing platform, where each annotator was provided with sentence, subject and object entities, and a set of relations to choose from.", "A recent study by Alt et. al. (referred to as TACRev) trained 49 different RC models and selected 5000 most miss-classified instances from TEST and DEV partitions for reannotation {{cite:3e615bb6ee83d43c4de93076ad8fc7e5328fa476}}. Their expert annotators were trained linguists. Out of 5000 instances, they ended up modifying 960 sentences from TEST and 1610 sentences from DEV. The revised dataset resulted in an average 8% improvement in f1-score, suggesting the noisy nature of TACRED dataset. However, their work has two main bottlenecks. First, their method has a fixed set of sentences to reannotate. Second, they ignore the label hierarchy and use only model confidence to select sentences for reannotation. We compare our strategies against TACRev by extending their method for the reannotation budget. Given a specific reannotation budget, we select top sentences from the dataset where RC models have the highest confidence.", "Another recent study by Stoica et. al. (referred to as ReTACRED) reannotated the complete TACRED dataset using crowd-sourcing {{cite:279db7cfc2344fa8408cf1b078b0aec52143d480}}. For more effective crowd-sourcing, they have made slight modifications to the original TACRED relation list. They corrected 3936 and 5326 annotation errors in the TEST and DEV partitions respectively. They have also eliminated 2091, 3047, and 9659 sentences from TEST, DEV, and TRAIN set respectively for various reasons. We use the data from their reannotation experiment to simulate our reannotation strategies. Our work cannot be directly compared against ReTACRED as they simply use the brute-force method of complete reannotation. However, for a meaningful comparison with ReTACRED, we extend their approach by selecting a random set of sentences for a given reannotation budget.", "Recently Parekh et. al. {{cite:c045f46b926f65c0668ee2e86c46d1db70a97ed9}}, have collected more than 600 relations between Person, Organization, and Location entity types from multiple knowledge bases and arranged them in a taxonomical hierarchy following is-a relationship. They manually created the relation hierarchy but did not show any applications of the generated hierarchy. In this work, we create a similar relation hierarchy for TACRED relation labels, and further use that hierarchy for calculating distance between ground-truth and prediction.", "The main takeaway points from the existing work can be summarized as follows:", "Deep Learning is the paradigm of SOTA RC models. Hence, we will focus only on Deep Learning based models for our work.TACRED is the RC dataset used by SOTA models. Also, reannotation data for the complete TACRED dataset is availble. Hence, for our experiments we will used the TACRED dataset.Both the existing works on RC dataset reannotation are rigid. They fail to provide flexibility about how much data to reannotate. Hence, we will introduce the concept of reannotation budget to overcome this limitation.The concept of taxonomic hierarchy of relation labels is not yet used for the RC dataset reannotation. We will build such a hierarchy of relation labels in the TACRED dataset and use it for selecting datapoints for reannotation." ]

[]

[ "The literature in risk evaluation for sequential decision making can be divided between those that apply a risk measure to a single cost random variable, and those that apply risk measures recursively to a sequence of cost random variables.The former approach optimizes the risk of a single random variable,which does not account for the temporal structure in what generatesit, while the latter optimizes the risk of sequences of random variablesin a dynamic framework as additional information becomes available.", "Several authors address sequential decision making problems by minimizing the risk of a cost over a whole episode.For example, {{cite:ce931119ad0347f54b375ff914a8c89e39d6f0b1}} focus on objective functions for variance related criteria, while {{cite:2aaca81584103f9b466daee27fdb8c5b190f6189}} take a risk-constrained approach with an objective function that includes a penalty on the conditional value-at-risk (CVaR).Some lines of research look at risk-sensitive RL problems using a broader classes of risk measures, such as comonotonic {{cite:6955bb1409af2dd4fe6013f313cbaf6e47bc6619}}, entropic {{cite:4598a70598d572a485aa8463f09dd9654eeb8527}}, or spectral {{cite:3104c4beb4a2833cc89cb2759e47cd5e0d28644e}} risk measures.{{cite:c15f737b9c70de99a4e9cfe14a95f31f825f4f01}} consider a risk-neutral objective function, but includes an additional constraint on the risk of the cumulative discounted cost.", "Other works extend optimization in Markov decision processes (MDPs) by evaluating the risk at each period.For instance, {{cite:16ecf2b895b6e517173c2986384742343c12099c}} evaluates the risk at each period using dynamic Markov coherentrisk measures, while {{cite:58f5a3705e35ae4b63058f1b7fa6b2e8e3409dda}} and {{cite:002eef471de88dc06bffed5f59e54626f10ac914}} propose iterated coherent risk measures, where they both derive risk-aware dynamic programming (DP) equations and provide policy iteration algorithms.While they focus on coherent risk measures, various classes of risk measures have already been extended to a dynamic framework, such as distribution-invariant risk measures {{cite:c4789b405fd3256888679b3acba46ca18707efd2}}, coherent risk measures{{cite:386e9659c31c4b67d2c0e657ea76d3bb3074fcea}}, convex risk measures {{cite:68eaa68845b6ac0f048f2d09fa1eaa3274104eda}}, {{cite:e534badda0a2d9f9b060fc3b4c79fb1d9b1933c6}}, and dynamic assessment indices {{cite:b6d961763403e8a6d942cc913025dab8e17ce47e}}, among others – however, these works do not look at how to perform model free optimization, i.e., they do not look at RL. For an overview of dynamic risk measures, see, e.g., {{cite:e8609d117c0730a5dc73f2c5d7185dcf70c21b84}}.", "Another way to account for uncertainty in the data-generating processis by allowing the parameters of the model, or the entire distribution itself, to be unknown.The class of robust MDPs {{cite:45dec5d605e68f340ce1170dae433edfe30baa3b}} focuses on optimizing the worst-case expectation when the parameters vary within a certain set.There exists relationships between risk-aware and robust MDPs, as shown in {{cite:d6850647d3c708b7a3413b367ebc4bf7c038e91d}} and {{cite:7700060ca1cc322c1eb451c7ec3307def4914178}}.Indeed, minimizing a Markov coherent risk measure in a risk-aware context is equivalent to minimizing a certain worst-case expectation where the uncertainty set is characterized by a concave function.Several researchers have developed algorithms to solve robust MDPs, for an overview see, e.g., {{cite:58d0bbaeed923009fd960d6f2cf8bc6282153e7e}}.", "While we consider a model-free approach, several researchers tackle related, but distinct, problems in a model-based framework.For instance, {{cite:6deaad688b9e3c367edf65ff6f08178270056b9f}} and {{cite:e69512550acf808e4caef166d5ee1390b4367c47}} use deep learning methods to solve non-linear partial differential equations (PDEs). Using the non-linear Feynman-Kac representation, they reformulate the PDEs as backward stochastic differential equations (BSDEs), they then parametrize the co-adjoint process and initial condition using an ensemble of neural networks, and use the mean squared error in the terminal condition as the loss function. Despite there being an equivalence between BSDEs and dynamic risk measures {{cite:58845bbc5fb40c8ee9d1acaa0c335c070b7176d8}}, {{cite:bd0cac17f20906afe654dcb8f64f526b0d7cfc3a}},the dual representation does not directly help when we aim to optimize a dynamic risk measure in a model-free manner.", "Other types of algorithms exist in the literature to find a solution to risk-aware RL problems.Among others, {{cite:e4941968b46df268ae4cce4583db37a31a9fe6eb}} use a multi-armed bandit approach, {{cite:5459a35cfe7124b6bbcbcd47dcbd2b8822518a36}} devise a risk-sensitive Q-learning method when optimizing utility functions, {{cite:1e4b19a44efadbdffbe04861c5aa02ef82e1b4c8}} use a distributional perspective to learn the whole distribution of the value function, {{cite:2812fa9d76d98933371030e1007803d747a2cbc8}} employ an approximate DP approach to devise a value iteration algorithm for Markov risk measures, and {{cite:84f12d80d3d96f181293f378d256727a70d3b649}} address risk-aware problems from a Bayesian perspective.The shortcomings of these approaches are that they apply to a specific class of risk measures and the developed methodologies are tuned to them." ]

[]

[ "Perceiving emotions in conversation provides affective information of the conversation partners; similarly, perceiving politeness in conversation provides cues about their social manners.The emotional states and dialogue acts have unique relations which is presented by {{cite:83ecf10c3310706a489261417e94d14ddd10166a}} in Emotional Dialogue Acts {{cite:18414da00dcd5941e583572efb2c10a104fcc751}}, {{cite:b41d0a7cfe6b569440776bb916662c1cc5f29291}}, {{cite:83ecf10c3310706a489261417e94d14ddd10166a}}.For example, Accept/Agree and Thanking dialogue acts often occur with the Joy emotion, Apology with Sadness, Reject with Anger, and Acknowledgements with Surprise.Similarly, politeness and emotion together are evidently discussed a lot in the literature as some of the most critical social cues {{cite:e21a506021a113a5f7d51200dd11b59f7f35ff75}}, {{cite:683bda62cc71b4e706429ce03c60b7098180c0ca}}, {{cite:6402e416bb21bd190d3e84a38b5b18aa93d20149}}.Some computational linguistic study shows how machines learn politeness, for example, please and could you words signals on the heatmaps of sentences {{cite:a90b9064af31341dcf431b276f6ca1945cb5af9b}}.", "In fact, {{cite:baf8bc446d312b581faeac7982060068b2778788}} mentions that the display of emotions or lack of control of emotions as positive politeness strategies or potentially face-threatening acts.Thus this study fosters analysis of linguistic politeness by understanding how ‘appropriate levels of affect’ are conveyed in the conversational interaction {{cite:e21a506021a113a5f7d51200dd11b59f7f35ff75}}.In human-robot interaction, politeness as a social cue plays a vital role to drive engaging social interaction with robots {{cite:216e9839ddb3edb2c6cb931a783da53210d0ec24}}, {{cite:bd03191c1268e2f5b0eba702a68799724cfc4eb5}}, {{cite:7416387213266e046e42e48766eb9b0029244d3d}}.However, this article explores relations between the socio-linguistic cues, finding out how polite the utterances are against their respective emotion and dialogue act classes." ]

[]

[ "RC, mainly composed of Echo State Networks (ESNs) {{cite:a2c0b235a9a2322f1e3b7ad3347c9fb8dc054de1}}, {{cite:c03875d0e037084ccb18e3fe44952b5acbcd6ad0}} and Liquid State Machines (LSMs) {{cite:dc6e145f3859edf543fb829f2192ff1a9b77e9ec}}, has been successively applied on speech recognition {{cite:15cf05383622297af1cd3a76321ebf593b3e1ea3}}, time series prediction {{cite:8769bf6430d4fc35e8723d0b867b01e7b40423ad}}, MIMO-OFDM {{cite:5fc6e381fbf1de9c9deff7779634e5a73723e7b8}}, smart grids {{cite:785fdaffc1155932b0edb76fd9dedef0c2e7b621}}, etc. Physical RC {{cite:98e0251a1bffbaa4c929fedeabcdd339390c429f}} such as photonic RC {{cite:d4a814072676da5f04df2eba96a4b003eae6f70c}} is amenable to hardware implementation using a variety of physical systems, substrates, and devices. However, few work have so far studied the behavior of such networks on computer vision.", "One of the related experiments has been carried out to classify digits of the MNIST dataset using a normal ESN {{cite:0fcbc1759ec1d5a5d79e427be2ed242d2edb37a9}}, which shows that RC can handle image data. More applications corresponding to images with RC are detailed in {{cite:059c6f5a7c6b7eab51555460d8a9f12acee4ae2f}}, {{cite:0fee289d83daa1048fdc78c62ed718bf9f61f81d}}, {{cite:f19e8a80fcdfcf1638c742b5245e0454b4d4f741}}. Combining RC with other popular algorithms such as CNNs or Reinforcement Learning (RL) is a valid approach to enhance the performance of networks for image tasks.Tong and Tanaka {{cite:9da64ab4cdbd85bb6a0020c836df6da2fc57e961}}took full advantage of an untrained CNN to transform raw image data into a set of small feature maps as a preprocessing step of RC and achieved a high classification accuracy with a much smaller number of trainable parameters compared with Schaetti's work. Then, a novel practical approach, called RL with a convolutional reservoir computing (RCRC) model {{cite:7208f9e2539cce1e12d05a4f910ad83d083c2a6c}} was proposed. A fixed random-weight CNN, used for extracting image feature maps, combined with an RC model, employed for extracting the time series features, is adopted for the evolution strategy of RL, which succeeded in decreasing computational cost to a large degree. However, such studies remain restricted because they simply treat RC as an auxiliary tool but not the core for image tasks.", "Based on the above issues, we propose an RC-based model specifically for image tasks, by exploring the internal topology of the reservoir. The proposed topology is originated from the Simple Cycle Reservoir (SCR), shown in {{cite:354c0f4c63410408270347ad7cbb9bf0ee55bc98}}, which performs close to the traditional ESN. Later on, the updated one-Cycle Reservoirs with Regular Jumps (CRJ) {{cite:ee455baa8bb659e2f969ce0557fffadeada87d15}} has been proposed based on SCR, which obtains superior performances in most time-series tasks.", "Recently, Verzelli et al. {{cite:ba1b919a430c5f9bdffb6d5dc0ba24018330ad9f}} made the best of the controllability matrix to explain the encoding mechanism and memory capacity of the reservoir with the topologies mentioned above, from the perspective of the mathematical characteristics of the matrix, i.e., the rank and the nullspace of the matrix. However, higher memory capacity usually does not correspond to higher precision prediction {{cite:5620bf1215a074f2a69e603d1ea52790e5061557}}, and the two are generally traded off in the \"edge of chaos\" state {{cite:7767b6afb883cc81722025aaba80603aaad36025}}.", "Another recent related model is the deep RC. Taking ESN as an example, the typical architecture of deep ESN can be classified into the series ESN and the parallel ESN, deriving other deep structures {{cite:f543c8d44b8b2f5aff446a0504979d34905dd64f}}, {{cite:e21ca8015fa877daebbda404d779606f032ffe69}}, {{cite:bb291b018c338f3682cb74c45509baa6e5834e4b}}.", "There exist few studies combining RC with transformer {{cite:8b0f574ec0a8215236c0e8ff2eab71a33514fc60}}. Shen et al. {{cite:1118d79b9c94b0eea4de730570913fc517d10f7b}} randomly initialized some of the layers in transformers without updates but obtained impressive performance in natural language processing. It only draws on the idea of initializing the parameters of the reservoir, not the \"true\" reservoir in essence." ]

[]

[ "Knowledge distillation(KD): KD{{cite:383053249f48b6c654b048e4ff5c09abe2a004fb}} is a technique used to compress a teacher network to a student network while maintaining the performance. The student model learns from ground true labels of the dataset and soft labels from the teacher model; using a higher temperature to shift the soft probability distribution enables a more useful soft label, and this a one method of KD from logits. And then, Hinton {{cite:9d070dd58272984b81f2d8910223e0c6fde0a361}} proposed to perform knowledge distillation by minimizing the distance between student and teacher network output distribution statistics. Furthermore, Fitnets first introduced hints, extracted feature results from the middle layer of the teacher network to train student networks {{cite:bcbe81b6b33e0912bc071da1ab8911c8f75afbcc}}. Further utilizing Fintnets, the flow of solution procedure (FSP) that fits the relationship between layers of the large model is also used to transfer knowledge between networks {{cite:fb8abc1af82a27c60051169490c55cea774b6e3f}}. And Zhou et al {{cite:bac2d8810f334c3f0c4356f714dd6cd34aa9f8a1}}. proposed each channel of feature maps corresponds with a visual model, so they focus on transferring attention concepts of feature map {{cite:c74f31ddc1fd7ef33525042efcc07ff90ce51448}}, {{cite:e2140dccdc4a2864e08d552eb85a644881abc9ee}} from each channel of intermediate layers However, some exiting approaches do not make better of different knowledge from different layers. Form each intermediate layer, the level of knowledge is different. The intermediate knowledge has large potential uses. In our paper, we will make more effective use of the dark matter in the middle layer so that the network can also use the dark matter that the model may not use.", "Self-distillation: Self-distillation means that the model distils itself, and we can say that the teacher model is an integrated version of the student model. There some approaches that introduce an informative thought in self-distillation. BYOT {{cite:ecabe6fc433ed94374ad70525b964e04a53204ea}} suggested that the last block of networks is teacher and the rest of the shallow blocks are students, where feature maps and soft labels generated from the teacher are used to train the student. Based on the training teacher network, it uses KD to train student networks. Furthermore the paper, the paper {{cite:c9bb295b93e8898bdde9a7e13a73723d0065cb2f}} posed the current model as a student and the previous model as teacher perform KD to train . Self-distillation greatly increases the accuracy of a model after training and shorten the training time of the model. Due to the advantages of self-distillation, we take this model instead of others. Self-distillation can accelerate the efficiency of training and easily extract knowledge from intermediate layers. Based on KD, the paper{{cite:c5f77eff7b8f3d4da00eda2dbe2c6b61c7f4b3b7}} generate refine feature maps and soft label for training network itself. In our paper, we use a different way than the way which use relatively refine knowledge to train the network, we use different multi-features from different network layers to provide rich dark knowledge to help network training.", "Deep Supervision: Deeply-supervised nets (DSN) utilizes classifiers at hidden layers to minimize classification errors and improve the performance inference. DSN adds the extra layer and trains this extended layer, supervising the whole networks we are training {{cite:74e209855daf7b983b42050310037a6d89b9c8c2}}. And this approach is used to solve the problems of DNNs training gradient and slow convergence speed. Applications about DSN are widely used, such as image classification and medical images segmentation and so forth {{cite:74e209855daf7b983b42050310037a6d89b9c8c2}}, {{cite:fc0a7e829c49c903a8e837280e3a4c9257e32fba}}, {{cite:2c79d0e2e4166a035e2f0d3ecfd509729fc7fee1}}. This method can effectively combine with other models. For example, in the paper Deeply-supervised Knowledge Synergy effectively combine with distillation; it is an improved algorithm for knowledge distillation in the middle layer{{cite:c94432b86555c9f559ad49157415bfec90d251be}}. Better generalization ability is obtained by adding additional supervision branches in some intermediate layers. In addition, a novel cooperative (snergy) loss function is proposed to consider the knowledge matching between all supervised branches through the distance between features. Therefore, we utilize deep supervision to interact with self-distillation and strengthen the model we create.", "Multi-view Data: In the real world, a data object often has multiple attribute sets at the same time; every attribute set constitutes a view. And there are often some views in a model, and we can utilize multiple views to improve our training. One of the co-training methods is a typical example of training Multi-view data{{cite:c94432b86555c9f559ad49157415bfec90d251be}}. There are some papers talk about utilizing multi-view. For example, one of paper considers that almost all perspective features will be displayed. However, in the input pixel space and in the middle layer perspective features exit{{cite:3fc6ca0fc5a89d0e217032c62279f743831b39e0}}. That is to say, we can extract multiple views in intermediate layers, and these views have different knowledge levels, which represent the thickness of knowledge. We can get more practical knowledge from multiple-view data." ]

[]

[ "In this section, we review three parts related to our method including top-down methods, bottom-up methods and point-based representation methods.", "Top-down Methods. Given an arbitrary RGB image, top-down methods firstly detect the location of human instance and then locate their keypoints individually. Concretely, the region of each human body would be cropped and resized to the unified size so that it has superior performance. Top-down methods mainly focus on the design of the network to extract better feature representation. HRNet {{cite:2aece2ab9440f0911574d2ee7a7edadb71bc9946}} maintains high-resolution representations and repeatedly fuses multi-resolution representations through the whole process to generate reliable high-resolution representations. {{cite:23494d7aaaacde1386bd3651b5e1ab15e2d806aa}} proposes a Channel Shuffle Module and Spatial, Channel-wise Attention Residual Bottleneck (SCARB) to drive the cross-channel information flow. However, due to inefficiency caused by the detection-first paradigm, top-down methods are often not feasible for the real-time systems with strict latency constraints.", "Bottom-up Methods. In contrast to top-down methods, bottom-up methods localize keypoints of all human instances with various scales at first and then group them to the corresponding person. Bottom-up methods mainly concentrate on the effective grouping process. For example, CMU-pose {{cite:74b94089be549c8486d58605ed0c8430132378ed}} proposes a non parametric representation, named Part Affinity Fields (PAFs), which encodes the location and orientation of limbs, to group the keypoints to individuals in the image. AE {{cite:4a918844dc24b177d5223b1af82d49c8a9a6cc72}} simultaneously outputs a keypoint heatmap and a tag heatmap for each body joint, and then assigns the keypoints with similar tags into individual. However, one case worth noting is that the grouping process serves as a post-process is still computationally complex and redundant.", "Point-based Representation. The keypoint-based methods {{cite:86ad3a279e97052331603e09f5bfdae0a1dfbef5}}, {{cite:47c8105b190d031a2043165bcbaf363fbc8b9470}}, {{cite:8e095a09f4c391c078abea5964b1d8487c871ba2}}, {{cite:3f83b246c73fbbd16bfc88eadaacc4ba93199a8a}}, {{cite:afbc738bc0eec3e876666b0063c6172279fe4cd6}} represent the instances by center or paired corners and have been applied in many tasks. They have drawn much attention as they are always simpler and more efficient than anchor-based representation {{cite:cd5cb4ebe308861a2232f1a245d8609d670f7cea}}, {{cite:54a4965f5d30ff5f2367e651dff0b78296705fe3}}, {{cite:1f375640d77815d76dee4dd954bcc702855d027a}}, {{cite:ddaada9abc14e5ccd6bdaaebbbdfc63197e1dd4c}}, {{cite:ae35470f02eb8c15b9cbeb77fa4363c02d245279}}, {{cite:72b729d4458dfb87bb52e623654eb477756b3c6e}}. CenterNet proposes to use keypoint estimation to find center and then regresses the other object properties such as size to predict bounding box. SPM {{cite:3f83b246c73fbbd16bfc88eadaacc4ba93199a8a}} represents the person via root joint and further presents a fixed hierarchical body representation to estimate human pose. Point-Set Anchors {{cite:ec3645ec7e410800bdc9ab67111fc2737a0b8cfd}} propose to leverage a set of pre-defined points as pose anchor to provide more informative features for regression. DEKR {{cite:74a8eb7fb71205b186db43807ae8c9f315d4021f}} leverages the center to model the human instance and use a multi-branch structure that adopts adaptiveconvolutions to focus on each keypoint region for separate keypoint regression. In contrast to previous methods that use center or pre-defined pose anchor to model human instance, we propose to represent human instance via an adaptive point set including center and seven human-part related points as shown in Figure REF (a). The novel representation is able to capture the diverse deformation for human body and adaptively factorize long-range displacements." ]

[]

[ "Community Question Answering.The current CQA task aims to automate the process of finding appropriate answers to questions in a community-created discussion forum {{cite:1a940dc886020e8b59c7c411bf6ffc8bc80c5ea5}}.Recent researches {{cite:45af7bb7e7037e6f009a4f7db3980b7360970af9}}, {{cite:6ad52b345029ae3ee940b0d834c95e910b96535e}} often regard the CQA task as a text-matching problem and have proposed various deep learning networks to learn the semantic representation of question-answer pairs.", "Different from the answer selection and span extraction methods, researchers propose the generative-based CQA methods {{cite:a9b4fd18da8b5792d995e900b29ae9946b11009c}} which can generate the natural language answers.Answering the custom question on E-commerce website is an important CQA application which has been addressed by many researchers {{cite:aba664398350bdd0942d3f6951bd6fe8e6510caf}}, {{cite:f2aec74ede00fb1802310681f3de5d15c3066ba4}}, {{cite:7e95c6c225877250afef84e7784ab64480efebf9}}, and most of these methods tackle this task as a knowledge-grounded text generation that leverages the custom reviews and product attributes to generate the answer.", "However, existing CQA methods do not consider the MIS content in the community, which can provide more related knowledge than using the articles along to answer the question.", "Heterogeneous Graph Network.Heterogeneous graph is widely used in many structure data modeling tasks: recommendation {{cite:d6cc2b35c8a3bb3cbe4d522b8a0d889db6ae13b8}}, {{cite:e2984fd01180536434266e3ba61afe541120ad00}}, {{cite:71183e089fdfef414ace12b0a3389f37025a3355}}, social network modeling {{cite:aaecc2053f473f08b2aabdd8a9ee075bcc660869}}, {{cite:37dc4cfbc054c882ab4f3a2861639ed59e20ba92}}, {{cite:e218296964a069d9ff24ef847cd31db49b80e125}}, text summarization {{cite:a1f5dd30ee704b721db64f6455294e76339204da}}, and knowledge-graph based tasks {{cite:b1c400c0675e5adce0736a0f73d10a6a2f1b3895}}, {{cite:3b6a9059543687190fdd8d2d4cd38818212380ff}}.In the CQA task, most of the researchers employ the homogeneous graph to model the contextual information, and only a few CQA methods leverage the heterogeneous graph model.{{cite:8035df15eca6b77ad5c0de2ceaf1985e1c1b5d8d}}, {{cite:cad2b4d07545abb764e325e2ad4aee88a84fb699}} propose to use the graph to model the multi-modal content of questions and answers with a graph neural network.{{cite:bf1bff63ae038bcd9ffd22ac8477a20f31a4211c}} propose a graph-based model to route the newly posted questions to potential answers, which models the relationship between question, tags, and users using a heterogeneous graph.However, how to model the different types of MIS text using heterogeneous graph has not been studied in previous research, and most of the previous methods only retrieve existing answers or select answer from pre-defined candidates.In this paper, we focus on incorporating MIS from the user community and generate a proper answer from scratch.", "Text Generation Pretraining.Recently, large-scale language models based on transformer {{cite:965c96ff300b4affaaabc1a5e0f7019079b0fb48}} pretrained with mask language modeling {{cite:5f360f4302f79e6977223a2793f6a6683e2b558f}} or text infilling {{cite:524a45a2ef23412ba0d5af75caf18c2b39179ba9}} have been explored further advanced the state-of-the-art on many language understanding and generation tasks {{cite:f946851d27cfd0054ce4ebeadc385127c0cd751f}}.We focus on improving the factual correctness and language fluency of the CQA model by using these pretraining models." ]

[]

[ "A landscape of the online process mining techniques was provided by {{cite:5aef2906a964b261a99e37e7382064065379b2e4}}. Being part of the process mining manifesto {{cite:567999f5e3aa35b4c838f8deac78c308413c3b17}}, OCC is receiving attention from the research community. Using regions theory {{cite:362558aeb0b64a2e6e5bdd1f4a0e09a322faf4bb}}, {{cite:0cdc3ee5749935226d99cf6f99f66c134a7f75e9}} added deviating paths to the process model and accordingly non-conformant cases following those paths are detected. Alignments {{cite:4ee42a6f1624f298ff33745ff31ef20b8695c66a}} is the state-of-the-art underlying technique for CC of completed cases {{cite:a551122dab78e8ee69494f2e429e2be639d1c678}}. Accordingly, prefix-alignments {{cite:8ee8c4eb4e743343ae241031c299da6b966b3d9f}} were introduced for CC of incomplete cases with a process model, without penalizing them for their incompletion. Decomposition {{cite:d34552bc7af5be019b7f0edbbec3a84676ead3d0}} and later Recomposition {{cite:f86334e7e04e46827b1b8868231fe05451042707}} techniques were proposed to divide and then unite the alignments computation for efficiency.", "Incremental prefix-alignments {{cite:72a22c8f164258ebae36f6de645dafe92ef78092}} combined prefix-alignments with a lightweight model semantics based method for efficiently checking conformance of in-progress cases in streaming environments. The technique of  {{cite:23c6e3ac16fd8f58a9601c38c9e50826747a582d}} determines the conformance of a pair of streaming events by comparing them to the behavioral patterns constituting the reference process model. Their technique is computationally less expensive in comparison to alignments and is also able to deal with a warm start scenario, where the event in an observed case does not represent a starting position. This approach somehow abstracts from the reference process model and markings there, therefore, it is hard to closely relate cases to the reference process model. The online conformance solution of {{cite:676d8ac2e6ab6cf02535eec87b14815f680d8825}} used Hidden Markov Models (HMM) to first locate cases in their reference process model and then assess their conformance.", "The availability of a limited memory, implying the inability to store the entire stream, has been sufficiently investigated in the data stream mining field {{cite:bf2a8e95d7a7715a80860eff67f81f5a247ce367}}, {{cite:fa117cc5dfee0e2188a53fabbc7c34a9c93f948d}}. In contrast, the memory aspect of online process mining in general and OCC in specific has attracted less attention in the literature. The majority of such techniques therefore assume the availability of an infinite memory. {{cite:02f88042d30ea5c62f98a89d605b3163ab809636}} generically suggested maintaining an abstract intermediate representation of the stream to be used as input for various process discovery techniques. {{cite:23c6e3ac16fd8f58a9601c38c9e50826747a582d}} limited the number of cases to be retained in memory by forgetting inactive cases. Forgetting cases on inactivity criteria may lead to the missing-prefix problem in many process domains. Prefix imputation approach {{cite:61773d89d32e3f4295759f8178dd96f9ea5960b3}} has been proposed as a two-step approach for bounding memory but at the same time avoiding missing-prefix problem in OCC. The technique selectively forgets cases from memory and then imputes their orphan events with a prefix guided by the normative process model." ]

[]

[ "The study of HAWVs is an emerging field in Mobile Robotics. In the last decade, the number of contributions to this field has increased significantly, with many proposed concepts and prototypes. Yet, it still has many challenges that need to be addressed, such as media transition control, underwater communication, sensor fusion, and planning strategies.", "From the first works regarding HAWVs {{cite:965c07fd1a8c8aec75a1b48cb276b6ed23996b0c}}, {{cite:e5555439741a688fdb14fb2002f206df3dfd3222}}, {{cite:33d24836c07f904b711679b606bd5a79adf501d8}}, {{cite:eda9179171330e77c7787448e47422942f8cebbe}} to recent studies {{cite:3150d338705f4f4ae6fe8c44737ec695cd08b32b}}, {{cite:cfcbce3b2fbd2e05b4df6cb1a8cc584fced5d729}}, {{cite:fc038bd961b9a2d5e30bb888826d81489e61e284}}, no major contributions were added to the high-level motion planning of quadrotor-based HAWVs branch. In recent work, Grando et al.{{cite:505fa069f447494bb2fb1d69e0288de10ff91c2d}} presented a navigation strategy based on reinforcement learning to HAWVs. Nevertheless, their method is limited to pre-trained cluttered environments. Planning is one of the challenges associated with this class of vehicles, given the need to efficiently navigate and transit between two fluids that are three orders of magnitude apart with respect to specific mass.", "Some studies have been conducted on modeling and control of HAWVs, as in {{cite:33d24836c07f904b711679b606bd5a79adf501d8}} and {{cite:fc038bd961b9a2d5e30bb888826d81489e61e284}}, wherein the authors propose a robust attitude control with provable global stability and a comparative analysis of nonlinear state estimation methods, respectively.Additionally, in {{cite:9e3c8d0051fc8f22b203e4bc3b0f27060d68c2dd}} and {{cite:cfcbce3b2fbd2e05b4df6cb1a8cc584fced5d729}}, the authors propose firstly a straightforward control scheme for a proof-of-concept prototype, and secondly a more advanced control system also adding the use of quaternions for orientation representation.Finally, in {{cite:eb6c695d678e9b2454116ffb04d63de68eb05a9b}} the authors present a hybrid controller designed for trajectory tracking considering the full system and preliminary experimental results, making this work as far as our knowledge goes, one of the few to address the smooth transition control. But trajectory planning is out of scope for all of them.", "The motion planning problem is not new to Mobile Robotics, having many classical approaches and proposed solutions {{cite:bae6ec2e4d91f52a93e66e8f73efa73b4fbd2175}}, {{cite:7679163bb98b1a6003727baedc507da7db35c1e7}}, {{cite:45038f423ee40f9bb8ea07e593a6ccd66962dc12}}.However, an approach that considers the vehicle dynamics is a key element to obtain a smooth performance, especially during the media transition. On the other hand, environments filled with obstacles, and more specifically with different navigation densities, are hard to address.", "One known approach that can handle nonlinear dynamics and complex workspaces is the RRT {{cite:57e41d47053a6e38e580d37c5210d84e35297d3b}}. In this method, the authors generate random trees that rapidly explore the environment by a robot model in an open- loop. This seminal work by Lavalle and Kuffner was later the basis for many approaches like the one with a closed-loop prediction by Kuwata et al. {{cite:491d5558cf912ad3d3809f0e0cf58f6f9b855cb1}}, called CL-RRT, and the less resource-consuming one by Karaman and Frazzoli {{cite:5f22f05237ad884763547294c2266c8d71f3c95d}}. Particularly, the CL-RRT is interesting because it allows real-time planning since it has an execution loop. It also is more robust to model uncertainties, external disturbances and even the presence of dynamic obstacles {{cite:9254e1b5876803fbd2198724f56b356ffbe0dc26}}.", "RRTs have already applied to hybrid systems, whose dynamics vary abruptly, often in the function of external events {{cite:408a5a692f2061b99233f428a6fb5d9161a7a833}}, {{cite:db105931cd4dbffb676bfe1eff8a8b4e57eaef9e}}. For some given conditions, they have proved to be probabilistic complete, an important topic for randomized planning methods.", "Other hybrid robots' studies have applied different techniques to deal with the motion planning problem. Wu et al. {{cite:318cde6517ee0b3fa9a0c717b4f548e11e87e458}} applied an adaptive and global-best guided cuckoo search (CS) algorithm, called improved CS (ICS) algorithm to plan for a water diving transition.Additionally, in a work by Wu {{cite:02d2005df8ca30cb49e1ec62c8d8d5875b5a0df3}}, a coordinated path planning between a HAWV and an AUV was developed using a particle swarm optimization algorithm.Finally, Wu et al. {{cite:183a887c89de5e68ae5ee8fd4dfc7d2f0ff595f2}} proposed an improved teaching-learning-based optimization (ITLBO) algorithm to strengthen the influence of individual historical optimal solutions in an environment-induced multi-phase trajectory optimization problem focused on the underwater target-tracking task.Recently, Su et al., {{cite:51a9bd2dd48181f4b16800b689420527dccd0227}} improved the (ITLBO)-based trajectory optimization changing the optimization function. The previous works in the field do not consider obstacles, and they are not suitable for real-time application. Thus, we are proposing a trajectory planning capable of (re)calculate in real-time and capable to deal with obstacles in the environment.", "Concerning the current literature in HAWVs, it is possible to see that most of the papers are concentrated on low-level aspects of robotics, ranging from platform design to control stability, especially concerning the transition problem. Therefore, in the sequence, we present a trajectory planning strategy, based on the CL-RRT, to deal with the trans-media navigation at a high level, with all advantages inherited from randomized planners." ]

[]

[ "Cluster analysis belongs to the family of unsupervised learning algorithms and is a technique to find groups in datasets {{cite:cc16b2f97063adf75a2bb29994b3d37e1772fba8}}.For swm data analysis, clustering can be used to segment users into groups with similar water-use behaviour {{cite:32ae5106c135b4eba8a5ee10a1f1ddfbf9bc81a9}}, e.g., commercial vs. residential or single households vs. multi-family homes.While customer segmentation was mostly focusing in the past on smart meters that measure energy consumption {{cite:3f209d8531bfc061c0bc03a4d51503e12dd361c7}}, {{cite:f0bb1845694f82cdef38abd4d55e3c5a2117ea96}}, {{cite:9714d8698fd166fea3448b2dad8e23feb4415397}}, {{cite:f91a321b0dd780667bcb5106b0c50f9f59750437}}, only few studies applied customer segmenation on swm data.Most research uses {{formula:06de1a5b-4fa6-4f5f-96fb-f0c0fde98ffa}} -Means clustering {{cite:fad62ef1c2a7a9a675b36ebc83de976c64013ebc}} in combination with different distance measures.{{cite:02744dc80c7f47341ece3de7313077eeedae3244}} clustered swm data into commercial and residential patterns by applying {{formula:4252e76a-8d84-4e60-bd49-c63eaccc0449}} -Means on features extracted through fitting Gaussian mixture models.{{cite:2da2f69674a044873d83a54b5a8d8732c3843127}} used {{formula:ea653482-3ebc-477a-b7ae-d166898f2df8}} -Means++ {{cite:ee69470d04901bd15366c773bfdcdcda2e578660}} with correlation distance to cluster data in residential and commercial groups.{{cite:00aa449628c25b021a7dd8e4c768462f2e24b240}} classified demand patterns using {{formula:77faf0ad-b7bc-409e-a49f-29ec65bae752}} -Means clustering.{{cite:78b049f4c060670106e9613fed3dc79688a93f89}} made use of Fourier-based time series models for clustering demand patterns.The patterns were qualitatively interpreted as residential, commercial, office, industrial or noise.{{cite:1023dc02a18922f9edebb1811750f64943b97ab2}} identified groups of similar households by features of their high-magnitude water use behaviours based on previous work {{cite:9dd658f5aa7d376d72394b577cc9ddecbf12a211}}, {{cite:f76fe77bee1a3ad537de9201b2abbce47d775b73}}, {{cite:6ad6b4fc15abc2a953463f501d810d9744b91c66}}.{{cite:032554c31b37ecde85637a7557615b7347870042}} applied customer segmentation analysis simultaneously on water and electricity data by clustering extracted eigen-behaviors and linked the clusters to a list of user psycho-graphic features.Recently, {{cite:32ae5106c135b4eba8a5ee10a1f1ddfbf9bc81a9}} coupled non-intrusive end-use disaggregation with customer segmentation to identify and cluster primary water use behaviors.", "Clustering techniques are also used as a prior step to demand forecasting.For example, {{cite:5927bba7c7121934668fa5e6bbf56749e6ff9b1a}} constructed clusters with {{formula:6050f63e-3825-43fc-8bb2-22a96d6aa6e7}} -Means according to their average weekly consumption before forecasting future demand.{{cite:046f8cc2c2eef7b4f998a4c57a77fa301767c06d}} clustered swm data using {{formula:4d4b7a76-8fb7-438e-9828-d1e63004f4a2}} -Means with cosine distance, first to split data into weekdays and weekends, then to split the data into residential, non-residential and mixed type clusters.", "In contrast to the studies mentioned above, this paper employs sdtw as time series clustering distance measure.This distance measure is capable of optimally aligning two sequences in time by non-linearly warping the time-axes of the sequences until their dissimilarity is minimized {{cite:46c5c2e51082b2733407f4b5deb830dfd46ea53d}}.The time when people use water is highly variable among different users with otherwise similar socio-economic characteristics {{cite:430e46c51ed0be20d66ce7d651c6b7427722eb21}}.The sdtw distance measure can expose similarities in daily schedules of inhabitant's water use that are shifted in time, which are not detectable by using linear time distance measure (Euclidean, correlation).Originally developed for speech recognition {{cite:ddc6e2249729d021738efd934fee8d3b73d03515}}, dtw has been applied in the field of water management in the past, but never in the context of time series clustering.Past applications of dtw in water management included burst detection {{cite:72d6c3fb1342378dee6e489b706d2bc94a68b1d1}}, analysing residence times in waste water treatment plant reactors {{cite:03a8b83da08e088e5023dfcb000ce7fba0fc2730}}, sewer flow monitoring {{cite:46c5c2e51082b2733407f4b5deb830dfd46ea53d}} or identifying water demand end-uses {{cite:fd5fb9b9e9868a06144c3bb3e6a46ee9139ab2fd}}, {{cite:d82115f4b9111d69b4190df5a2aa34ab544cc2f5}}." ]

[]

[ "This section gives an overview of related approaches that focus on assignment of disease codes using NNs trained with clinical text, bypassing numerous equally interesting results of applying these networks to other clinical domains {{cite:637a2c1d1fe37d383229378bd7cb381c955973fc}}, {{cite:fd55e6640111b361adfbbe86449f0cd67ac087b2}} and tasks, e.g., medication extraction {{cite:aad1fb08468fdcb42b38fa9be12d30d234795a17}}, {{cite:fe4cd846127fcf3b2fa3f4beb161d4c2ee217466}}, {{cite:f9cec03f51ff8cb3f2699d48a2272eda40934748}}, {{cite:dca152fb60ac6266d643f944535aa408a6915449}} or document de-identification {{cite:620754ab011217199dd01fddcb30a23afe821a95}}, {{cite:3b81d429e2e3b15582dd61d9eade1dec79a3348d}}.", "Emphasis to process this kind of data had also been given within several task-specific academic challengeshttps://n2c2.dbmi.hms.harvard.edu {{cite:251c80c83ba2e6542cfae3cdcb6a5e317e638134}}, {{cite:eca671507e42747e28091ef45c5be82dd4e4235b}}, {{cite:3a6086423e5c434480a7d660a9fbb10f6221a669}}. A recent systematic literature review regarding ICD coding systems was published by {{cite:5f66bc9cf04cd6eaf7cd4e07aa706eec7e094f63}}.", "{{cite:119ea7d57cefa3dbbb6e77c63d534e5eab1d870b}} applied convolutional neural networks (CNNs) to the recognition of ten disease phenotypes using 1,610 manually annotated discharge summaries from the MIMIC-III corpus. CNNs outperformed n-gram models and concept-based prediction models using the clinical information extractor cTakes in combination with random forest and linear regression models, yielding a macro-average {{formula:52901e1f-e6e7-44c2-aef8-a9028d4af861}} -measure of 0.76. In a salience analysis, they showed that the NN was able to identify phrases that were meaningful for their corresponding phenotype class.{{cite:58b2cdd7d123799e8952cf87e72ae1ca2e945fb4}} compared CNNs, LSTMs, both with and without attention, and an LSTM with a hierarchical attention approach. Using the MIMIC-III data set with 5,000 documents and the seventeen most prominent “level 1” ICD-9 codes, the attention-based CNN outperformed the other models with an {{formula:0174caa8-4b79-4955-a955-340729443d4d}} -measure of 0.73. Using the twenty most common “level 5” ICD-10 codes with 52,600 records, the CNN performed best, with an {{formula:1a52d808-d13a-4bdb-ae11-341de7fff80a}} -measure of 0.73. For the top seventeen “level 1” ICD-9 codes on the same data set, an {{formula:b0b9821f-6494-4c99-8015-404900ef0dd8}} -measure of 0.80 was reached using the CNN with attention.{{cite:8a02c2cd1d581c9f2ea548fe396c3d86ccbdb71a}} applied hierarchical attention networks (HAN, {{cite:5a6ab6051661214fe84a7d3264b0e47fdf886e0f}}) for information extraction from cancer pathology reports using 942 de-identified texts. Twelve ICD-O topography codes and four histological grading stages were recognized by the network. Their HAN outperformed conventional machine learning (ML) and deep learning (DL) methods, reaching {{formula:56c6c586-3024-4bf1-a0b2-b8eebde784e1}} -measures of 0.85 and 0.71, respectively.", "How to process the textual content of large EHRs using NNs exploiting billions of data points and to yield results comparable to human experts was demonstrated by {{cite:0fb2867598dc46ec62afddeb02fe29c80fe38f70}} and {{cite:752e1cf3d40debbe49310a83b5286bd7c26994fe}}. The latter achieved a macro-averaged {{formula:44c23e0a-20aa-4f35-b561-f33ba3252d64}} -measure of 0.89 by classifying thirteen disease phenotype classes, and {{cite:0fb2867598dc46ec62afddeb02fe29c80fe38f70}} reported the following weighted AUROCs: admission diagnosis - 0.87, diagnosis 24 hours after admission - 0.89, and diagnosis at discharge - 0.90.", "The main difference to the aforementioned studies is that our experiment uses very short, ICD-10 annotated problem list entries in German that were created as part of routine documentation by physicians." ]

[]

[ "In this section we investigate some of the available plagiarism detection systems.", "An Arabic plagiarism detection tool called Aplag has been introduced in {{cite:4513087bcb3184392f4d6b7e1bed1b4114feb3df}}. For developing the plagiarism checker system, they have extracted the fingerprints on document, paragraph and sentence levels for saving the computation time. If the similarity between hashes of the two documents is above a specific threshold, then the process continues to paragraph level and so on.", "In a work accomplished by Alzahrani et al., an intelligent plagiarism reasoned named as iPlag, has been designed {{cite:275a93ef02460092bc0e8ee82ee0dde5319bbec3}}. Scientific publications in same fields usually share same general information and have some common knowledge. Besides, each publication should convey specific contributions. In this system, they have processed various parts of a manuscript and weighted them based on their importance in plagiarism detection. So, the PD system pays more attention to the parts of a manuscript that has more contributions and lower weights given to less important parts.", "Meuschke et al. proposed Citeplag, a plagiarism detection system based on citation pattern matching {{cite:43c65fcb012ae19e8031fb6e5ed2f08b915d3878}}. They search similar patterns of citations between source and suspicious documents to find cases of plagiarism.", "In a work accomplished by Leong and Lau, they have developed a document plagiarism detection system named as Check {{cite:566a1c6cbc877aecbaddeb99d2771cb8014db470}}. They try to eliminate unnecessary comparisons between documents with different subjects and so reduced the computational costs.", "A word-similarity sentence-based plagiarism detection tool on Web documents, named as SimPaD has been developed in {{cite:f80297f5bad01dd3e7b798826ebec15fb21883ea}}. They measure the similarity between sentences by computing word correlation factors and then generate a graphical view of sentences that are similar.", "Collberg et al. developed a system for self-plagiarism detection named as SPLAT {{cite:90c092a4eb9cd9fa71ffad7c53f08beb46a07034}}. The system crawls websites of top fifty computer science departments and downloads the research papers. Then a text comparison algorithm compares all of the papers for instances of text re-use.", "It should be noted that most of the tools for detection cases of plagiarism are only pay attention to instances of verbatim copy plagiarism and cannot identify paraphrased passages of text which require semantic similarity detection methods. Our contribution in this paper is to use specific features in Persian to detect cases of paraphrased plagiarism." ]

[]

[ "The work of {{cite:fd4f5a2f5054d0dafbf8ff4d89039004e45e7243}} is closest in spirit to this paper. They also establish dynamic regret bounds scaling with a generalized notion of shifts based on changes in the gaps between mean rewards. However, their notion of significant shift is weaker than ours for they admit changes that do not switch the best arm's identity. Also importantly, their algorithms need to know the number of such shifts, whereas ours does not.", "The work of {{cite:e18af5c28fd314abace87ba317b3f999c9a8e516}} also studies dynamic regret in the randomized adversarial bandit problem. They provide two algorithms that get regret {{formula:ffb7d0b4-9a47-4932-bf7e-d38973e3caf2}} and {{formula:feb27130-0d3c-4a3c-b2f6-c05a681ad71f}} , which do not match the optimal regret of {{formula:bacea969-e6f6-42c9-acd4-5c7309c60a11}} . Furthermore, their procedures either require knowledge of {{formula:0e439be1-9166-4da1-a1ba-e444f667aea2}} or structural assumptions to allow easier detection of best arm changes. We impose no requirement on the structure of changes, and achieve faster rates than {{formula:c1164f31-c5e2-45bf-b211-9c94a0974157}} .", "In the aforementioned switching bandits setting, several non-stationary variants of UCB achieve regret {{formula:4dba8b57-cccd-4406-8ece-c50371aae2bc}} , albeit with knowledge of {{formula:7647d5f9-3733-46b8-a2d1-927acacb649e}} {{cite:e38b35622434ab5ee3e9283d2e9676055e95ad44}}, {{cite:90a79f5ec27159fb74916a84b7228a895a8061b5}}, {{cite:6a6617b6fdb6fa061855eaadb72d81dd7419e28a}}. In a similar vein to this paper, distribution-dependent regret bounds, of order {{formula:ed0fe174-54dd-4844-8745-16545db60cc5}} for environments with large gaps, have been achieved by {{cite:9d27774a0f8858809a4eb47f9c350f53600796a6}}, but with strong structural assumptions to ensure easy detection of shifts.", "Another line of work concerns drifting environments, where reward distributions may slowly change over time. In such environments, a switching regret of {{formula:4b30a9f7-1b2c-4215-b244-61c4531b16f7}} is pessimistic when {{formula:8d2d3fd2-64f9-457c-8f9a-2af55cdb7782}} is large, while each shift involves small changes in mean rewards. Such changes in mean rewards are therefore captured in aggregate over {{formula:9b7cbd19-625c-4ac1-8cf5-7f367c05a2f4}} rounds, by a so-called total variation quantity {{formula:6260d929-cb32-4d8d-8f40-dc1a6ecfa0f1}} , yielding an optimal rate of the form {{formula:02de2246-a812-4d41-abd0-4a5f1f0553f7}} {{cite:116e6d5bd97dd151628e8be2ae2ea78d67e7f014}}, {{cite:9776405e23cd32ecbaf294cedb9271b7ab008e34}}, {{cite:403d88313673f2f416409c497526dfdc23eba677}}, {{cite:bb0ad8a18883471b5dab90c88a6c9665324a91ae}}, {{cite:a7f3be1b9be486fa9f4939c9378649952e171100}}.Comparing to our work, we note that {{formula:e2f27369-260f-48ea-9f9e-e27c7141181a}} accounts for all changes in the mean rewards, while our notion of significant shift only tracks those which are truly severe and long-lasting. As a result, a significant shift necessarily corresponds to a large magnitude shift in the aggregate and thus we recover the {{formula:07bfadb7-0206-4f98-b12d-56f02ab2e71d}} rate without knowledge of {{formula:45da5552-f46f-4676-bb67-7a3a8cd31268}} . Going beyond this, we achieve improved rates {{formula:ce5f5206-cb4a-423d-a6b7-2028a0590b74}} in environments exhibiting a small number of significant shifts {{formula:48606368-8910-4633-9ad9-a69f68b44a6c}} but large total variation {{formula:cfa81670-9f2c-4220-900e-bb46949165d7}} ,", "A different body of literature considers more structured changes in reward distributions. In rested rotting bandits, the reward of an arm decreases depending on its amount of play {{cite:9a7a13a80f5a2875fef806485654e52e90cd00ed}}, {{cite:2151a9f8e6da588a7bf47718ab2982e2dc781e7a}}, {{cite:6784d704c4954c1ad78e7ee00bd3b0cc469001e3}}, {{cite:71115110c95c70f8818106a27534193b78c8f262}}. {{cite:15e052d231066b21ee4547311409f3436d51defc}} study a setting where the rewards follow a Brownian motion across time. Several works also studied a subcase of the above mentioned drifting environment: slowly varying bandits, parametrized by a local limit on the change of rewards between consecutive rounds {{cite:b134e56cfa1f2a568f321ed0c5c87a267f770767}}, {{cite:8d59b58359dbf200e5a4ca4a271c6a1601d2c877}}. Generally, such stronger structural conditions can yield faster rates at times, often of the form {{formula:3644f165-e7be-4b02-8a98-608e5348bb5d}} , for some problem dependent quantity {{formula:0f5a1406-be6e-40ba-bf38-58b212b12408}} .", "Finally, various works combine adversarial and stochastic approaches to simultaneously address both settings {{cite:345ffc67651f4b9534dbfd1b7aec11650535c1fb}}, {{cite:1a285019f0caad90133fd708923485fc86a27a48}}, {{cite:b5d1aadef82ba7575ed08c8ac245426353a46d2f}}. However, these approaches measure regret to the best arm in hindsight, whereas our results hold for the stronger dynamic regret." ]

[]

[ "Deep-RL surveys in robotics tasks for the navigation problem have already been performed to demonstrate how effectively we can solve this problem with learning approaches {{cite:49d9a9f3f05b45dc307c1c9277b9331bc974afe2}}, {{cite:cf7dbb9b2dc7e79e659e5a4a86528a579b34b50a}}, {{cite:529491322c13029afad0acac4b56cc0cf8b27416}}, {{cite:3ea8b6b5953921f49a25ab782676099acb9ae386}}.", "Tai et al. {{cite:404016cc4067f3ba6aef7e71e5add9325d02b18c}} used 10-dimensional range findings and a goal location as inputs and continuous steering commands as outputs for a 2D mapless motion planner when applied to mobile robots. They concluded that an approach based on the DDPG method with a fully-connected ANN of three hidden layers would effectively train a 2D mapless motion planner. Furthermore, the process of completing the task of reaching a predetermined goal was completed successfully.", "Zhu et al. {{cite:5556eb95d537321cb9daac19be5e874d3e44efdc}} used an actor-critic Deep-RL model to tackle the challenge of driving a mobile robot. It took the current state observation and an image of the target as inputs, producing the action in a 3D environment as output. Chen et al. {{cite:e16b98f5af7a8490485fd4e966e2b0506e3376d0}} discussed the issue of navigation in a pedestrian-rich environment. It proposed a Deep-RL approach to control a mobile robot that could navigate in a speed close to a human walking speed. It also performed in a scenario with multiple humans simulating a crowd. Ota et al. {{cite:14127e7a104accd93265de719081dea0d1cefd36}} developed an approch based on the SAC technique to efficiently navigate. It was also done in simulation, with the agent learning the course.", "For UAVs, however, Deep-RL has been addressed fewer times. Rodriguez et al. {{cite:81d98b5bebda631cf086d6be844e7be50a812818}} used DDPG to solve the problem of landing on a moving platform. On the Gazebo simulator, they used Deep-RL in conjunction with the RotorS framework {{cite:87b778edd4334eed3e21be1412739ece419bed90}} for aerial vehicles. Sampedro et al. {{cite:b6701625beba5d17a36183ad67bd88b2f443b98e}} suggested that an approach based on the DDPG technique could perform a task of Search and Rescue in an indoor environment. Sampedro et al. {{cite:b6701625beba5d17a36183ad67bd88b2f443b98e}} made the use of a real and simulated UAV to perform this navigation task. Kang et al. {{cite:0e13db9a661253e3d33f8315cc8026dc4b25d073}} focused on tasks realted to avoid collision with obstacles. This work also used visual information and real and simulated data.", "He et al. {{cite:b3c25dde5171727265bbf3143120975bc9dbc4cc}} combines imitation learning and TD3 technique and applied on the challenging 3D UAV navigation problem using depth cameras and sketched in a variety of simulation environments, obtaining great results. Li et al. {{cite:d4f82bee212dcba9723f70a5bcd486460c57b3e1}} combines Deep-RL with meta-learning and proposes a novel approach, named Meta Twin Delayed Deep Deterministic Policy Gradient (Meta-TD3), to realize the control of an UAV. Thereafter, Li et al. compared your method with DDPG and TD3, obtaining a substantial improvement in terms of both convergence value and convergence rate.", "Differently, our method addresses 3D mapless navigation and waypoint mapless navigation tasks by relying solely on laser sensor readings and the vehicle's relative localization data. Since the TD3 and SAC techniques are state-of-art for terrestrial mobile robots, we based our approach on them. We also make use of a LSTM-based structure instead of a MLP structure, showing that RNN based structures are better suited for this problem. We also differ by imposing environmental difficulties such as wind, which makes the overall results worse but more realistic." ]

[]

[ "The determinization problem we consider in this work assumes that the deterministic target model is also that of OCNs. An alternative approach to simplifying a nondeterministic OCN is to find an equivalent deterministic finite automaton, if one exists. This amounts to deciding whether the language of an OCN is regular. This problem was shown to be undecidable for OCNs in {{cite:b3f881567f2c52def76dc615d93c598889864abc}}. Interestingly, the related problem of regular separability was shown to be in PSPACE in {{cite:e6b63dab7991ffa76bfb179cd28e0510572c9a2f}}.A related result in {{cite:0dd2171bdc0efaa9550082cc5e6c01a3b9f187f7}} describes a determinization procedure for “unambiguous blind counter automata” over infinite words, to a Muller counter machine.", "From a different viewpoint, determinization is a central problem in quantitative models, which can be thought of as counter automata where the counter value is the output, rather than a Boolean language acceptor. The decidability of determinization for Tropical Weighted Autom49ata is famously open {{cite:890120e69bb0c27c521e32a6461b3c1e868fcb99}}, {{cite:aefb7bd26ba2adc4d8f6bdd23fb3a99ee4b1082b}} with only partial decidable fragments {{cite:aefb7bd26ba2adc4d8f6bdd23fb3a99ee4b1082b}}, {{cite:7f690f522c4e68d27eaaf7354ebfd1e978c4c1e4}}. A slightly less related model is that of discounted-sum automata, whose determinization has intricate connections to number theory {{cite:4d359bc1efcb662f770271a0c92e543c24cd9ee2}}." ]

[]

[ "Sublinear-time approximate {{formula:5a682bc6-e0d8-465f-8eeb-9078019191a9}} -median clustering of a space in which the diameter of points is bounded was introduced by Mishra et al.  {{cite:aae50a22421855fe4e69b27a5078d14b633335f6}}. They modeled clusterings as functions and studied the quality of {{formula:3440f4ae-0394-42eb-aadd-0a3c9e9c85aa}} -median clusterings obtained by random sampling using computational learning theory techniques. For a metric space, their work shows that if we sample a set of size {{formula:305c35d3-f16d-4a34-b4fb-24f708a3bf10}} and run an {{formula:a3952503-c7eb-4d13-9070-20570763ca4b}} -approximation clustering algorithm on the sample, then with high probability, the set of centers outputted is at most {{formula:2de2d449-2b27-421d-b6d7-bc76e134754b}} . Their sampling model was adapted by Czumaj et al.  {{cite:6405692178a356e8b99799b751626744518a0781}}, who achieved a sample complexity that is independent of {{formula:38a1134e-507a-48fd-b2c7-82c5a65ba6c4}} for the {{formula:e2ae8c42-7a2d-42f1-9cbf-948b4ff5c8b3}} -median clustering problem in arbitrary metric spaces. They also extended the random sampling model and their analysis to give sublinear-time results for clustering variants such as {{formula:5fefa0f9-4c9a-4a0a-96ed-d87d01849ba5}} -means and min-sum clustering.", "Private clustering was first studied by Gupta et al.  {{cite:91da3633e46c73bc0671572eeca5b335d65cb37e}}, and Feldman et al.  {{cite:6ef96a2386ffa79137016d44d25faba6f8742a88}}.Gupta et al.  {{cite:91da3633e46c73bc0671572eeca5b335d65cb37e}} modified the local search algorithm for {{formula:d036d80c-5f22-4f5a-a374-8ee40dccba8c}} -median by Arya et al.  {{cite:c7c8c2a79ea74637426f8d948fb20169172d17ef}} to choose candidate centers in each iteration via the exponential mechanism {{cite:b6f1ad984899f10c32768c30f9ec7ebd432a0dca}} and produced a polynomial-time algorithm that achieves {{formula:690e9fd1-d899-49d1-9924-bf01b7062628}} -approximation ({{formula:82d1b25a-5e9c-4cf9-bde6-0e8c902fefcc}} is the diameter of the space) in discrete spaces. However, their algorithm is highly inefficient in Euclidean space (see {{cite:1dedf5a489ba616c38df6e0203e99d3ed4b43456}} for a detailed exposition).A recent line of work has focused on producing an efficient polynomial time algorithm for clustering that achieves a constant (multiplicative) factor approximation in high-dimensional Euclidean space by adopting the techniques of Gupta et al. while maintaining efficiency {{cite:61086ae5a02fef6b84cd937071979b1f8fe53afc}}, {{cite:1dedf5a489ba616c38df6e0203e99d3ed4b43456}}. A different approach to private clustering was taken by {{cite:6ef96a2386ffa79137016d44d25faba6f8742a88}}. They gave an efficient algorithm for {{formula:7fad59b1-f8f7-4189-abd3-abf4a2ada0e0}} -median and {{formula:1ad31ec8-af7a-43d5-935e-05e36d7b6e7d}} -means in Euclidean space by introducing the notion of private coresets. A recent line of work has adopted their techniques to give clustering algorithms with better approximation guarantees and efficiency {{cite:6dfbdbac8bbfc151aa5bff3f2bcc5aade3acc389}}, {{cite:5b9f9dd624c5f2f797028520eaa60f0508464b60}}, {{cite:2052f6c4bc989932bd32d3d28df7e63117b4c7f2}}.", "Privacy amplification by subsampling has been formally studied by Balle et al.  {{cite:7297011e357800afa58b40a3fc9db70a0995fb89}}. Our result is a simple observation that tailors the privacy amplification achieved with respect to group privacy for a generic sampling algorithm that runs a private algorithm as a black-box in the sampling step." ]

[]

[ "Grounded Language Acquisition from Text. In robotics, language is grounded in real-world actions and percepts, whose applications include following task instructions {{cite:6e6fc7ea96f880bc250d8bbb2bc811cbd78764cc}}, {{cite:2921420efce86a9a6cdb8eafddfa9ccaa6c6ab68}}, {{cite:27795fe1ab93055c6efc853cccf55d848191c298}}, navigation instruction following, {{cite:2261a4694c079d320a12746261f2fe619b4207b7}}, {{cite:99666e0a37adacb80cdc6060e0656ef4a4cf9da3}}, and learning groundings from human-robot dialog {{cite:fbcb3d5d194f54e0ddc9f6f71152b45108dd2477}}, {{cite:e31381e6f969e7a9ccd4ea4bcfe99d2b0bb4e0d9}}, among others. While some of these approaches use text derived from ASR, none use speech directly.The focus of this paper is on learning language groundings directly from speech; we demonstrate this work on the common grounding problem of object retrieval {{cite:59ad55b5f43ba331a3b39105b7558ae9698d7f64}}, {{cite:a21d2231a15cdeb7af40358182e523cfe0ed10ee}}.", "In vision, grounded language typically refers to how language refers to existing images. From image and video captioning {{cite:a3e84e766286b1056a3f731cc32d45dccb622ab1}}, {{cite:263eb6761c6a088270184bf30be69fca194ba3c9}}, {{cite:587fe89280236105e2f000e7a323beda6190f4a6}} to large-scale pre-training {{cite:6fa5b8b5b4028c1be8aac071600a809a45404e59}}, learning from vision-language pairs is an active field of research.In this work, we use the manifold alignment approach of {{cite:59de0e2c93b5c7685751c9624166c45a16f77c50}}, in which language and vision representations are projected into a shared manifold, which is used to retrieve relevant objects given a natural language description. The novelty of our work is not in the triplet loss learning method for multi-modal alignment but in the comparison of transcription-based versus raw speech methods, and analysis of performance for end-users.", "Spoken grounded language learning. While the majority of existing grounded language learning is performed over text sources (either typed or transcribed), there are exceptions that demonstrate the importance of learning directly from speech. In early work, {{cite:8131d85d02d7f27a1bd67bdaf19a598ca5629c93}} presented a grounded speech learner that segments words from continuous speech. Our problem is more complex, in that we aim to ground full descriptions rather than words. The work most closely related to our research explores using audio-visual neural networks to learn semantic similarity between single images and raw spoken utterances {{cite:86a99b84b8c6d4832622c0139db2ac07d5d79cf3}}. By contrast, we focus on multi-frame RGB-D percepts gathered from a sensor, aiming to identify individual objects rather than entire images.Our work is most similar to that of Chrupała et al. {{cite:245a0f78c554664ea627f7858961879e125262a9}}, {{cite:905e55e35caf0d47dc8c058e71b06c58af0fe8a2}} and {{cite:e928384497519f50012820f25fb2e30764befc4d}}. However, the speech corpora in those works are collected by asking speakers to read captions of images aloud. This may remove grammatical constructs, disfluencies, and speech repair, effectively gating the complexities of speech through written language. The dataset used in our work consists purely of people describing objects.", "We are not aware of previous work comparing grounding from raw speech to the widely used transcription-first approach. Additionally, we show how to create a speech-based grounding system based on complex perceptual data using a comparatively small number of data points, which is consistent with the requirements and available resources for implementing on robotic systems. Compared to previous work, we leverage depth information and pretrained speech representation models to ground naturalistic spoken language in a model which converges with fewer data pairs; {{cite:86a99b84b8c6d4832622c0139db2ac07d5d79cf3}} used 402,385 image-caption pairs and {{cite:905e55e35caf0d47dc8c058e71b06c58af0fe8a2}} specifically mention that the Flickr8K dataset of 40,000 image-caption pairs is small for the speech task.Finally, we avoid the computational overhead of fine-tuning the model for specific domains, which may change as the robot experiences new environments.", "Language and Speech in Robotics. The role of language in robotics is wide-ranging {{cite:a073f02af94f51512c3ed46d97c8e0005860ea44}}, and the role of speech in particular is starting to receive significant attention {{cite:3099ccad8df59f45b5eeab6f5b7dd0bb938170c2}}.Natural language is widely used in HRI tasks, for example in dialogue with assistive robots {{cite:9a14d6d05e3c329725dca26b1daf8362264c0b8e}} or to facilitate human learning {{cite:171cb7886bcff40c1fc1df07522615d4ea4e5d6c}}, {{cite:0696eb750b1fef3a8ec2ee4119b7a49c911033bc}}, {{cite:7ae7af8563f18ae532707daa96ec790280a031d6}}, {{cite:63dbe507286e5f24e651f3250f011f111f23bbe7}}, {{cite:e8360dbb70283a9638868e55c53a855071a2efcb}}.Speech-based HRI, in particular, has been applied to a wide variety of problems, such as emotion recognition {{cite:1c6e4fa30aabaab34ac8ce194c6207cbb5da6d54}}, {{cite:f95a3d3f3c1f2b726e8fd5985f684223a0fef68c}}, social robotics {{cite:0dead0f0039ea8e9958d99195ad92424642ee97f}}, {{cite:d6714d24de96dd85da874e2e293daac9a38543c4}},and speech recognition. {{cite:e3ee12e3c826de2b8595128916249621ac81d3f2}} determine how a robot should position itself for optimal speech and gesture recognition, complementing work on how people expect a robot to react when given instructions {{cite:ea365c3a639edfae03e5d3c429f8c4f9b778f5ac}}. Speech is also an important source of insight into how different groups interact with robots, for example in assessing the communications of dementia patients through speech features such as pitch {{cite:a111dfbf1ebe7052233dfe28f002daae64ddba2e}}.", "Speech Processing Bias. While much previous work relies on ASR systems, these systems have known biases in their ability to recognize speech without errors. For now, most widely available ASR approaches depend on large-scale data {{cite:9d454a43a0e93cc9df09e26f2439e6efd0021333}}. These large datasets are usually derived from fairly heterogeneous groups {{cite:c67306dbfc09bf68a2e4f2fbdc1a69bae2790d4e}}. Given this, we seegender {{cite:55e6ba7f190e661a8f90a78f07a3714b8f55d596}}, {{cite:e8f33baa661c5026c5bd597e1d6d76825ac67a16}},race {{cite:a2983e052c7208a942639c9c75d4641e46d8ba3e}}, disability status {{cite:b3139bbd1649a46fadcac9bf33ea5d00c80b5a26}}, and native language/dialect {{cite:99edddfbadeaa8b0a72449472d4ac26a6f5984b3}} disparities in successful ASR, reducing technology accessibility for under-represented groups. Contemporaneous work {{cite:83d63fb8d5f4ec525db8dc826c90aec14ca99ef7}} introduced a dataset to measure ASR performance across age, gender and skin type. Technical remediation approaches remain scarce {{cite:b701844f1bdab9608beecc38d7b184cb5125cbca}}, {{cite:7e7054988cb59c6e1c2b8b5f761ff8fffd43c5e9}}, although there has been some work on multilingual grounded language {{cite:6c021e37e07aa334bdcd2add2182a20e18bc1219}}.To our knowledge, no previous work in grounded language acquisition has examined the impact of these factors in speech." ]

[]

[ "Multiagent RL. The simplest form of multiagent RL is Independent Q-learning (IQL) {{cite:5163ebf69d7fd37abde33e8d53d4e17097a90f35}}, where a group of agents each learns on their own and treats other agents as part of the environment. While IQL delivers decent performance, the agents frequently face the challenge that the environment appears unstationary, as other agents simultaneously learn and update their policies. To address this issue, centralised training, decentralised execution framework was first proposed separately by {{cite:ab200b292b82f15a1aa81ff63738e37cccf97da3}} and  {{cite:d211b2d8aba37b907b701a0215aa26affd6f9b05}}. This training paradigm allows the agents to access the global information during training and has been a standard approach to recent multiagent RL algorithms since then. {{cite:ab200b292b82f15a1aa81ff63738e37cccf97da3}} proposed an actor-critic algorithm that addresses multiagent credit assignment explicitly using the counterfactual value of each agent and is applied to domains with discrete action space such as Starcraft. {{cite:d211b2d8aba37b907b701a0215aa26affd6f9b05}} proposed a multiagent policy gradient algorithm where each agent receives a separate reward and uses their critic. This method can apply to competitive settings but does not address the multiagent credit assignment problem. Another line of work using centralised training, decentralised execution, and Q-learning are the implicit credit assignment methods such as {{cite:8918e05d49e89c0ac379658c7dc6995300370733}} {{cite:1f967acecf9cb2ce355182417ca87f65f05cab2f}} {{cite:e731a0c8af376499c47fe59f5812125789c13f2c}}. More recently, {{cite:7c3a56a68e7920bddb9f2f1b4e0f98961ad4b826}} showed the effectiveness of PPO in multiagent problems for discrete action space domains. Their model is optimised using a centralised reward and demonstrated to outperform baseline methods such as {{cite:d211b2d8aba37b907b701a0215aa26affd6f9b05}}{{cite:8918e05d49e89c0ac379658c7dc6995300370733}}.", "Game-theoretic Credit Assignment for Multiagent RL. Several works have considered the Shapley value for credit assignment in multiagent RL with discrete action spaces {{cite:6c00941a4a6b21f05dc47d982ea98012fddfee98}}{{cite:e85caa9dc1bc158f2222d51e8569cb9a13573904}}{{cite:3e15c408cde8a3c73456beb8ecbbb35187a31dc8}}{{cite:9051ccdf1c876e2abaf0b674eb8b76f9b3e562a0}}. Most notably among these works, {{cite:6c00941a4a6b21f05dc47d982ea98012fddfee98}} has shown state-of-the-art performance on Starcraft. In our experiments, we adapt their credit assignment framework from discrete domains to the continuous multiagent PPO. Unlike most prior credit assignment methods, which only consider one of the common values (e.g., Shapley value), we introduce semivalues to credit assignment in multiagent RL and define a more generic game-theoretic framework that encompasses a family of common solution concepts and analyzes the relation between them. Moreover,we use the new framework to address the problem of multiagent continuous control, and show that our model-based RL module can better estimate the coalition values for credit assignment and improve the sample efficiency accordingly.", "Model-based RL.Model-based RL approaches typically alternate between fitting a predictive model of the environment dynamics/rewards and updating the control policies. The model can be used in various ways, such as execution-time planning {{cite:902c96f5ad3e1181e01c7def36f606a010d1c984}}, {{cite:95eadc1537e210600c8749551c6869547f755d44}}, generating imaginary experiences for training the control policy {{cite:7e00f324803749b9b051f61b487fb8bc361c4567}}, {{cite:e01ef82cb6e6d73b11b604f9c8cc3521d3c1924e}}), etc. Our work is inspired by {{cite:0c6e4a89dbc956cec36dccdd97325e3a42c38e44}}, which addresses the problem of error in long-horizon model dynamics prediction. {{cite:0c6e4a89dbc956cec36dccdd97325e3a42c38e44}} presents a hybrid algorithm that uses the model to simulate the short-term horizon and Q-learning to estimate the long-term value beyond the simulation horizon.Unlike the previous model-based RL works, our work is the first to introduce model-based RL for enabling game-theoretic credit assignment in multiagent continuous control." ]

[]

[ "Electronic prescription systems operate in a multi-stakeholder environments. It requires the integrity and transparency of information to avoid illegal drug sales while preventing patients' health problems due to drug overdose. Besides, the application of privacy-preserving techniques for medical records is another requirement to avoid misuse of sensitive information present in prescriptions. {{cite:b2721e609fe087cb56773f02128ffaf80d8e1e9f}} proposed SecureRx {{cite:b2721e609fe087cb56773f02128ffaf80d8e1e9f}}, a blockchain solution using the Ethereum platform to maintain patient records and prescriptions. Similarly, {{cite:97924920676ada6b92992217bc9763e4b262bdc1}} suggested RxBlock {{cite:97924920676ada6b92992217bc9763e4b262bdc1}}, a solution using Ethereum to manage prescriptions and avoid drug overdose. {{cite:d1b77866a766a00efdc434bcefb9807309590150}} {{cite:d1b77866a766a00efdc434bcefb9807309590150}} proposed a decentralized e-prescription system using smart-contracts on a BFT platform. However, these solutions do not manage consent and focus on writing records to an immutable ledger without providing mechanisms to track who accessed the data and for what purposes while protecting patient's sensitive information.", "Other research works investigate approaches to ensure the integrity and privacy of medical records by preventing tampering and data leakage. {{cite:07de86ee0770915016035483e0b1e98b177ea37b}} {{cite:07de86ee0770915016035483e0b1e98b177ea37b}} presents a model to preserve patient data between different institutions using asymmetric encryption, particularly proxy re-encryption (PRE) on permissioned blockchain. Similarly, {{cite:40598a41f36e65984c4f95e2b876747d654d58bb}} proposes SPchain {{cite:40598a41f36e65984c4f95e2b876747d654d58bb}}, a blockchain and PRE-based solution for sharing electronic health records (EHR). {{cite:2c980a9be49b8eb115e4b7bcce28c4e5355ed243}} {{cite:2c980a9be49b8eb115e4b7bcce28c4e5355ed243}} investigate an electronic prescription system with the adoption of the k-anonymity method based on differential privacy for data protection. {{cite:5225a8e331a4fa60845a45821b6c2c924edbbc4e}} introduced DMMS {{cite:5225a8e331a4fa60845a45821b6c2c924edbbc4e}}, a solution that exploits blockchain technology for medication history management and electronic prescriptions. {{cite:74e907812c19517b70094099c06f8c4ec30b5c77}} proposes MedBlock {{cite:74e907812c19517b70094099c06f8c4ec30b5c77}} and {{cite:197da6631632a7728d9f6c2fb0ad9f0cfb48bc91}} {{cite:197da6631632a7728d9f6c2fb0ad9f0cfb48bc91}} presents a blockchain-based solution for sharing medical records with privacy and threat analysis. Table REF summarizes the related works and highlights the gap." ]

[]

[ "Forecasting solar PV output is akin to forecasting solar irradiance, since the former strongly correlates with the latter  {{cite:b0b12750444ec847c77bd40228bd715991d31c7b}}.Numerical Weather Predictions (NWP) algorithms {{cite:f4ec50c2d9417ab18fdb27adf777e0d6812cb2f5}}, {{cite:78806d6a71872cd06ba7de38879c188ec83feccb}}, {{cite:905813240d0063eaf09ef692ba1d7b77325c61e1}}, {{cite:a89f6c1ffce30445ee7c89d460cafc19dc63283f}}, that mostly leverage physics-based modeling, are often used for solar irradiance forecasting.These physics-based models are most appropriate for forecast horizons on the scale of hours to days, and not near-term forecasts on the scale of minutes to an hour {{cite:4a3619e8d3af2eac05fe65f425645d6d3ebf518f}}, {{cite:f4845bc869071960bbda6191a10159bfcf4d7de8}}.Over long-term horizons, the complex and non-linear evolution of climate patterns can be difficult to model, requiring knowledge of climate processes and the history of many atmospheric events over time that can cause subtle changes.", "On the other hand, at shorter time scales of 5 to 60 minutes, machine learning approaches have the potential to implicitly model local changes directly from observational data {{cite:f4845bc869071960bbda6191a10159bfcf4d7de8}}, {{cite:bc80b94d779830f2a2f16cf52fc1d4f86f264111}}.While there is recent work on analyzing images from ground-based sky cameras {{cite:63c193c6481ac1d957d944a9f4df6dbc94da2f23}}, {{cite:12d6877135892cc5a4fa96a0641bd6ad04ba8ed3}}, {{cite:32c609c9fd97753c1656d78d4a0cc1bafa0885bc}}, {{cite:b94d294c86d943f8873eb178b7f9375fa0df7fc7}} for near-term solar forecasting, it requires installing additional infrastructure at the site.Another alternative is based on estimating cloud motion vectors {{cite:83af445c23d823eb9dd7096f3931f38bea527172}}, {{cite:e94b567c13436b8e67c7ebc5e92deae1ae2898a2}}, {{cite:e605670536b0cb94e599bec3b2cf3508e2bbb326}} from satellite images, however ML approaches that more directly model solar irradiance tend to perform better {{cite:75b3ccc5484048d6346bfa42f6e5e274bee06962}}, {{cite:e97905e68b2ce91654994029cf0f0f73e75100df}}. Our approach differs from recent approaches in solar nowcasting by forecasting solar irradiance values from multispectral satellite data using a combined CNN-LSTM, which can forecast changes in spatial features over time. We then combine these solar irradiance forecasts with a model that predicts a site's solar output from solar irradiance.", "Our approach is a self-supervised approach, where we directly use abundant satellite data for modeling. Such methods have gained increasing popularity in computer vision recently {{cite:eda2aedb4bfd0fd6a89e8f4a44e03a944b0bd68c}}.While such methods have been widely successful, their applications in remote sensing have been limited and their applications to solar modeling have not been explored before.{{cite:704ac2f0ee42685bc27f04408d40b47ef752441e}} similarly uses self-supervised learning over Landsat images, although their approach is designed for classifying geographical regions and not applicable here.{{cite:014b0097e8a067138b34315733df3dd981c5161f}} reconstructs visible bands from other bands, in a colorization task, to learn useful representations for land cover classification.Related to our work, {{cite:daad234662a2bdf8b96a2543ea223eefe9d97e1f}} uses temporal information for constructing positive-negative pairs for classification of remote sensing data. Unlike this work, they ignore complex spatio-temporal dynamics and auto-regressive modeling, which are more crucial for forecasting.Recently, parallel work {{cite:0b6c04daf9fd15fd49b6338bcb5aa7a20c0d8f3c}} utilized generative modeling on radar data for precipitation nowcasting using generative adversarial networks. Compared to this, we focus on a different application of solar nowcasting and are able to demonstrate utility of a simpler model for this application where directly predicting the next instant values is sufficient as they directly correlate with solar irradiance {{cite:7249d4d72e62e484248018b6064cd7f93c92c412}} – hence do not necessitate requiring a complex discriminator for learning as in GAN models {{cite:b4ee085d9b3586f235f3b9a2689c1e20e6cdaca1}}.Finally, our work is also related to auto-regressive language models, that learn to predict the next word given previous words, which have been very successful for natural language processing {{cite:5be607ebc54153fed2eff0d3106bce8dd231c87c}}." ]

[]

[ "Vision Transformers. ViT {{cite:b1d340542b29e6741df51e5a534be0bf98f7d8dc}} is a pioneering work that uses only a Transformer to solve various vision tasks. Compared to traditional CNN structures, ViT allows all the positions in an image to interact through transformer blocks, whereas CNNs operate on a fixed-sized window with restricted spatial interactions, which can have trouble capturing relations at the pixel level in both spatial and time domains {{cite:c07c7c0de096282cd118b4d0ca804840495a8496}}.Since then, many variants have been proposed {{cite:bd70ed7a624a628db93fb93807c50a39728893ae}}, {{cite:5d29fa9007daef6492c94c8d52785e7b0896c0ac}}, {{cite:5de7a37e5b2b6172b809b05eeb6efc3e53beb72b}}, {{cite:99a1bb2da36fbf35ea0311f4aea314ddf2742a6a}}, {{cite:c756f46dc501586d4e5b21c0608e35b7dc7c7e9a}}, {{cite:dd155cb961dbbc546f210b5733a4db26e9b13a11}}, {{cite:7791a32654b202dc0395fa4705058c433b4b01d0}}, {{cite:1e9b66922b3df034d1644c4bd4c3a2276243aeeb}}, {{cite:2858340aa4221c8fc74dd17214a0227c643e2da3}}, {{cite:ec87c7b63e2826b29bf7cc747182075695a13ff8}}, {{cite:4bd17aef71ff6bc1bce15587aedc709077b83722}}, {{cite:f2906a42f2cd9ac3407db6d574a98081d3a95206}}. For example, DeiT {{cite:d05fc1fc7b486450e4bf407d9fdf6066f7c6f5a9}}, T2T-ViT {{cite:d7fa2f420d30633f11cd671eaf6da16897179374}} and Mixer {{cite:9f90d24fe99339dcd39f15a32862882308796c9c}} tackle the data-inefficiency problem in ViT by training only with ImageNet. PiT {{cite:39f91d8c5abaa24a6bda991fce660271731d24e8}} replaces the uniform structure of Transformer with depth-wise convolution pooling layer to reduce spacial dimension and increase channel dimension. LV-ViT {{cite:0688c050b93b62d3bf0dc0b4dc9ac10ea4cd7a1c}} introduces a token labeling approach to improve training. PS-ViT {{cite:249474b456eb660e2ef2ecf55b5cbc2716e4389f}} abandons the fixed length tokens with progressive sampled tokens.", "Efficient ViT.The huge memory usage and computation cost of the self-attention mechanism serve as the roadblock to the efficient deployment of the ViT model on edge devices. Many works aim at accelerating the inference speed of ViT {{cite:ab724de48012a1743234a4819beb813364269014}}.For instance, S{{formula:0b3c6e5c-cd42-4dc8-a330-3e2c18632bfd}} ViTE {{cite:3801b6e56489a3316f22ddae1c4e459bd5ea6a09}} prunes token and attention head in a structured way via sparse training.VTP {{cite:fa3daa8896909a3ff2c0024f1c90932d74aa664d}} reduces the input feature dimension by learning their associated importance scores with L1 regularization.IA-RED{{formula:7ee4e910-5cf3-4d55-b7dd-cb8327230ff8}}  {{cite:fa24c5c437e9101ae58ddafc0478bd23f2c3440e}} drops redundant tokens with a multi-head interpreter.PS-ViT (T2T) {{cite:e0c855e8e1fc067af0f0d4255ddd5131cb140058}} discards useless patches in a top-down paradigm.DynamicViT {{cite:99311a2f7ff888b73688729bcb09b45e5b14d4d9}} removes redundant tokens by estimating their importance score with a MLP {{cite:3fb9d2fec80d6f9ee08ed3345044eeb0641b2865}} based prediction module.Evo-ViT {{cite:fdf549a01d3d041318abd2b3f7ce05358ee60786}} develops a slow-fast token evolution method to preserve more image information during pruning.TokenLearner {{cite:d62405835c822da1101cfd20c4860475594b4baa}} uses spatial attention to generate a small set of token vectors adaptive to the input.However, to the best of our knowledge, our idea of considering actual edge device deployment and acceleration has not been investigated by any existing ViT pruning approaches.{{table:f7ff3bfe-723b-46f1-a934-60243228a5c5}}" ]

[]

[ "Several systems have been established to standardise the reporting process and structure of common vulnerabilities (e.g., CWE , CVE and OWASP ). However, identifying the root causes of the reported vulnerabilities is still a time-consuming and expertise-required process . Recent work employed information retrieval, data mining, natural language processing, machine learning and deep learning techniques to characterize vulnerabilities reported in CVE and CWE systems , , . However, privacy vulnerabilities were not addressed in those previous work.", "Substantial research have been done for detecting software vulnerabilities in software systems , , , , , . Early vulnerability discovery approaches evolved from static analysis, fuzzing, penetration testing and Vulnerability Discovery Models (VDMs) , . Later approaches applied more advanced techniques such as machine learning, deep learning and neural networks to improve the accuracy of vulnerability detection , , , , . These approaches heavily focus on detecting security vulnerabilities in software systems, while detection of privacy vulnerabilities were overlooked.", "Recent work have also studied and used the CWE and CVE systems. For example, the work in {{cite:8a16339351fb6ba6d53693edd34abb67ce76becc}} collected CVE records with their associated CWEs and code commits. The collected information was then analysed to produce insightful metadata such as concerned programming language and code-related metrics. This work can be applied in multiple applications related to software maintenance such as automated vulnerability detection and classification, vulnerability fixing patches analysis and program repair. proposed a formulation to calculate the most dangerous software errors in CWE. They used this formulation to identify the top 20 most significant CWE records in 2019. Again, these prior work only focus on security vulnerabilities.", "Limited work has been done in identifying privacy vulnerabilities in software systems {{cite:732450bf5a9ec1e1578918f192d707eb4a22eb79}}, , ). {{cite:732450bf5a9ec1e1578918f192d707eb4a22eb79}} proposed a taxonomy of privacy goals based on Internet privacy policies. The study employed a content analysis through goal mining process to extract privacy goals from 25 privacy policies in e-commerce industries. The process consists of three steps: goals identification, classification and refinement. These identified privacy goals were classified into privacy protection and privacy vulnerability goals. The privacy vulnerability goal addresses a set of information processing that may violate consumer privacy (e.g., information monitoring, collection and transfer).", " introduced a framework to detect privacy leakage in mobile applications. This study identified several common privacy vulnerabilities in Android such as unintended sensitive data transmission and local logging. discussed a privacy vulnerability in mobile sensing networks which collect mobility traces of people and vehicles (e.g., traffic monitoring). Although these networks receive anonymous data, it was proven in the study that these data can identify victims. These studies have confirmed the occurrence of privacy vulnerabilities in multiple types of software systems (e.g., web/mobile applications and sensing networks). However, most of the existing studies only focused on security concerns when investigating software vulnerabilities, thus overlooked privacy-related concerns in many contexts.", "A large number of existing studies have investigated methods to preserve and protect privacy in software development. Several studies have proposed approaches to derive privacy requirements from organisational goals , data protection and privacy regulations , , {{cite:abaf47e3afebd1ef3eeb0b43e4a219534d5a6fe6}}, or privacy policies , to ensure that software systems comply with those restrictions and/or constraints. and presented frameworks for designing privacy-preserving architecture in software development. Recent work have examined how organisations or service providers have compiled with individual rights in data protection regulations. For example, studied on how iOS and Android app vendors respond when users request to access their personal data. This request is a mechanism to execute the right of access, one of the individual rights covered in GDPR and other privacy regulations. The study found that only half of the app vendors responded to the user access requests. This violates the right of access as the users must be able to request and get responses to access their personal data.", "A number of studies have proposed taxonomies of privacy threats (e.g., , , {{cite:5ece28e400026498ce1cdd63092917bc0919e180}}). developed a taxonomy of privacy threats for mhealth systems (mobile computing for personal health monitoring). It emphasises on identity, access and disclosure threats of patients and their personal health information. The threats in this taxonomy can be caused by patients, internal parties and external parties. {{cite:5ece28e400026498ce1cdd63092917bc0919e180}} proposed a taxonomy of security and privacy threats for IoT. The taxonomy identifies the security and privacy threats posed by IoT objects (e.g., sensors and cameras). It adopted four groups of threats developed by to analyse privacy concerns of IoT objects. These two studies cover limited areas of existing software applications. One of the most well-known privacy threat taxonomies was proposed by . This taxonomy covers the harmful activities that can violate privacy of individuals. However, the taxonomy does not focus on privacy threats in software systems. Thus, we have adapted and extended this taxonomy to address privacy threats in software engineering." ]

[]

[ "To the best of our knowledge, we present the first work that tightly integrates specialized compute units into the last-level cache of a CPU to perform stencil computations. In this section, we succinctly compare prior proposals against Casper.", "A wide body of research has focused on studying and analyzing stencil computations {{cite:ce218d9c41372a4e14282e4a68f78750fb61f9ae}}, {{cite:da75f856dc22ce231c8c17d30d327b52b0d8d4fc}}, {{cite:6712c83336c004184b55d4acb4f8bed3a5f9b127}}, {{cite:8de03c08e50b8b6b25012a964804879897109988}}, {{cite:4cecdb08dcddcb5e47a483e6435d585b912222f1}}, {{cite:fe359ee920bd41289b3c412aff7b97374d966a9d}}, {{cite:3f7d576f3caa351d3b7aa1d083097ab51d1f0162}}.PIMS {{cite:ce218d9c41372a4e14282e4a68f78750fb61f9ae}} is the most closely-related work to Casper. PIMS exploits high-bandwidth provided by 3D-stacking (e.g., HMC, HBM) to accelerate stencils. Casper, being a near-LLC accelerator, can be integrated with any commodity processor without the need of costly interfacing using through-silicon vias.Szustak et al. {{cite:83d8afd4e56a0a9441722a1a89d4778f6264d2b8}} accelerate the MPDATA stencil kernel on multi-core CPU while Thaler et al. {{cite:599d2ef4bab79bc48817e4dfb9f157778fbdcf31}} port weather stencil kernles to a many-core system. GPUs {{cite:98de80ea30051819ce941def9e357695b90bba8c}}, {{cite:fe359ee920bd41289b3c412aff7b97374d966a9d}} have been shown to increase performance due to the high degree of parallelism present in the computation. Wahib et al. {{cite:c8cb66a649eb50aab25a949ddc30a30ab693e193}} develop an analytical performance model for choosing an optimal GPU-based execution strategy for various scientific stencil kernles. Gysi et al. {{cite:da75f856dc22ce231c8c17d30d327b52b0d8d4fc}} provide guidelines for optimizing stencil kernels for CPU–GPU systems.", "More recently, the use of FPGAs to accelerate stencils has been proposed {{cite:8b3aad89f12950ab85337054f2f06ccd9dee44a6}}, {{cite:381cfa1393697a035ac1ba39467461f09e5f7a14}}, {{cite:e63856fe4be8b7c052e4ff14c98fe1f8e8505f1a}}, {{cite:58fbc92bc4e1fe20b215252f900759d42b2c83cb}}. Augmenting general-purpose cores with specialized FPGA accelerators is a promising approach to enhance overall system performance. However, data still needs to be moved to these off-chip external accelerators. Moreover, taking full advantage of FPGAs for accelerating a workload is not a trivial task as it requires sufficient FPGA programming skills to map the workload and optimize the design for the FPGA microarchitecture. In contrast, processing close to memory tightly integrates compute units close to the memory. This integration avoids unnecessary data-movement." ]

[]

[ "Y. Polyanskiy first introduced a framework named URA {{cite:8ef4eaa1c0fb512f569a55b574d1ccb70089d8c3}}. Specifically, in URA, all the users share a common codebook, and the decoder only needs to decode a list of messages transmitted from the active users. The error probability is defined as the average fraction of mis-decoded messages over the number of active users, including both missed detection and false alarm. It is obvious that the message recovery at the BS can be formulated as a compressed sensing (CS) problem due to the sporadic activity in mMTC, which is similar to the conventional grant-free random access schemes {{cite:01a17a83a79ea0ba60726cbc9e4f0106e59ff735}}, {{cite:7cda029df4a59e25dd192562984f5f97b56f33d6}}. However, the size of the common codebook grows exponentially with the number of information bits. In practice, even if a short packet is transmitted, the size of information messages is typically at the order of 100 bits, which makes the CS algorithms computationally intractable. In this context, V. K. Amalladinne {{formula:fd7ff55f-97d2-4d9b-a56b-1c4a8e20810a}} proposed a coded compressed sensing (CCS) scheme for URA communication {{cite:1865507e2a31b60aed776d16c81238ed553d4705}}. In particular, the messages from active users are first divided into several sub-blocks. Then, a systematic linear code adds redundancy to those sub-blocks. Once this is achieved, each sub-block is mapped into a codeword in a common codebook and transmitted in a certain sub-slot. Then a standard CS algorithm implements the detection of the sub-blocks. Finally, the sub-blocks transmitted in different sub-slots are stitched together to obtain the original messages. Build upon the findings in {{cite:1865507e2a31b60aed776d16c81238ed553d4705}} and the structure of sparse regression codes (SPARCs), A. Fengler {{formula:0adf5243-420e-4ffd-a91f-36111f30e7a3}} provided an improved inner decoder, and a complete asymptotic error analysis {{cite:7d09584897fddf63385aaaf0416f17448c2153e0}}.", "Apart from the above works, the study of massive multiple-input multiple-output (MIMO) URA has also attracted much attention. A. Fengler {{formula:ab155c0b-1501-4157-a7d2-d97cee8d6f8f}} extended the URA model of {{cite:1865507e2a31b60aed776d16c81238ed553d4705}} to a block-fading MIMO channel by using a low-complexity covariance-based CS (CB-CS) recovery algorithm {{cite:a50b77cb6ba99bec2fc6a387c53c148feb7912bc}}. Considering the low code rate and spectral efficiency of the CCS scheme, V. Shyianov {{formula:6064fcd0-49dd-4fb5-9831-a4fa3ece8278}} proposed a new algorithmic solution to solve the massive URA problem by leveraging the rich spatial dimensionality offered by large-scale antenna arrays {{cite:e243f21853036ba5be1e31bf741244a34d8f86ba}}. Besides, without requiring a separate activity detection or channel estimation step, A. Decurninge {{formula:79dc9ec7-cffa-4a52-880b-81ef7b263b33}} introduced a structure that allows the receiver to separate the users using a classical tensor decomposition {{cite:bcfc4b612ad0cdb2343e3b09cf3b4e4f354d47a3}}. As URA is a special scheme of grant-free random access, A. Fengler {{formula:34c6cf5a-7235-4257-a5dc-05d0c1489e76}} presented a conceptually simple algorithm based on pilot transmission, activity detection, channel estimation, Maximum Ratio Combining (MRC), and single-user decoding {{cite:06e0d9972a44ea3db441a01762ada3aadd23831f}}, which is similar to the existing grant-free random access schemes {{cite:c8ff6b00a90d33cc8d0990c44c1619ca080b71c4}}, {{cite:01a17a83a79ea0ba60726cbc9e4f0106e59ff735}}. The difference is that they use a pool of non-orthogonal pilots where every active user picks one of them pseudo-randomly. Furthermore, X. Shao {{formula:4462e33b-d26a-432e-b86b-59b8acebbc55}} proposed a unified cooperative activity detection framework for sourced and unsourced random access based on the covariance of the received signals for the sixth generation (6G) cell-free wireless networks {{cite:29d0a306cb8b22c87bbdc5d2bfe8deeae5306b5e}}." ]

[]

[ "Mutation testing has been established as one of the strongest test criteria {{cite:072c0641b0fb61935838a22d54619052a82b747b}}, {{cite:480051aebc776be92de13f101faef5419dc03331}}. Despite its potential, mutation is considered to be expensive since it introduces too many mutants. To this end, random mutant sampling {{cite:1a723c72f167626b211d6d850b3ad10eb797c18e}}, {{cite:28c0f523818257da1980a4e6966e7c50cbee6ea9}} and selective mutation {{cite:1dd7d86208ad3469e7ad5d7d7d9dabd2e8c0ef2f}} (restricting mutant instances according to their types) have been proposed as potential solutions. Unfortunately, these approaches fail to capture relevant program semantics and performing similarly to random mutant sampling {{cite:fd1415d5da23e7c86fb198da59c103f6367201dc}}, {{cite:0dad1008c1b4c87b9a4d2134fa0a9eea17d2c6d0}}, {{cite:8e5a5a3784ec176354ff362dcfd8f6581b29c7a6}}.", "Other attempts regard the selection of relevant program locations, which should be mutated. Sun et al. {{cite:a7e8dbc30b5b764c610bd3fdf159fb78670d358b}} proposed selecting mutants that reside in diverse static control flow graph paths. Gong et al. {{cite:2818d4e48cfae4dfb31598a0adaa4ebe19cf927b}} identified dominator nodes (using static control flow graph) to select mutants.", "More recent attempts regard the identification of interesting mutants (pairs of mutant types and related locations). Petrovic and Ivankovic {{cite:a851c8c9d9ab91841f7e4d55a42faeb19d7a2380}} and Just et al. {{cite:b68a779c15b6d2f1a62c3e02fae5c792953e458d}} proposed using the code AST in order to identify “useful” mutants. Petrovic and Ivankovic used what they called arid nodes (special AST nodes), while Just et al. used the AST parent and child nodes, in order to identify high utility mutants. Mirshokraie et al. {{cite:6c1c4600d531ceceaab6d0f8060e3a5f93e778cc}} employed complexity metrics together with test executions to select killable mutants. Similarly, Titcheu et al. {{cite:8e5a5a3784ec176354ff362dcfd8f6581b29c7a6}} employed static features, including data flow analysis, complexity and AST information, in order to perform mutant selection, wrt mutants linked with real faults.", "In our analysis we approximate the performance of the above approaches through the two baselines we adopt and show that our approach significantly outperforms these. Random mutant sampling is performing comparably to operator mutant selection {{cite:fd1415d5da23e7c86fb198da59c103f6367201dc}}, while the supervised baseline we consider simulates the AST-based and complexity-based approaches.", "Perhaps the closest work to ours, is from Marcozzi et al. {{cite:d16701b95420af269248ccd71eec51174a2e0b51}}, which attempts to identify subsumed mutants using verification techniques (such as weakest precondition). While Marcozzi et al.'s approach is particularly powerful, it targets weak mutation and not strong as we do. This results in several false positives in the strong mutation case due to failed error propagation {{cite:072c0641b0fb61935838a22d54619052a82b747b}}. Moreover, Marcozzi et al.'s approach is time consuming, requires complex computations and infrastructure while Cerebro is fast and simple. Nevertheless, future research should attempt to combine these methods.", "Tufano et al. {{cite:844724ff148cad40ffe0aeb7d97dbf7bcefaa031}} proposed using Neural Machine Translation to learn mutations from bug fixes with the aim of introducing mutations that are syntactically similar to real bugs. Cerebro relies on the same technology, though it targets a different problem; the identification of high utility mutants, among those given by regular mutation testing tools, while Tufano et al. aim at generating mutants regardless of their potential. This indicates that Cerebro can complement Tufano et al by selecting relevant mutants. Nevertheless, we focus on subsuming mutants, that could help measuring test adequacy and designing test suites, which are unlikely to be supported by Tufano et al. as there is no notion of subsumption in the bug-fixing sets they use. Moreover, we make no assumption on the availability and repetitiveness of historical bugs and their fixes.", "Predictive mutation testing (PMT) {{cite:f8208bdcbd40536e0c6261983155ffc600c9818b}} attempts to predict whether a given test can kill a given mutant without performing any mutant execution. The approach relies on a set of both static and dynamic features (relying on coverage and code attributes) and achieves relatively good results (on average with 10% error). Though, PMT mainly targets intra-project predictions, while Cerebro targets inter-project. Nevertheless, PMT is incomparable to Cerebro since it aims at evaluating test execution results, while we do mutant selection prior to any test execution. In other words, we aim at identifying the mutants to be used for test design/generation, while PMT to verify whether mutants are killed by some tests. Therefore, the two methods target different but complementary problems.", "Evolutionary Mutation Testing (EMT) {{cite:ae7fec56b105cb56518401020a464970acbee4a5}} utilises dynamic features (execution traces) in order to identify interesting locations and mutant types. As such, EMT requires tests and user feedback, which make it different but complementary to ours; Cerebro can set a starting point for EMT or integrate its predictions within EMT's fitness function. Higher-order mutation {{cite:688a425cad63f29be14da546b967ec3eaf01e0a9}} aims at dynamically optimizing mutants based on given test suites. This means that Cerebro can be directly applied to support test generation prior to any test generation, while higher-order mutation is only applicable after test generation. Perhaps more importantly, Cerebro does not introduce any expensive dynamic mutant execution, while higher-order mutation introduces major mutant execution overheads." ]

[]

[ "There are numerous accounts of basic domain theory in several constructivesystems, such as{{cite:30b991963215933c68c1767553400dc5fbd11f1b}}, {{cite:5beecabbd283c0f284035fc4e040ec46b3959932}}, {{cite:12f6cd2941d037588b620705dd95d27c5186a548}}, {{cite:2d60b9002523194819bb4390e1cee6fe86709d57}}, {{cite:7123248f0ee0aded70efbac764d0a6d2d688bfaf}}, {{cite:03fc744a220337c1f966e87e22a16c6fd6276f03}}in the predicative setting of formal topology {{cite:9e95849d06f4451ec1a457fe438dec6d1d86408e}}, {{cite:08492c9f93bd509cfd76e675e044849ad4a19a55}},as well as works in various type theories: {{cite:2f0cda7d3c15e4a6e3eb96a430cf0a272a9337e5}} in (a version of)Martin-Löf Type Theory, {{cite:abceabb771735eae036bc932a5b379085ea7b481}} in Agda,{{cite:7bf60fdac1e3969fc3d685fd05f2371348097fbf}}, {{cite:851a926514f4305d1b54e1339eb9dee762fdebe8}} in Coq and our previouswork {{cite:7bc51f54f75f20fe19af285016eea5ea7b30169b}} in univalent foundations. Besides that, thepapers {{cite:fac082847ef1c77f96f080f2885ddf9c7360e0ae}}, {{cite:2e2f7aea3ef5d81bcbd7fa19c91dc379092214b6}} are specifically aimedat program extraction.", "Our work is not situated in formal topology and we work informally in(impredicative) set theory without using excluded middle or choice axioms. We alsoconsider completeness with respect to all directed subsets and not just{{formula:76b2c050-0128-4948-8668-b8e9e59e48b1}} -chains as is donein {{cite:fac082847ef1c77f96f080f2885ddf9c7360e0ae}}, {{cite:2e2f7aea3ef5d81bcbd7fa19c91dc379092214b6}}. The principal contributionsof our work are the aforementioned notions of intrinsic apartness and sharpelements, although the idea of sharpness also appears in formal topology: anelement of a continuous dcpo is sharp if and only if its filter of Scott openneighbourhoods is located in the sense of Spitters {{cite:959339f596c068ce9f2053a4871007e028f984f4}} andKawai {{cite:7123248f0ee0aded70efbac764d0a6d2d688bfaf}}.", "If, as advocated in {{cite:bd4b552c570cfff84e8fb89fa3c922a9bd76742d}}, {{cite:625188e82f998a3744252e6774a52f897d176929}}, {{cite:c11f7580ddb8542962e00a022ec174a96929662a}}, we think of(Scott) opens as observable properties, then this suggests that we label twopoints as apart if we have made conflicting observations about them, i.e. ifthere are disjoint opens separating the points. Indeed, (an equivalentformulation of) this notion is used in Smyth's {{cite:b0a3572a1799bdb19aa6c5c80578b8d2b89ffcb6}}.While these notions are certainly useful, both in the presence and absence ofexcluded middle, our apartness serves a different purpose: It is a positiveformulation of the negation of equality used when reasoning about the Scotttopology on a dcpo, which (classically) is only a {{formula:cf546266-fb69-4f72-9005-cea3b970b7d1}} -space that isn'tHausdorff in general. By contrast, an apartness based on disjoint opens wouldsupposedly perform a similar job for a Hausdorff space, such as a dcpo with theLawson topology.", "Finally, von Plato {{cite:32ed267898eae5fdba9e43d5f54fc4f13105e852}} gives a constructive account ofso-called positive partial orders: sets with a binary relation {{formula:43b9f2c1-4e40-4390-b26f-1ac0962017ff}} that is irreflexive and cotransitive (i.e. if {{formula:edf5c0c2-81de-4042-aad0-0e1b1eb54de4}} , then{{formula:3d72671e-1871-4949-bea3-976f8d295866}} or {{formula:7cbe423a-19bb-40e6-8ae9-9ab812828fa3}} for any elements {{formula:b96ebae6-91c1-44ae-b83a-4f3405f04953}} , {{formula:ef2002e9-b326-4db3-8ac4-15e7127eda94}} and {{formula:41ce8acb-dd7f-4fe5-9c2a-633decb62bbe}} ).Our notion {{formula:701f1e01-e50e-43c1-8f14-52d6ac1b4289}} from Definition REF bearssome similarity, but our work is fundamentally different for tworeasons. Firstly, {{formula:1efceb4d-7dd0-4645-a826-7960aa411c0f}} is not cotransitive. Indeed, we cannotexpect such a cotransitive relation on nontrivial dcpos, cf. Theorem REF . Secondly, in {{cite:32ed267898eae5fdba9e43d5f54fc4f13105e852}} equalityis a derived notion from {{formula:85bd7f77-7f75-481a-be87-bee4386c1889}} , while equality is primitive for us." ]

[]

[ "Most proofs of applicative similarity being a congruence use operationalarguments {{cite:b758dc7ca68444ee29d0b696817dd316b93877cc}}, {{cite:734b0440d6bdacb962d0bc3504aa2285b928a0bd}}, in particular Howe'smethod {{cite:3691bce56d82cb7f7dd9d61ac97db0f1da11e548}}. More recently, an abstract version of Howe's methodhas been developed {{cite:18d170810618278b26c3e1f451dae8acc5b9dac4}} to handle languages with algebraic effects ina uniform way. This method uses domain theory to handle recursion. It would be interesting tosee if the method described here generalises to a similar uniform method for computationaleffects, but this requires first developing a theory of algebraic effects in guarded type theory.", "Step-indexing and guarded recursion based operational techniques have previously been used for languageswith non-determinism. For example, Schwinghammer et al. {{cite:92c75f3eee87feb4ad7bfc786ea1bba7b311d15e}}construct an operational model for reasoning about a typed programming language with recursive types,polymorphism and non-determinism and use it to prove contextual equivalences of programs. Bizjaket al. {{cite:5caac59093c81e03fcdaf8b0a9b02c38f47b35e3}} show how to construct a similar model using guardedrecursion and topos logic. These works use complex operational techniques including {{formula:2191827b-2ea1-41d7-bdc0-58e62989307d}} -closure.Our goal is different, namely to develop a theory of denotational semantics in a type theory with guardedrecursion.", "The above mentioned works onnon-determinism {{cite:b758dc7ca68444ee29d0b696817dd316b93877cc}}, {{cite:734b0440d6bdacb962d0bc3504aa2285b928a0bd}}, {{cite:92c75f3eee87feb4ad7bfc786ea1bba7b311d15e}}, {{cite:5caac59093c81e03fcdaf8b0a9b02c38f47b35e3}}study countable non-determinism, rather than finite non-determinism.This is generally considered a harder problem. For example, this forces the step-indexing used bySchwinghammer et al. {{cite:92c75f3eee87feb4ad7bfc786ea1bba7b311d15e}} tobe transfinite, whereas the underlying model of the Clocked Cubical Type Theory is based on natural numberstep-indexing. Likewise, defining powerdomains in domain theory for countable non-determinism is much harderthan the finite case {{cite:1097015ca05b52d20383c2671907c49b2861bf40}}, {{cite:797afac4e8d2f0daafa0a7bda1ca116b376f0d8e}}. We discuss the possibility of extending ourapproach to countable non-determinism in Section .", "As described in Section , our partiality monad {{formula:15555363-c94c-43e9-83bc-1ec6d9bb072a}} is strongly related to thecoinductive partiality monad, and our use of it is similar to previous uses in semantics ofrecursion {{cite:04268ad59715c27b59c0fe2cbee93bbd751cc0e2}}, {{cite:a3f2c206cd5f73e3e1ed4a52cd3602cada91fd4b}}, {{cite:5e9cd9de13e586ca588de957d101939cc22474fd}}. Interaction trees {{cite:4c18b09b259e6128b2f08bb84ade4fc00bc2f812}}are a general data structure combining the coinductive partiality monad with computational effects. Our guardedpowerdomain monads can perhaps be seen as a form of guarded interaction trees for non-determinism, exceptthat the use of HITs allows us to consider these up to an equational theory." ]

[]

[ "Long-tailed classification.Early works focus on re-balancing the contribution of each class in the training phase, which can be divided into two strategies: re-sampling the data to balance the class frequency {{cite:b11647a6fc65af60dc685e0522c5561dec9b7cf2}}, {{cite:41c4a449d049b4bbb66a4617991fe4cb139972b0}}, {{cite:b41cf09ca8aa7d131ae678d325ac7b50bd211d48}}, {{cite:0059f96f3377f2126f5e459cd33fe679929349cb}}, {{cite:f906b30681463b991f9add1cc04a63d30d36e68c}}, and re-weighting the loss of classes {{cite:7032ad37a2b34d0d1f2b69c45421f2da27c49cf8}}, {{cite:7520b4c26e7174ab4ccccf685f0d2fb6cf433d66}}, {{cite:8f1c522bd98f784dcf1df032725b4fc8607d98c8}}, {{cite:a9b52277eab53f68949f91ca65d092b921d759b0}} to increase the importance of tail classes. Nevertheless, both strategies suffer from under-fitting/over-fitting problem to head/tail classes. Recent studies shows a trend of decoupling long-tailed classification into two separate stages: representation learning and classifier learning. Decouple {{cite:2749b6b46e8aeb475afe71252ab9486e0c350605}} learns high-quality representation with natural (instance-balanced) sampling, and achieve strong classification performance by adjusting classifier with balanced learning. {{cite:2b9d11f2f4bad85aed731dbaeaa1bc68ecf8b24f}} arrives at a similar conclusion by proposing a bilateral-branch network, where one branch is trained with natural sampling and the other uses balanced sampling. The decoupled learning idea is also adopted in {{cite:23a80ae728150084bf4277a572ece668a97b6096}}, {{cite:5035c230f6285b891e08398989391e55504f9b9a}}, {{cite:51b132179ff58a22386aee8db0f5056aca9c46b9}}, {{cite:99973ba5626d4e7464af8152e4a1985aab5386d0}}, where different classifier adjustment methods are proposed. There is a line of post-hoc logits adjustment by subtracting the training prior from the predicted logits {{cite:23a80ae728150084bf4277a572ece668a97b6096}}, {{cite:5035c230f6285b891e08398989391e55504f9b9a}}. {{cite:51b132179ff58a22386aee8db0f5056aca9c46b9}} removes the “head” causal effect to get the adjusted logits. DisAlign {{cite:99973ba5626d4e7464af8152e4a1985aab5386d0}} modifies the original logits by adding an extra learnable layer to the output layer.Apart from decoupled learning, {{cite:bb8a2d608c61039cfcfacbbca938109096552ab3}} addresses the long-tailed problem by transferring head distributions to tail; distilling balanced student models from imbalanced trained teacher models {{cite:1e84565c66cf2d3877fbaaf1624fe998f7ca3e35}}, {{cite:d9606efd6afda2566cbf04b9a11e7d534f75a001}}. Although all the above methods manifest an overall accuracy improvement, as we discussed in Section , they are indeed biased.", "Causal Inference. Causal inference {{cite:9499e7714107e61cee9cab6ec3201c87e7119c08}}, {{cite:c79a9cfa0f87c25f863b3d45130f0efca83c5e49}} has been studied in a wide spectrum, such as economics, politics and epidemiology {{cite:ac3da5276f8d7ce457e5ffa97e567e6f9a780a0f}}, {{cite:90c424e8c2318e800606a8976bafc34ba6ad3b80}}, {{cite:8e6c8a4e89f0ea938d703c5315a4518eabf25d23}}.Recently, causal inference has also shown promising results in various computer vision tasks, e.g., pursuing pure direct effect in VQA {{cite:3bfada53bc1d72d816378c72660bcd9b35aa684a}}, back-door and front-door adjustment for captioning {{cite:abf35558868efcb04ae48e776f2af5bd16631125}} and counterfactual inference for zero-shot recognition {{cite:dfb8ef2207f065a132f4037f599d5ce08626a5a1}}. Our algorithm can be viewed as a practical implementation and complement of the recent empirical risk minimization of learning causal effect {{cite:4d8974f099eb039a59798a657f3b1af4b535548c}} for confounding effect removal.", "Bias mitigation. A bias mitigation work related to us is LfF {{cite:2c3c9bf3727b556ce949e8abd09484ebce3a94a0}}. LfF used a reweighting-based method to deal with bias in out-of-distribution robustness, and used the cross-entropy loss to calculate the weights. The main differences are as follows. First, LfF focused on biased dataset with spurious correlations, while we focused on long-tailed task with label prior shift.Second, LfF weights the losses at the sample-level, while xERM weights the losses at the supervision-level. Third, LfF calculates the weights using a biased model and a target debiased model, while xERM calculates the weights using fixed balanced and imbalanced models which can be merged into one." ]

[]

[ "Improving the screening process. Several approaches have been proposed to reduce the costs of the expensive screening step in SRs {{cite:5223798125f07f5f737d918bd9e41d4cd71e5de5}}. They are mainly divided into two categories based on its end goals. One category is to directly help the screening step to be more effective, either by decreasing the number of candidate documents to be screened or by reducing the number of SR experts needed. This direct goal is pursued by training a classification model. However, adopting a trained classifier is prone to fail to find all relevant documents, because it is hard to ensure the nearly perfect level of accuracy and also different SRs are meant to answer different clinical questions. That is, each SR has its own topic and it is hard to learn a general classification model for all SRs. Moreover, the small number of relevant documents in each SR makes training a model more challenging. With this regards, several studies suggest active learning setting through relevance feedback {{cite:abf069a98a6cc6d2404330d10f680bc38e1ebf89}}, {{cite:938040eb1e7c0b966e4330445b1c490644d6c2d0}}.", "The other category is a relatively indirect way to improve the screening process. While the number of documents to be screened remains unchanged, it assists SR experts during the manual screening process. An advantage of this approach is not to imperil finding all relevant documents. Screening prioritization is an approach in this category. As aforementioned, it makes the screening process efficient by allowing abstract screening and full document screening to be carried out simultaneously. Another approach is to visualize groups of similar candidate documents and provide their topics. In this way, SR experts are able to screen similar documents together, instead of documents in a random order.We refer to {{cite:5223798125f07f5f737d918bd9e41d4cd71e5de5}} for more details.", "Ranking documents using a query document.Using an example document as a query is an important research problem in various domains, such as patent search and literature search {{cite:61e26e23280fa144279613472adb1121ce62e0d4}}, {{cite:5d8acf44e49141c843a44abcf334a4e538192afa}}, {{cite:21a0ade53a4f9d716fb0867ed8a02e45431dd30c}}, {{cite:97963dd03529e16ab41c7ccfc0ff4b9cf9b7c59e}}, {{cite:394e3b8dbe9143fde7f7d4c26503100d77a0e282}}, {{cite:275c87fb629567c1a47be2a7957ea49e43d41643}}. Most current solutions first transform a long query to a short query, and then apply a retrieval function. As most retrieval functions are optimized for short queries, verbose queries often become a challenge in computation and query understanding {{cite:8b9d17c21be4e15e20462d1ac76910081f3aebc4}}.", "In screening prioritization, formulating a short and meaningful query from a known relevant document is non-trivial because of the detailed relevance conditions. There are a few possible approaches: keyword extraction and PICO element extraction. Keyword extraction techniques can be adopted, but they are prone to lose detailed yet important relevance conditions such as age or gender of patients in P-element. Extracting medical concepts can be an alternative {{cite:275c87fb629567c1a47be2a7957ea49e43d41643}}, but this approach has to rely on the accuracy of toolkits for medical concept extraction. PICO element extraction can be employed to extract relevance conditions appeared in a query document {{cite:3311969f56ad74fd9d3881fc3931b20830420f52}}, {{cite:9fc26258e44395fe2fb059e5266fba5118e4ee8e}}, {{cite:5d72df2fc74595edb0c15d34429aee7c62c0a16d}}. Extracting PICO elements without loss of relevance information therefore becomes a challenge.", "Document matching approaches have been less explored in ranking by a query document. Computing pairwise document matching scores in large-scale corpora online is costly. Several studies propose ways to improve the efficiency in pairwise similarity search problems {{cite:3b9aa444565ce27d4c480907ead0158a8e170dd4}}, {{cite:293a275831dfe35ae02b40ddb07fe70b388f205d}}, {{cite:a290ad56cce46daff2f0d20ffab4492307e084eb}}.However, in screening prioritization, candidate documents are typically in the scale of several thousands documents. In the CLEF TAR 2019 dataset, the average number of candidate documents is about 6,000. The computation on a few thousand documents is small compared to the costs of manual screening which takes place on prioritized documents for weeks. The effectiveness of prioritizing documents is more crucial than the efficiency in improving the manual screening process." ]

[]

[ "Extensive work has been done on detecting word usage change across corpora that predated the alignment-based methods {{cite:c07d0b257fbd63ae399b3322fb063324fa6bb2b7}}, {{cite:65cf9e681bbdb140a68933428ae293674a358b74}}, {{cite:683c66fd36f0445ca397537461225d1a72363767}}, {{cite:fa3ab3a59a9f59ba3e1912d3802daa378b3c32ad}}, {{cite:fddc330f9f773ea270ad505855b0f8ab412796b4}}.", "In addition, two works are more closely related to our approach.In ADB17, the authors also use the neighbors of a word in order to determine its stability (and therefore, the extent to which it changes). Their best model combines the traditional alignment-based approach with weighting the neighbors according to their rank and their stability. The algorithm is iterative, and they update the stability of all the words in the vocabulary in each update step. Our method uses the neighbors of the words directly, does not include an iterative process, and does not rely on cosine-distance in the aligned embeddings. In addition, their method requires computation for the whole vocabulary, while other methods, including ours, usually allow querying for a single word.", "Another work that considers the neighbors of the word in order to determine the extent of change is that of HLJ16b, in which they suggest a measure that is based on the changes of similarities between the target word and its neighbors in both spaces. They find that this method is more suitable for identifying changes that are due to cultural factors, rather than linguistic shift. This may serve as another motivation to move from the global measures to a local one.", "Two other relevant works have come to our attention after publishing this work, both using some variation of our method as a part of a different analysis, where identifying word usage change is the mean and not the end goal. Both have a very limited experimental setting compared to ours {{cite:bb642ede31930889808a2fcddcacac33b9f58c31}}, {{cite:191d933cbfb89519a94c38d0015af5953c64584c}}.", "In a concurrent work, RR19 model theevolution of language in relation to world events. Their work deals with creating timelines of words based on semantic change and suggests, among other methods, to use the intersection of neighbours to detect relevant time points.", "Recent works {{cite:d0aa79836a30e28cbfaf3ce6b0de3b1007dac2a1}}, {{cite:20853d12de3ae4fc64d50d5d8ab034c8113bdbd5}}explored the possibility of modeling diachronic and usage change usingcontextualized embeddings extracted from now ubiquitous Bertrepresentations {{cite:b44669e4301f157b51a4cfe6b3e5ea66867e2538}}. Focusing on the financialdomain, montariol:2020:bertDiacChange use, on top of Bertembeddings, a clustering method that does not need to predefine thenumber of clusters and which leads to interesting results on thatdomain. Another approach from hu-etal-2019-diachronic relieson the inclusion of example-based word sense inventories over timefrom the Oxford dictionary to a Bert model. Doing so provides anefficient fine-grained word sense representation and enables aseemingly accurate way to monitor word sense change over time. Mostof those approaches could be easily used with our method, theinclusion of contextualized embeddings would be for examplestraightforward, we leave it for future work." ]

[]

[ "Our work complements existing research, but manages to occupy a distinct niche in a broader landscape.Some previous work labeled as preference revision {{cite:2aec3670f161e6f4bcb878e23fe9643f0fc4bc3e}}, {{cite:dc72b681aca373639b73e37fcca1778cfeef04a7}}, {{cite:97551829d4134f2780b4d975a4ad0463080da684}},looks at changes in preferences prompted by a change in beliefs.Here we abstract away from the source of the new information, choosing to focusexclusively on a mechanism that can be used for resolving conflicts:the rational thing to do when knowing that, for some reason or other, one's preference has to change.Other work {{cite:339544ab7a9a28d2f3d42f93af55793a6d367ce5}} describes preference change whenpreferences are represented using CP-nets {{cite:6a5d759755aa4238206fcff904e14c8bdcfb5f54}},or dynamic epistemic logic {{cite:f59dcd05f1bd11c61e004f181ff9487307491fec}},in the context of declarative debugging {{cite:96f4bc6387ca007dc84dca3cf90682dc3673654e}},or databases {{cite:3d310947a3a5d11da1a4ae09cebf5d6c3cb0b1da}},and therefore comeswith additional structural constraints.In contrast, we have opted to represent preferences asstrict partial orders over a set of items: we believe this straightforwardformulation allows the basic issue signaled by Amartya Sen {{cite:657cae4d273f82777fb0f143e0cee1b96268912a}},to be visible and tackled head on.", "Apart from the issues raised in the Economics literature about second-order desires{{cite:d60b8332fa8a1852a3498e3de2ab451415f4f9cd}}, {{cite:7d41909599594419950944fa51adb6e077e19b6a}}, {{cite:657cae4d273f82777fb0f143e0cee1b96268912a}}, {{cite:2d6c85c8dcffec780595a01f36f8620b2dcfaa7e}}, {{cite:74ddf933f3e4d1aea07a627818151006d8b7aa5b}},the basic phenomenon of preference change has also been raisedin explicit connection to belief change {{cite:b34ff5292fdba53f1bb68f77cfb3fd60089b503d}}, {{cite:bec455811b615e2aadfc69fca58fb03014b193ce}}, {{cite:659a326ea119289d29a10a47f2a6c22442d7b5b9}},but a representation in terms of preferences on the comparisons present inthe preference orders, along the lines suggested here, has, to the best of our knowledge, not yet been given.Much existing work proceeds by putting forward some concrete preference revision mechanism,possibly by shifting some elements of the original preference around,and occasionally with a remark on the similarity between this operationand a belief revision operation {{cite:085c8dce1554f00715c8e00979c59498e721ad2b}}, {{cite:bc3a60aaa00e9639ef2e7f10544da0e58253c3e4}}, {{cite:97551829d4134f2780b4d975a4ad0463080da684}}, {{cite:2618f21d854795c29c8bfa2b6356d46262470d12}}.What our work adds to these models is an analysis in terms of postulatesand representation results.", "The postulates we put forward for preference revision bear a distinct resemblance to the AGM postulatesemployed for belief revision{{cite:0ed9850eba18f5a75e0bb4216589b9200da9652c}}, {{cite:9adbe46388b22c882451d24639bcc7562c123030}}, {{cite:a1f1abd46a604569a84b020e72e831e47eb04fbf}}:given that changing one's mind involves choosing some parts of a belief to keep and some to remove,this is no coincidence.Indeed, the two problems are similar, though the structural particularities of preferences(in particular, the requirement that they are transitive)mean that transfer of insights from belief revision to preference revision is by no means straightforward." ]

[]

[ "In this section, we review existing work point process and uncertainty quantification.", "Hawkes process: Point processes {{cite:77aeb3ab1a97f039eb687a572e05b319d3dfd495}} are useful to model the distribution of points and are defined using an underlying intensity function.A Hawkes process {{cite:2cc2736eb2d245d134a7144a88f71974355c7fc4}} is a point process {{cite:77aeb3ab1a97f039eb687a572e05b319d3dfd495}}, {{cite:16cfa359568a13db979c47078929b64e0fdf6710}} with self-triggering property i.e occurrence of previous events trigger occurrences of future events. Hawkes process has been used in earthquake modelling {{cite:f6ae11ed821582f73a4656cadb753d4af342338e}}, crime forecasting {{cite:2f132691fac0c3751065f7c4a0dafc691dfe599c}}, finance {{cite:f5271bdb6f34b35ace34f6449702ac1ba63e0fc9}}, {{cite:0bbddd86ced3c356055440ccf130b6e63754f6f1}} and epidemic forecasting {{cite:7ba07c6656be65c09334943900062b5962a44610}}, {{cite:ed6218050cecbb9073f730578aa384cd37db0e35}}.However, parametric models may not be capable of capturing event dynamics in some cases.To avoid this, various research works have proposed {{cite:6e8ff2e6646d488f796b660f78448b3f0c1c3064}}, {{cite:df59889e0232f73e1adeee06fd9b8f65f86b2e4c}}, {{cite:105b1633e70d91c8b43a2d558d0949396a815ee5}}, {{cite:04b1b2aea8691d95121ec67537b091c5e20e2647}}, {{cite:2519a037768b8118a0bc60e96d73bf96f3199b1b}} where intensity function is learnt using neural network. It has been proved that neural models are better at learning arbitrary and unknown distributions. {{cite:04b1b2aea8691d95121ec67537b091c5e20e2647}} uses the neural network to learn the cumulative hazard function,which is highly flexible and doesn't involve any approximation for the calculation of integral. There are few works {{cite:95d51e137a5300b2fb0ceaaabe35966ff86cc62e}}, {{cite:f5cd86c808e2c0d1f80dbd399740d7bcad9bfd85}}, {{cite:5f137722978804ac3b13ffc6ab0ce6a06bf06f12}} where authors have proposed bayesian parameteric hawkes process models where they find uncertainty over the parameters of hawkes process. However, our work proposes bayesian neural hawkes process to model the time of occurence of events along with the epistemic uncertainty.", "Bayesian Neural Networks: Bayesian Neural Networks (BNNs) {{cite:c31eb9819be1b6bc61f71ae3175d0d104be46f02}}, {{cite:ccf364f1de00c0c759b6fa1f3098f2ca5f16a02f}} are widely used framework to find uncertainty estimates for deep models. However, one has to resort to approximate inference techniques {{cite:b40c34b52a0e5182ba8f5b871e404e6a1d65ed5c}}, {{cite:c31eb9819be1b6bc61f71ae3175d0d104be46f02}} due to intractable posterior in BNNs. An alternative solution is toincorporate uncertainty directly into the loss function and directly optimize neural networks by backpropagation {{cite:dc6087db174b2db4cb1c900fd6197d2109189285}}, {{cite:ba43c02e5d8b9d01686e64fb205d3ccce23f4c52}}. Monte Carlo Dropout proposed by {{cite:cf129230b4cc742494393c0bc938db2c1e0c9bd1}}, {{cite:d348ef8ea2fb28bafd0e51955869c83d112d265a}} proved that dropout regularization {{cite:3859788a2c8194bff6d88c1d68dfc96c78cac994}} can act as approximation for the Bayesian inference over the weights of the neural network.", "There are some other efforts in the literature related to uncertainty estimation for event modeling. A related work for event modeling is {{cite:c2728ff86b86368421760f872830983464dc134e}} where authors have proposed an adversarial non-parametric model that accounts for calibration and uncertainty for time to event and uncertainty prediction. However, neural point processes are widely adopted model for event modeling due to its theoretical and practical effectiveness in various applications. Therefore, the goal of this paper is to augment the capabilities of neural point process to predict future events with the uncertainty estimation capability. To achieve this, we propose a novel approach which combines neural point process and Monte Carlo dropout for performing time prediction along with uncertainty estimation.{{figure:a7c550a0-0b06-4dc4-b7c1-48850bddf676}}" ]

[]

[ "Pioneering studiesthat automatically classified camera trap images relied on manual feature extraction and smaller datasets {{cite:057c6c9bfe6519b65d55df9f1f2433abf1449eb2}}, {{cite:87f5407b78e240f574b09ed1c53a5549e003824b}}, {{cite:c14db2dcd54754f15f4f36de84203fb921baaec4}}, {{cite:be38a31c6e3fae15f61bfec754d84119c4395f2b}}. Better and more scalable results were later achieved with deep CNNs and larger datasets {{cite:388c7d00e75bdb94f978b0f6b95c993dbbda3b45}}, {{cite:27f05e45470f630e4636981f93307a98e0c3c0d2}}, {{cite:1c3325a4c6a12556bda147e2c843233af43ccb1d}}, {{cite:46f14ce4b409b0f7b4a7b055df08992e10374bf6}}, {{cite:afc714a026e5e36d5b7fd36a30d15b6949155924}}.Generally, models trained by these scholars achieve accuracies well above 90%, but the models are biased towards majority classes, which severely affects their class-specific performance.Especially theperformance for rare species is poor.Scholars dealt with this challenge by removing the rare classes from the dataset {{cite:388c7d00e75bdb94f978b0f6b95c993dbbda3b45}}, {{cite:46f14ce4b409b0f7b4a7b055df08992e10374bf6}}, with confidence thresholding and letting experts review the uncertain classifications {{cite:46f14ce4b409b0f7b4a7b055df08992e10374bf6}}, with weighted losses, oversampling and emphasis sampling {{cite:27f05e45470f630e4636981f93307a98e0c3c0d2}} or by using a combination of additional image augmentations for rare classes and novel sampling techniques {{cite:afc714a026e5e36d5b7fd36a30d15b6949155924}}.Although {{cite:27f05e45470f630e4636981f93307a98e0c3c0d2}} managed to greatly increase the accuracy for a few rare classes using oversampling, none of the aforementioned techniques systematically improved accuracy for most of the rare species. It can thus be concluded that dealing with class-imbalance in the context of camera trap image classification is still an unsolved issue.", "Two categories of methods for mitigation of class imbalance in deep learning exist: data-level and algorithm-level techniques {{cite:d70cefc3478d10ec54b745ae13be518dc29534d6}}, {{cite:e8f7e90a74fafad6bf0e279c6156097873610443}}.The former refers to techniques that alter the class-distribution of the data, such as random minority oversampling (ROS) and random majority undersampling (RUS),which respectively randomly duplicate or randomly remove samples to obtain a more balanced dataset.More advanced techniquescan also be used to synthesize new samples {{cite:6b9115aaa2f737d9b8c0ea89282b412cbfec85f0}}, {{cite:f2a1e7471643861a8dd8e0aa77ca404a2a0b1335}}, {{cite:f7ccbfe79740d106fb40f31f0d7beb0159912326}}, {{cite:df08a1e64bdfb7cbda8444cf8458fc3ab368cc6b}}, {{cite:c8bec1d7ddf67035716880580f7aca308d89b4e8}}, {{cite:22246ed39c0babd7ced0979cb2c67d91e27f31df}},but these are computationally expensive, and they require a large number of images per class and images within a class that are sufficiently similar.Algorithm-level techniques are techniques that work on the model itself, such as loss functions or thresholding {{cite:f24f366b0b181ce2ef09824f22138318e5fa9630}}, {{cite:4f4806e359195912549296a96e3928666a7d33b9}}, {{cite:d70cefc3478d10ec54b745ae13be518dc29534d6}}, {{cite:e8f7e90a74fafad6bf0e279c6156097873610443}}, {{cite:ec6fbce3acf6a6edede35e4d3cd92105eea37b0a}}, {{cite:d70cefc3478d10ec54b745ae13be518dc29534d6}}.Two-phase training, a hybrid technique, was recently introducedand shown to obtain good resultsfor training a CNN classifier on a highly imbalanced dataset of images of plankton {{cite:5dbd0ebcd72eb9c3ad5378dd39285162443de56a}}, and it was later used by others for image segmentation and classification {{cite:32367df17a0c65e64a64bb5759510cf5a39249a0}}, {{cite:d70cefc3478d10ec54b745ae13be518dc29534d6}}.Because of these promising results and the broad applicability of 2-phase training, wetest 2-phase training for camera trap images." ]

[]

[ "Communication games have been used to study the emergence and evolution of language theoretically {{cite:3b991c862948b3296e3cd22d211d8a449a79bc82}}, experimentally {{cite:227bda9131b0b24b3cb215ef8d8efe49bf91ac77}}, {{cite:b330181218847893d36bc10ec4f6c09eebdcd3fc}}, and computationally {{cite:643df594e797e0085305a4318442f92ebe3243a6}}.Artificial intelligence (AI) research has also emphasized the importance of learning to communicate through interaction for developing agents that can coordinate with other, possibly human agents in a goal-directed and intelligent way e.g.,>mikolov2016.It has been shown that by playing communication games, artificial (robotic) agents can self-organize symbolic systems that are grounded in sensorimotor interactions with the world and other agents e.g.,>steels-1998, steels2001, steelsbelpaeme2005, bleys2009.For example, in a case study with color stimuli, simulated agents established color categories and labels by playing a (perceptual) discrimination game, paired with a color reference game {{cite:ef45cd12a2a97d15db146105838d1adbe4f84059}}. bleys2009 extended these findings to robotic agents, demonstrating that successful color naming systems emerge in spite of differences in the agents' perspective.These studies are mainly interested in how a categorical repertoire can become sufficiently shared among the members of a population to allow for successful communication.Our analyses, in contrast, assume that successful communication will emerge, and focus on how visual representations and language shape each other.", "Over the past years, research using communication games to study language emergence in DNN agents has been gaining popularity {{cite:83064c77c34f8f908fdaadb7e3b402fa4210a55f}}.Some of these models skip any form of perceptual processing by using symbolic input data e.g.,>bouchacourt2019, kharitonov2020, chaabouni2020.Even though other models implement a visual processing system and work with image data e.g.,>lazaridou2017, havrylov2017, they have rarely been used to explore the relation between language and visual perception.Notably, rodriguez2020 examined the effects of natural differences in object appearance (such as frequency, position, and luminosity) on emergent communication.Apart from that, bouchacourt2018 measured the alignment between agents' internal representations and conceptual input properties to determine whether emergent language captures such properties or relies on low-level pixel information.Still, these models usually extract object representations from fixed, pre-trained CNNs.As a result, they make claims about how the emergent language relates to the input, not the visual perception of that input.In our work, we exploit the flexibility of modern setups and introduce systematic variations in the agents' visual processing, such that we can establish a relationship between differences in visual processing and differences in emergent protocols." ]

[]

[ "While the sheer number of graph-structured data have been found particularly useful in many learning tasks in recent years (e.g., financial risks analysis for costumers social connection graphs), designing effective and efficient graph neural networks has become a popular topic in the graph representation learning field.", "{{cite:284a6188ab7595cb7bd33c4b8101502652b9d276}} first introduces the convolution operation into the graph domain and propose both spatial construction and spectral construction during the convolution process. To speed up the graph convolution computation, both {{cite:ea739e432fd7d9d1e299012c464af50474ef8042}} and {{cite:3a4c8ae4ad85b175a0ac7203efe0e700b6f60e22}} design localized filters based on Chebyshev expansion and verify their approach on relatively small datasets, where the training is able to happen in full batch. Apart from that, {{cite:e51ac2d442202252461146dc1779a958c9752b29}} adopts the convolution by using the spatial connections on the graph directly, where a wight matrix is learned for each node degree. Additionally, {{cite:271e8cb4d37ede29e10fc501241093ba5a521665}} develops the channel-wise convolution to learn a series of graph channels at eachlayer, which is able to shrink the graph hierarchically to encode the pooled structures. These GCN-based models have achieved state-of-the-art performance at that time and outperformed the traditional node embedding approaches such as node2vec {{cite:c92d15aa97b894150b580c0d5ee02dae0b2921c5}} and DeepWalk {{cite:963a77bab236d067d33ae8dd402b9a2e28b342b8}}.", "However, despite the success of GCN-based models, one major limitation is that these models are all based on knowledge of the entire graph Laplacian matrix amid training, making it unpractical to large-scale graph data with millions of nodes and edges. Following that, various sampling approaches are proposed to avoid operating on the entire Laplacian matrix {{cite:b644f38d5dbad4247257c54c47a90162fa42e811}}, {{cite:cc6b2c8b54235db76606627dceaf7a757c7a20ca}}, {{cite:2ef25f8d03ba85b51f3800e664523b9b2d82ab7f}}, {{cite:82300b91fab5f96a1202b91ee6d75f4c414e4dd3}}. Specifically, {{cite:b644f38d5dbad4247257c54c47a90162fa42e811}} first introduces the uniform layer-wise sampling algorithm, which applies a fixed size sampling window to each layer from inside out iteratively. {{cite:cc6b2c8b54235db76606627dceaf7a757c7a20ca}} takes advantage of random walk and considers sampling from different perspective. It considers nodes with highest visited count as important nodes and constructs the neighborhood based on this. Both {{cite:2ef25f8d03ba85b51f3800e664523b9b2d82ab7f}} and {{cite:82300b91fab5f96a1202b91ee6d75f4c414e4dd3}} focus on the variance reduction-based sampling. While the former proposes a condition-based sampling where the lower layer is sampled conditionally on the top one, the latter samples at the graph level, suggesting to construct different subgraphs as the minibatches while decoupling the sampling from the forward and backpropogation processes.", "Another line of work aims to improve the model capacity. For instance, {{cite:00a80343af764462b5e7b6d444367dc67797b4f6}} first integrates the attention mechanism into the graph representation learning. They introduce an attention-based architecture and the use self-attention to assign edge weights to the neighborhood. Following that, {{cite:187d97ec7e94c70ee21aa938aba687db404d18de}} adopts gated attention network to further enhance the control over attention heads. While our proposed method shares the same concept of quantifying the correlations to the neighbor nodes, it differentiates in coefficients calculations, the adaptive policy-driven sampling for neighbor nodes, and the forward message passing process." ]

[]

[ "In the research line of sharing models while protecting the corresponding training data privacy,previous work mainly use DP-SGD to train privacy-preserving models {{cite:51d705f43803ac9eb86e0a3fcf1d482ff5e06898}}, {{cite:76377f66ca7fc1953f678e636bc3c971feccd906}}.{{cite:e32fb1081b17d5b513b9d1522360e644b5ba0b8e}} propose a tight privacy accounting leading to a reasonable level of privacy cost for DP-SGD.Following that, variants of DP-SGD have been proposed to improve the model accuracy such as clipping based on quantiles {{cite:0c0951a66233bd5a69b43b3c53ad5a87403279d9}}, dynamic noise power and clipping value {{cite:c6edf73bef75bb1e8304d70e7ce543da55a1f02c}},and dynamic learning rate {{cite:3d4bc7758014e119cf1eac4dbc4dc6c6f6feb0b5}}, etc.DP-SGD and its variants, on the other hand, have had limited success in large deep learning models due to their high computation cost, huge memory overhead, and significant performance drops.", "In order to reduce the huge memory cost of each DP-SGD step,{{cite:326690838d3b46cdf44e33559af7c38ba8d3349c}}, {{cite:68996e0527dabc6bf6c8bfbdc43f5950b1a9b296}} study an effective use of JAX primitives inconjunction with the XLA compiler.Later {{cite:0655a28f77007eb358d3541e6d5e4ea98905b226}} mitigate the memory cost by a ghost clipping method, but still has 2 times computation throughput cost due to per-example grade clipping.{{cite:05a2bf4fff9ded2310b907c031f42c98363ce789}} reduce the memory cost via rank decomposition.", "To improve the model accuracy,anil2021large successfully implement mega batch size, e.g., a batch-size of 2M, for DP-BERT for model accuracy improvement, but still has a 10% accuracy drop compared to the non-private BERT model.{{cite:dc2407a7110c3537e69209dfb811420eb56e5390}} investigate private BERT fine-tuning, but reported results with {{formula:688f87d2-01a9-467a-91b0-8dec26c81e4d}} at least 100.Very recently, studies in {{cite:9fbcb8cb37d9f25940fc18da887a1df98ee5b33a}}, {{cite:05a2bf4fff9ded2310b907c031f42c98363ce789}}, {{cite:0655a28f77007eb358d3541e6d5e4ea98905b226}} show that the DP-SGD's model performance drop can be mitigated by the use of large pretrained models. Event though, there is still an obvious performance gap to the non-DP counterpart due to gradient bias and additive noise." ]

[]

[ "Profile guided optimization is a well-established technique for improving compile time optimization decisions {{cite:a4c583069aa268fdfa73a612b3f1316970599887}}. Profile information iscollected using instrumentation or sampling of the executable, and the collecteddata can only be used to optimize the executables on which it was trained. Compilerssuch as LLVM {{cite:66a9d6b4a5046684496929f2a8992b42c7fd4458}} and BOLT {{cite:7e87f15c8dceec9418e8880dc919f05c05d4b872}} use profile guided optimizationsto optimize and perform efficient layout of code.", "Despite the performance improvements offered by profile guidedoptimization, it is cumbersome to integrate into build systems and adds significant overhead. Significant engineering infrastructure is needed to provide deep integration with large evolving codebases {{cite:4f9220f5f1981f1a505856273a2508f618f7407c}}.", "We propose a novel technique to overcome the limitations of profile guidedoptimization using machine learning. The application of machine learning to compiler optimizations is well studied and shows promise in eliminating the human developed heuristics, as surveyed in {{cite:c5f35a0ec41d7afbdec0077925115a29b7577e12}} and {{cite:2ed88a3ca3f60b05ddca0a33442879ec82cc459e}}.", "The most commonly used technique is supervised learning. Supervised learning hasbeen applied to a range of problems such as loopunrolling {{cite:e379678184cd3cfca85a1c44ea95b8445b62ad34}}, {{cite:0ffb70aa64725049c5eb4c4b7618e87803232389}}, instructionscheduling {{cite:22f2837e65eca7821d9b33b7f7a9e505787d8f35}}, programpartitioning {{cite:2fae523d587fff0169aa4fe79844e2ec44746709}}, heterogeneous devicemapping {{cite:e7e1b899ecf254b2258666d7538300f8e94e3dba}}, {{cite:3ecf48fdeb529f323e1a09d430b9d34782df9467}}, function inlining {{cite:b95f4411b23d7e7664aa6d0686d9c5ce3e6f8ad1}},and various optimization heuristics in GCC {{cite:4ee48f027494b1a99903ee2b7bd9d07453f1c9fe}}.", "Supervised machine learning algorithms operate on labeled data, but it's noteasy to extract the labeled data from compilers. In some specific domains, suchas code generation for linear algebra primitives, there is a fixed compilationpipeline and a program that is reasonable to measure {{cite:7a2e37201a1104a08751ccc46272cc7b4fd4a447}}. In traditionalcompilers, the size of the optimization space and the complexity of theoptimization pipelines may make it prohibitively expensive to collect trainingdata, as it requires exhaustively compiling and measuring each program withevery combination of optimization decision to determine the best performance. Incontrast, our approach enables every measurement of a program to be used toproduce ground truth labels for branch probability.", "Another challenge is that measurement noise makes it difficult to evaluate theperformance of program. This is especially significant for very smallmeasurements such as the runtime of individual basic blocks {{cite:9d6bebbd2e9e446fcf7fbda7fba05022102135e9}}.In {{cite:4aa7057b910c4470f4dd097389e28502d1e42bd1}}, noise in measurements is used a prediction target.Some projects rely on proxy metrics such as static analysis of the generatedbinaries or code size {{cite:3088b1355e61682a72e0937bcc618cfa02985e33}}, {{cite:92a952de08a36939e56b69302e56b389eb5ad528}}. In contrast to these works, ourapproach enables noise-free ground truth measurements to be collected.", "A key challenge in supervised learning is feature design. The quality of alearned model is limited by the quality of the features used. Many prior worksused hand-crafted vectors of numericfeatures {{cite:4ee48f027494b1a99903ee2b7bd9d07453f1c9fe}}, {{cite:3088b1355e61682a72e0937bcc618cfa02985e33}}, {{cite:22f2837e65eca7821d9b33b7f7a9e505787d8f35}}, however selecting which valuesto include in a feature vector is a time consuming and error prone task. Anautomatic approach to feature selection was proposed in {{cite:4612720724f4dc7199bd8917219d0d696d6b7b11}} butthis requires a cumbersome grammar to be written to describe the space offeatures to search over. More recently, deep learning techniques inspired bynatural language modeling {{cite:0ffb70aa64725049c5eb4c4b7618e87803232389}}, {{cite:e7e1b899ecf254b2258666d7538300f8e94e3dba}}, {{cite:bdc3e1b935d13ff9803955a943ebdfa4c4c91ef2}} andgraph learning {{cite:3ecf48fdeb529f323e1a09d430b9d34782df9467}}, {{cite:d9ced6d45aa77e18ddbfdfda7eae44860abaa1b4}} have been proposed to simplify the task offeature engineering by automatically inferring high-level features fromlow-level input representations, however these techniques sacrificeinterpretability as the inferred latent representations are hard to reasonabout. Our approach aims to strike a balance between interpretability andfeature engineering cost by pairing a large number of the readily availablevalues from LLVM's static analyses with a machine learning algorithmthat automatically ranks and prunes features.", "There is active research around the differentiation of wholeprograms {{cite:41537272a4ec1f80a5b23ac2b1ef77c5d082d9ba}}, but as of today the whole compilation pipeline is notdifferentiable. A different approach is to optimize compilers usingreinforcement learning. Cummins et al. {{cite:92a952de08a36939e56b69302e56b389eb5ad528}} formulate a suite ofcompiler optimization problems as environments for reinforcement learning. Ameeret al. {{cite:60580f21a6d93bff8884aea6cf04e679d50163ba}} use reinforcement learning to make vectorization decisionsin the LLVM vectorizer {{cite:3898c0ead5ffd8e298d99f8b7fb15e56bb02fb36}}. The MLGO project {{cite:3088b1355e61682a72e0937bcc618cfa02985e33}} is aframework for integrating neural networks into LLVM, targeting the functioninlining heuristic. ESP {{cite:ba7741e01f7c27052510bd96a34e73306d0deda2}} is an earlier effort to apply machine learningto the problem of predicting branch probabilities. ESP uses neural networks anddecision trees to predict the likelihood of a branch to be taken.VESPA {{cite:35876d296d57b34d4852e6d2c29d4a8007855e9c}} extends BOLT and allows the use of machinelearning techniques for predicting branch probabilities. Their approach is similarto our work in that they predict branch probabilities that are later used for codelayout and other binary optimizations. Compared to these works which target individualoptimizations at a time, our technique enables a single learned model to benefitthe entire compiler by predicting branch weight metadata that is available toall optimization passes." ]

[]

[ "There are many general smart contract verification tools available that scan a contract for multiple security vulnerabilities. The usual techniques in these tools are static analysis, symbolic execution, and fuzzing. Some of the most well-known tools that use symbolic execution, static analysis or a combination of both are Manticore {{cite:3e55cef836be9fe5d89595ffe76f7166429586c0}}, MPro {{cite:d0c0cb69a49dc036e5c21ce8c80aba1d4ce4c186}}, Mythril {{cite:9c04e22eac8a03ad469dd9fb4eccc7011f99029b}}, Securify {{cite:7de0c58bb5c1c7fca13a0039f1c2dbfcf4a8fe87}} (deprecated), Securify 2.0 {{cite:2fdd1d6aa63b20b97a0970859b9242a27885e600}}, Slither {{cite:2d17ade7178a0220c010b8bc16cc1a1c0eaba745}}, SmartCheck {{cite:531e184c5a17c02a45ed82559538be2e966aff06}}, Verx {{cite:310e94f065a881a97d2a9e132812baa88742242c}}, and Zeus {{cite:8cdb6074e9a1b0f265918c531b080510cd6740ee}}. These tools can detect multiple security vulnerabilities; however, most of them either cannot identify gas-related vulnerabilities or if they can, their results are not reliable due to their high false negative rate (See Table REF in Section ). Fuzzing tools like ContractFuzzer {{cite:30dd1c4b0f23a467bd71c774a0899cc42dd92951}}, Echidna {{cite:869dad6300b10b161dd9a1605739e0759ab82a24}}, and Harvey {{cite:22eb90de68b315e9197070d48445b952e87aae1b}} do not discover gas-related vulnerabilities, to the best of our knowledge.", "Meanwhile, extensive research has focused on gas-related vulnerabilities of smart contracts. GASTAP {{cite:eba981840ffb2f7ffd8a6557e33577f7fc9ffef3}} derives gas upper bounds for all public functions of smart contracts via inferring size relations, generating gas equations, and solving the corresponding equations. Gasol {{cite:194593ce26bd7b2851d5149e84f6a1f7341d4f5c}} is an extension of GASTAP. It replaces the multiple accesses to the same storage location with one access. Also, GasCheker {{cite:6d2a81725e3e0398380586dc60d3c8f8c2fe1de4}} mainly focuses on automatically identifying gas-inefficient code in smart contracts based on symbolic analysis to save the gas usage for the creation of a contract or the invoking process of the functions in a smart contract. GasFuzzer {{cite:6194530cc3c19e2811fb097f3c50d01f7e1341f5}} proposes a gas-greedy strategy and a gas-leveling strategy to detect Exceptions Disorder kind of security vulnerabilities. Madmax {{cite:dc2c8abe7a490641ae0152609194b6e722b7aa88}} uses a static program analysis technique to detect gas-focused vulnerabilities automatically. It first decompiles EVM bytecode using Vandal, then detects gas-related security problems from the decompiled code in Datalog language, which is a logic-based language. Additionally, eth-gas-reporter {{cite:f80b3445077a9590970d2934a9da85e08083fbed}} reports gas usage per unit test in a Solidity smart contract. It can be used to get the amount of gas used by calling a smart contract. However, most of these cannot detect gas-related DoS vulnerabilities directly, or they do not provide any information to fix the problem.", "Gas Gauge can find all the loops in a contract reliably and quickly. Also, it identifies the exact functions and variables and provides an Out-of-Gas instance that helps developers investigate the problem further. Finally, to the best of our knowledge, Gas Gauge is the only tool that accurately and reliably finds the threshold values and provides more useful information like each loop's type and the variables affecting it. This information is helpful in order to repair the code and prevent gas-based attacks like DoS with Block Gas Limit." ]

[]

[ "As we mentioned in the introduction, there are quite some methods for the automated reasoning in modal logics. But relatively few implementations are available. The matrix method and the functional translation method require specialized unification algorithms. In contrast, it is very easy to write a relational translator.", "Another advantage of the translational approach is its generality. It can be applied to many modal logic systems. Changing from one system to another, we need only modify the axioms. On the other hand, many other methods and tools are designed for a few specific systems. For example, semi-functional translation {{cite:dd257f7272d685e06074ca23f636a4632d978491}} and the technique described in {{cite:7ff581f8283666a166d935b49a148a0413a8dfcc}} are most suitable for serial modal logics like KD and KD45.Modal resolution rules are discussed in {{cite:1f8ea5d7bd6530080716e4f7e4e2178768406d01}}, within 5 systems, i.e., K, Q, T, S4 and S5.The program LWB has built-in proof procedures for 5 selected logics, and KSAT {{cite:366d46f68423fb59c5783086aefdc07dffbb9634}}, {{cite:b9d79ebfe92c3ce8ace8d3de00262b8c4169cfd5}} works only in K.TABLEAUX can deal with many modal systems. But if one was to extend it with new systems, new procedures (encoding the semantic properties) would have to be added.", "Tableau-based decision procedures can be used both to prove theorems and to show the satisfiability of formulas. As we point out in the previous section, the models found by our method are usually different from those generated by tableau procedures. An example in linear time temporal logic has been given to show the difference.For another example, the formula 4. {{formula:46aa430c-3137-482f-94c0-9ec71e5fef1b}} is not valid in K.Catach {{cite:db6e54ef764c608ba4436d7e2649b783356aa5b3}} gives a 3-world countermodel produced by the tableau-based procedure. But with our method, we shall first find the smallest model having 2 worlds:{{table:3931f7f1-daf0-47c4-a3ba-c21c52cfe2d3}}", "where {{formula:1132f290-8d8c-42f7-983f-3d5411090902}} is the real world.", "According to our experiences, many satisfiable formulas have very small models. Thus it is practical to use model generation tools in the classical logic to show the satisfiabiliy of propositional modal formulas. However, for some modal logics (like K), there are some specially constructed formulas {{cite:96fdfdc6ed5a892cb8793db04bc57a3d068e7f17}} which are satisfiable only in exponential-size models. They present much difficulty for methods based on the Kripke semantics. The program KSAT can deal with such formulas efficiently {{cite:b9d79ebfe92c3ce8ace8d3de00262b8c4169cfd5}}." ]

[]

[ "Our paper builds off a line of work on trend filtering, which, recall, enforcessmoothness by penalizing the {{formula:c9060b02-0f66-42f8-90db-0dad9f890865}} norm of discrete derivatives of a givenorder {{cite:dfa4f1675572f4ada261ca3b965fef09b9b57258}}, {{cite:33c6f9e9f9440828a6f667f7e8109ed6b0180861}}, {{cite:af01f61b48830c2173c8b61982f528857c41de8b}}, {{cite:a73434541b628d7282c65613864b36bcf1acef06}}, {{cite:882c6fbb70df530e5b529e17aebafedf094a8870}}. This can be seen as a discrete analog of thelocally adaptive regression spline estimator, which penalizes the TV of agiven order derivative of a function {{cite:84f8454335eb91948af9625354a1108a03f92b18}}, {{cite:9271dbc479ab147a203afb83ebbc206adcd468b3}}. For an overview of trend filtering, its connection toclassical theory on splines and divided differences, and related results thatbridge discrete and continuum representations (such as that connecting trendfiltering and locally adaptive regression splines), we refer to{{cite:e3f15247f91f6690c223de0e423efe5ca9c2ab44}}. Trend filtering was extended to general graphsby {{cite:0e95e534961a08e1e14f7d4bcc7cc8f0c089b936}} (more on this below), and was studied in an additivemodel setting by {{cite:fbd3790acfbce8dc12a584be8a374520c0eb5366}}.", "Kronecker trend filtering was proposed in {{cite:9b0ee3f632bf469fdc6442e158168d0f28bcfe38}}. Minimaxtheory for KTF has been developed in different special cases, distributed amongseveral papers. Minimax results for TV denoising on lattices can be found in{{cite:0f8bbaa71064d53abdb84ba1b9c46a6381624b18}} (upper bounds) and {{cite:bdb9725921ca507048af8c65f96264f22d21e315}} (lowerbounds); with respect to KTF and KTV classes, this covers the case of {{formula:27b970d6-2183-417f-aff9-4c34b2717408}} andall dimensions {{formula:2b833398-cc4f-450b-a6cb-b10edb92db2c}} . Meanwhile, {{cite:9b0ee3f632bf469fdc6442e158168d0f28bcfe38}} derived minimaxresults (matching upper and lower bounds) for all {{formula:7aa8b6b7-9515-41a3-b5c5-6f5b6aa70e0d}} and {{formula:0f5bbe15-fad7-4901-b2d3-ab560585ff35}} . The currentpaper completes the landscape, deriving minimax theory for all smoothness orders{{formula:384c2cff-90b2-49ea-9a7e-c73219aade9a}} and all dimensions {{formula:ff99c5e7-0dca-48a2-b627-ac08fdb392c1}} . The majority of this paper (including the minimaxtheory) was completed in 2018 and can be found in the Ph.D. thesis of the firstauthor {{cite:d88f75990c07838840d9be7fb908746fd371a1c9}}." ]

[]

[ "The methods to solve Autonomous Driving problems can be divided into two categories i.e Rule based and learning based algorithms. Although rule-based techniques have had some success in the past, learning-based approaches have also shown their effectiveness in recent years.", "Many conventional solutions are based on explicitly written rules and rely on state machines to shift between specified decision behaviors. For example,the \"Boss\" that is trained by CMU {{cite:8939c8016f8288cd3482b51c96488d034e8d6970}} discusses a rule based approach, also team of researchers at Stanford {{cite:b0fafb2cc7c244bedb45f450f5daf69d4404d822}} used reward designs to determine trajectories.However, reliable decision has high amount of uncertainity in rule based approach.", "Learning-based techniques, as a key AI technology, can deliver more advanced and safe decision-making algorithms for autonomous driving. Recently, NVIDIA {{cite:168596ad4b678c6ae8794b359fbacedd77227ef0}} researchers trained a deep convolutional neural network (CNN) to train images directly from camera to cluster control. The trained model was able to handle the task of lane change and driving on a gravel road.", "In addition to supervised learning, reinforcement learning results have significantly improved over the years. Wolf et al. {{cite:c3db3c34b3fa20e28e7e3f51a22d2552c82ac795}} have presented a method by using DQN to drive a car in stimulated environment. Only 5 action were defined each corresponding to a different steering angle along with training images of size 48{{formula:5157eb31-b709-47f9-8e02-9fe0a32ca46c}} 27. The reward function was calculated by taking into account the distance from lane center as well as certain auxiliary information (such as the angle error between the vehicle and the center line). Hoel et al. {{cite:6fe0b6dc8b7fd07348f8889c9b413ce2d8ee10fe}} used DQN to solve lane change along with vehicle speed. They used a one dimensional vector to define many components including speed, surrounding vehicles and adjacent lanes instead of using front facing images. Wang et al. {{cite:ce14b45c970cce22ea22fedd1280dd18ab57c1c4}} used safety rate to measure the quality of their model which was calculated based on collision frequency which provided clear idea about stimulation.The results show that the models which considers the speed along with other factors usually outperformed the other models that did not consider the average speed of vehicle." ]

[]

[ "Private set intersection is a well-studied problem (e.g., {{cite:939d25edaf34d8f1613e9be83afa48f3b7dc95b8}}, {{cite:3466e592579ad794cead95167fae01ba525082ff}}, {{cite:e38bfbcbcd79a01b58667a8030fe617dcb536aca}}, {{cite:bd8feddcb70afaff753e2d1d0c865da05b30cf56}}, {{cite:b5559b887332db9d620caee4d6a5712b3aaa35ff}}, {{cite:fdc0b3252220251f6a73b0dadfee0f8cdecab17d}}, {{cite:7de6af7ae72cb31d117b53a50f5ca87fc441aa03}}, {{cite:47f3d6a15e3ae35f1401b42953654fa18d399c0a}}).We refer to these for details on general PSI and describe primitives necessary for our protocols.", "Threshold PSI:Threshold PSI is a relatively unexplored topic. In this line of work, the goal is to reveal the intersection of two setsonly if the number of items in the intersection exceeds a certain threshold. Note that this is different from the notion introduced by {{cite:3466e592579ad794cead95167fae01ba525082ff}} in the context of multi-party PSI where the goal is to include an item in the intersection only if a threshold number of parties share this item.{{cite:6f76a9be32ce525e718e00080fcbdb72f40c3727}} designed a threshold PSI protocol based on a threshold key encapsulation mechanism. However, the communication cost of this protocol scales quadratically with sets sizes. {{cite:071ddb8222215e760a70b707f0c53b50cee23b46}} proposed a threshold PSI protocol with communication cost linear in the set sizes but expensive computational requirements due to the use of single-server PIR. The state-of-the-art protocol is due to {{cite:198b7325f25bad8ef27b9e69fc6ce86d44d12597}}, which builds on the idea of set reconciliation by {{cite:360b9d51c8917b3cc89f45d2a074104fe01881b0}} and uses oblivious polynomialevaluation as the underlying secure two-party cryptographic primitive. A similar construction by {{cite:485dee419bd0b2b594cd8620890110623ecadce8}} replaces OPE with an additively homomorphic encryption scheme.", "Privacy-Preserving Proximity Detection:A loosely related task to DA-PSI is privacy-preserving proximitydetection {{cite:605d195c346a3d816ab1215b7867ffa9811f382e}}, {{cite:d22b502b4922703caf6456596aaed6df04b32e48}}, wherein a trusteduser of a service (say Alice) wants to learn if another user known toher (say Bob) is currently nearby. One may envision building aDA-PSI like protocol with this primitive such that two points areestimated to be within a distance threshold of each other when one ofthe points is contained within a “ball” of radius equal to thethreshold centered around the other point. However, current techniquesdo not generally support this idea; proximity detectiononly works in 2D Euclidian spaces and extending these ideas even to 3Dwhile ensuring similar privacy goals is non-trivial. In contrast, ourgoal is to design general protocols for DA-PSI over metric spaces ofarbitrary dimensions.", "Private Collaborative Blacklisting:Collaborative blacklisting (e.g., {{cite:99bc0a3c784ebed7a6328155671a745441ec8d60}}) allows parties to share information observed in their networks with others and jointly identify malicious traffic identifiers, e.g., IP addresses. This information may be later used to train models to detect likely malicious connections in the future. For mutually untrusting parties, this process is carried out with a PSI protocol {{cite:f5a1c649ce8f3fabc915917ca5c6032eba8dbc55}}. The PSI protocol can be configuredaccording to desired functionalities; e.g., if only the number of common items is required then a PSI-CA protocolis used instead. Based on this, parties may either proceed to share their information or abort.", "Iris Recognition:Iris recognition for biometric authentication reduces images to binary vectors and then compares two images based on a Hamming distance computation{{cite:3368214e45096fd9bfd46c4ac0c8ae20b5fadf21}}. The vectors required however are long: more than 10000 bits vectors are required. A privacy preserving version of this problem relies on partially homorphic encryption to compute the Hamming distances but the really long vector sizes are a communication bottleneck {{cite:c379c453ca133a243844a989dd1f38e74712547b}}. On the other hand, a match can be detected with low Hamming distance thresholds.Therefore, a Hamming DA-PSI protocol where communication depends on the distance threshold, and independent of the vector lengths (as the one presented in the paper) is crucial." ]

[]

[ "In this section, we review the related work to dynamic network and activity state models, as well as temporal motifs.", "Dynamic network models.There is a body of work on modeling dynamic networks, mostly by extending a static model to the dynamic setting {{cite:ef6d166acdb020c7d35216c8a3439033cf233572}}, {{cite:fe133087f7b2b7e5e8189973d8475ad5aa9d67f5}}, {{cite:28b288a4b7afce77ab130743ebdf5408a8e0143e}}, {{cite:dbf13506a3680f68555c4b7ebd21ab39ee60ea01}}, {{cite:5e77a6cb894dc10f03c750ea1782b3c64a660751}}, {{cite:0e0d1cfade41293fdcbd70d9a44ca8dcda811492}}.Among the existing methods, temporal extensions of stochastic block models (SBMs) are the most relevant to our work. In particular,a dynamic SBM, in which a transition matrix specifies the probability of nodes to switch classes over time {{cite:dbf13506a3680f68555c4b7ebd21ab39ee60ea01}}, {{cite:28b288a4b7afce77ab130743ebdf5408a8e0143e}}.However, in contrast to our work here the above works both assume that edge probabilities do not change over time.More recently, models in which the network snapshots are modeled using SBM {{cite:0e0d1cfade41293fdcbd70d9a44ca8dcda811492}}, {{cite:5e77a6cb894dc10f03c750ea1782b3c64a660751}}, and the state evolution of the dynamic network is modeled by a stochastic dynamic system have been proposed.Other temporal extensions of block model include {{cite:c9567fff25455aad0d29a200d235b90835729358}}, which presents a non-stationary extension of the SBM; and {{cite:9a14a17fd77bcc0fd90dedb494966de28fbbf485}} that performs clustering on temporal networks, and uses Markov chains to model node groups.Our work here differs in that it focuses on temporal graphs, the model allows for linear time fitting and provides closed form expressions for temporal motif counts.", "Temporal motifs.Paranjape et. al. {{cite:1a14020a4d53df71a6233e5cdd06fb924f758b2b}} extended the concept of static motifs to temporal networks and proposed a framework for counting the exact number of relatively small temporal motifs. Very recently, {{cite:3bdb7f554bc6d3a55326b7edd816dceba66faa4e}} proposed sampling methods for approximating the number of temporal motifs.Existing methods either do not account for ordering of the temporal edges {{cite:eff75810f16838f63fc8a12a1733cadd75516188}}, or require temporal edges in a motif to arrive consecutively to a node {{cite:eb4d2d2732c5e0a01ecd5e7dceb397e8f8c36d54}}. Among other examples are the algorithm of {{cite:21a90ea62f0251cad7841428e1b032662b5e10f9}} that uses ideas from sub-sequence mining to identify patterns in temporal graphs, and methods of {{cite:d0c655cf7939e194485dee18e1c2a2abbd9a8921}} for finding temporal isomorphic subgraphs.While this line of work aims to accurately count or approximate the actual numbers of motif instances in any graph, the goal of our work is to determine the expected number and variance of motif instances in a graph, given its underlying statistical model.", "Currently no statistical model of motifs in temporal graphs exists. However, there have been several heuristic/empirical techniques proposed.Such approaches are based on generating a large ensemble of randomized temporal networks and include shuffling and reversing of timestamps {{cite:2f50c72062c63538a2d14f6f1f1003c00716584d}}, {{cite:53832a55fff469983ecf4c93bb50fbf15271234d}}, {{cite:27dccabab1f97871d36446e84a20c5290124bc34}}, {{cite:eb4d2d2732c5e0a01ecd5e7dceb397e8f8c36d54}}, {{cite:ddfeed9b95b3a866a4b3d9c07fba4de113e6c49b}}.Unlike these heuristic/empirical methods, we develop a statistical model and closed-form expressions for motif counts and their variances. Crucially, our approach does not require expensive shuffling/simulation of network ensembles, and this scales well to large temporal networks." ]

[]

[ "Unconditional image synthesis and GANs. Unconditional image synthesis is a generative learning task which seeks probabilistic models that are capable of explaining away data under the maximum likelihood estimation (MLE) framework. To this end, introducing multiple layers of hidden variables is often entailed, which in turn popularizes DNNs. There are two types of realizations in the literature: top-down generator models {{cite:d8a5377355d2480ef815636a2a762ebfc49c8550}}, {{cite:df8ebc6de038a94b3d0275fb113c9139abcd585b}}, {{cite:a9c6427ff190366829ad6de982b13892d33ead78}}, {{cite:da0050ee9f48ec8ec011977ae527f82872409878}}, {{cite:a20c05acab1bdbf38a9d3dc04e25ba792a0b47de}}, {{cite:19872fd6905aa5149dd95bf37ece5c17f179a7bd}}, {{cite:c3f7cf4a7f443e40607b63193af79f24fcdb657e}}, {{cite:9a76a9ec3aa828a10a1d81e4a94f4845cbcce25c}} that are non-linear generalization of factor analysis {{cite:dd8649c767055195f54ab38a4c7aacd700318e14}}, and bottom-up energy-based models {{cite:87ebd1d1c71eaa98195f01e1a50a769f85a94b2c}}, {{cite:f39430760fa85c1309701b512b65cae9a264b342}}, {{cite:f1166198449bacfb8c9bcbf0bd36e3ba4c50f5b5}}, {{cite:bf47aeb4b7a5134b22fed2d9e33bef0adbbec365}}, {{cite:8335cca7b64bec5ee9107983caceff49811cb1ac}}, {{cite:a2fbefe507a8026c3c95b3c3c516b7276292f5cf}}, {{cite:8881d7767ebfa56a78dbebb58fc0c21ee45b10c6}}, {{cite:30ab31a6187fedd7f9e5edd5494279a6b3d5e7ed}}, {{cite:54c5069256bbbd4996b901c70e97293498afb2ab}} that are in the form of a Gibbs distribution or Boltzmann distribution. It is well-known that training a DNN-based generator model individually is an extremely difficult task, mainly due to the difficulty of sampling the posterior for real images. Adversarial training is a popular workaround and GANs {{cite:d8a5377355d2480ef815636a2a762ebfc49c8550}} are widely used in practice, which are formulated under a two-player minmax game setting. The proposed SLIM module can be used as a drop-in component in the generator network of GANs.", "Light-weight GANs with Few-Shot Learning. Compared to the extensive research on light-weight neural architectures in discriminative learning for mobile platforms, much less work has been done in generative learning {{cite:31bcd185133fb124077f22694345ed2f9908ef24}}. The residual network {{cite:b0821bdd5bbd01659beafd50d10d61ac9321b270}} is the most popular choice, on top of which powerful generative models such as BigGANs {{cite:3254ca92f96973fa43a5fbe036af7c9c277efea2}} and StyleGANs {{cite:063970539201a52b25409109aef455ec130fd5f8}}, {{cite:be7aa758509a250385eb9a59b838a5cec0bca3ca}} have been built with remarkable progress achieved. For high-resolution image synthesis, these models will be very computational expensive in training and inference. In the meanwhile, training these models typically require a big dataset, which further increase the training time. Few-shot learning is appealing, but very challenging in training GANs, since data augmentation methods that are developed for discriminative learning tasks are not directly applicable. To address this challenge, differentiable data augmentation methods and variants {{cite:f7d728c3bae6a307bb02fb28039751a020963759}}, {{cite:59cc72b59a60ccf410b0b0f271aaf73209f4902f}}, {{cite:1fb8ac0e891d923916d4f16de20d8f22a3b4d1e9}}, {{cite:b51e8f71137ddba4700e5791acc23dbe641a83b2}} have been proposed in training GANs with very exciting results obtained. Very recently, equiped with the differentiable data augmentation method {{cite:f7d728c3bae6a307bb02fb28039751a020963759}}, a FastGAN approach {{cite:9420a8de32caf92929b3d0f7cc24239333518c0d}} is proposed to realize light-weight yet sufficiently powerful GANs with several novel designs including the SLE module. The proposed SLIM is built on the SLE in FastGANs by exploiting the well-known Google's Inception {{cite:78010e6a052bad5c207c82bfb43d40afb5320211}}, {{cite:f8a1f4d29755cbddf4b0bd2115621303a850b382}} building block design and the Átrous convolution {{cite:0e13411da6280e38991743e387ac68174932f6b1}}.", "Learning Unconditional GANs from a Single Image. There are several work on learning GANs from a single texture image {{cite:fea715246288106b6dc28cc50e562c8edd810008}}, {{cite:97ad906f2d020bebd4b820a5133ae21bd40250e5}}, {{cite:b184c2378ce25ba6a50d19153f1c1b8a1ea7ea08}}. Recently, a SinGAN approach {{cite:8743dafbc24ce0e45e21d49175774f9e019dd0e2}} has shown surprisingly good results on learning unconditional GANs from a single non-texture image. It is further improved in ConSinGANs {{cite:2b4c77d5d3d97ab9d5dc9f6241723abbeade829d}} which jointly train several stages in progressively growing the generator network. However, it remains a challenging problem of preserving image structure in synthesis. The proposed SLIM is applied to the vanilla SinGANs {{cite:8743dafbc24ce0e45e21d49175774f9e019dd0e2}}, leading to a simplified workflow that can be trained in a stage-wise manner and thus more efficient than ConSinGANs, and facilitating a stronger capability of preserving image structures.", "Attention Mechanism in Deep Networks. Attention reallocates the available computational resources to the most relevant components of a signal to the task  {{cite:ced6e42b79e44c235e64f98620751153bcbb612d}}, {{cite:9ecf5b1841a41dd27f53e84eaebd831ce0812f7e}}, {{cite:c6e3cdae3fc58cffc0693bd47171ceb99979cfe3}}, {{cite:a4b940ba0d2323356cc9abfd03d45bf536c455bb}}, {{cite:ecfb01bb96812f7ebec5312523c54f00f2f02482}}, {{cite:89e71c406e71cded32f873d33027d46b9eae6712}}. Attention mechanisms have been widely used in computer vision tasks  {{cite:b75257da4fecae6b71933cd67874dcadde6ad66a}}, {{cite:576ecbefd017861697420c6bb030f1206d142916}}, {{cite:4624bc31cf46fe5dfcc09c6826aca7894f32dd75}}, {{cite:5689e3197eecc0de4e2f2b1dd060cecefc2bc774}}. The SE module  {{cite:27f00f4bb90b189da9a123f6e42919c0683ce732}} applies a lightweight self gating module to facilitate channel-wise feature re-calibration. Our proposed SLIM module incorporates spatially-adaptive feature modulation, while maintaining the light weight design, improving the representation power for efficient discriminative learning in addition to generative learning.", "Our Contributions. In summary, this paper makes the follow four main contributions in unconditional image synthesis using GANs and image analysis: (i) It proposes a Skip-Layer Inception Modulation (SLIM) module to facilitate better learning of generative models from few-shot / one-shot images, and to provide visually-interpretable visualization of the skip-layer information flow. (ii) It shows significantly better performance for high-resolution image synthesis at the resolution of 1024{{formula:f7908bf2-6f9c-4d89-a029-364eb8c7a971}} 1024 when deployed in the FastGANs {{cite:9420a8de32caf92929b3d0f7cc24239333518c0d}}, while retaining the efficiency. (iii) It enables a simplified workflow for SinGANs {{cite:8743dafbc24ce0e45e21d49175774f9e019dd0e2}}, and shows a stronger capability of preserving image structures than prior arts. (iv) It further presents a Same-Layer Inception Module for image classification, which serves as a stronger new realization of the ResNeXt-like models when deployed in the ResNets.{{figure:a054b4cc-19d5-41a9-a0ea-b7ffdb5a0c68}}" ]

[]

[ "Connected Vehicles and Infrastructure:Connected vehicles promise great opportunities to improve the safety and reliability of self-driving cars. Vehicle-to-vehicle (V2V) and vehicle-to-infrastructure (V2X) communications both play an important role to share surrounding information among vehicles.Communication technologies, e.g.,, DSRC {{cite:7afd4d23f8b1759541a846c27655fd667e5db7c8}} and LTE-Direct {{cite:bb2efc0885cc88efcc21f3701ee712866c10892c}}, , provide capabilities to exchange information among cars by different types of transmission, i.e.,, multicasting, broadcasting, and unicasting.Automakers are deploying V2V/V2X communications in their upcoming models {{cite:e0a2b92c22d4a8d0d3498761a9b8850d73a8d8d6}}, {{cite:6bc000ac471133ccb1b46dcd783a8e166742e8ec}}). The academic community has started to build city-scale advanced wireless research platforms (COSMOS {{cite:c75cfcc7b462b882077c3e657769dd551ed83a86}}), as well as large connected vehicle testbed in the U.S. (MCity {{cite:68c428724728543f1265967a7ea461926a3d27b0}}) and Europe (DRIVE C2X ), which gives an opportunity to explore the application feasibility of connected vehicles via V2V communications in practice.", "Sensor/Visual Information:Collecting visual information from sensors (e.g., LiDAR, stereo cameras, etc.) is a major part of autonomous driving systems.These systems rely on such information to make a proper on-road decisions for detection {{cite:be065e0ac010b0449ecca67e7f4cf41749570aae}}, tracking {{cite:e070479e4f404017a5813c5a9be63332608b1dd8}}, and motion forecasting {{cite:3d455b139881b1b2dd5f14123e5a94139f894f7d}}. In addition, there is a large body of work that explores vehicle context and behavior sensing {{cite:8f9d7f354cbc41948e2bcb60bf3185fe34ac1113}}, , to enhance vehicular situational awareness.Thanks to advanced computer vision and mobile sensing technologies, all the sensing capability can already be leveraged efficiently in a single car setting {{cite:7ac6d09f2a889a35822a0fb40ccc244fce308e45}}. This work, and several related works discussed below, take the next step of designing how to share this information among nearby vehicles.", "Vehicle Sensor Sharing:Past research has attempted to realize some form of context sharing among vehicles. Rybicki et al. discuss challenges of VANET-based approaches, and propose to leverage infrastructure to build distributed, cooperative Traffic Information Systems. Other workhas explored robust inter-vehicle communication , {{cite:6fcf79ed574736ec4a8e024b2af0d925468d9a4b}}, {{cite:eabc72d39068e066f43edd44cc97cb7b9d547004}}, an automotive ontology for intelligent vehicles {{cite:8df1dfd3bdac220147f05aa3178e2cd235007c10}}, principles of community sensing that offer mechanisms for sharing data from privately held sensors {{cite:ee4d40c8b2b68079e6e1ad95c3d642c681adba9d}}, and sensing architectures {{cite:cf02826c484614ae3a577d2f16c01601f5ac9c26}} for semantic services. Motivated by the “see-through\" system {{cite:46b3e64c2f8da021872236c98c0ac1cabbd4f981}}, several prior works on streaming video for enhanced visibility have been proposed {{cite:2021c02ad4571b8d99be9c26e760b1771d959837}}, , {{cite:f36b1633ebd575ae40c945e09e0919c0c4ff7cab}}, {{cite:31e0def2732ccce4e6d7bdaa0a11ef86dcf096a0}}, {{cite:6040e6741bf6697b743d8c0c5cadb3088ac349ce}}.The above line of work has focused on scenarios with two cars, where a leader car delivers its whole video to a follower vehicle, which in many scenarios is not sufficient.", "Most recent work , , {{cite:95238df4e17dcf2667013a2e4fc71d1f46d43764}} demonstrates the feasibility of sharing point clouds, but with limited scale and infrastructure support. In this work, we focus on enabling clusters of vehicles to share sensor information at scale in the absence of edge servers.We design an end-to-end cooperative perception framework to compute when, which views, and to which cars to deliver views over time.While enabling on-demand cooperative perception, when infrastructure is available, our approach can naturally extend and work as-is. Because of the design of spatial reasoning and metadata exchange, as shown in §, autocast can scale better than existing baselines." ]

[]

[ "The use of artificial intelligence for the security of cyber-physical systems is not a new endeavor, with several other works also presenting anomaly detection methods {{cite:decf5cf0ac8ae0920279f9f0e85b3434b6ae7526}}, {{cite:952454b5c94a8e9a4820537cd9eecf8c2e36a0e3}}, {{cite:787893c5f4e82dc052be2deb13949c5cb1940e7b}}.", "In {{cite:787893c5f4e82dc052be2deb13949c5cb1940e7b}}, Dmitri Shalyga et al. compared the abnormal detection rate of different types of neural networks, for identifying attacks on a CPS. The future sensor readings were predicted based on a previous window of a fixed duration. An anomaly is detected if the expected error between predicted and actual values is greater than the 99th percentile. Furthermore, with the intention of improving the quality of the anomaly threshold, several techniques were employed such as exponentially weighted smoothing, mean p-powered error and weighted p-powered error. Finally, due to the unbalanced nature of the dataset, a disjointed time window technique was applied, where the window used to forecast at given time point might not be the one immediately preceding it. With this setup, a multilayer perceptron (MLP) based model achieved an F1-score of 81.2%.", "In {{cite:952454b5c94a8e9a4820537cd9eecf8c2e36a0e3}}, Moshe Kravchik et al. used a 1D-CNN to identify anomalies in the SWaT dataset. A threshold is estimated from the model's prediction error in the training dataset. Using such a threshold, the authors were able to distinguish attacks from normal behaviours, through a z-score function. To approximate and select the local optimal hyperparameters, a grid search was employed. In this manner an F1-score of 87.1% was achieved. In {{cite:65efedc73cab772c5df55fc7116b563b11294eb4}} the same author applied the previous approach for anomaly detection, using a mix of grid search and genetic algorithms for tuning the hyperparameters of an autoencoder network. The author obtained an F1-score of 87.3%, an improvement from his previous work.", "In {{cite:59387e6690c3093a0eb5f9a6d2a9d2e4f105f098}}, Giulio Zizzo et al., achieved an F1-score of 81.7% with a Long-Short Term Memory (LSTM) based model. Using grid search to compute the best values window duration and error threshold.", "Jun Inoue et al. in {{cite:decf5cf0ac8ae0920279f9f0e85b3434b6ae7526}} have investigated the detection capabilites of two unsupervised machine learning algorithms, a Deep Neural Network (DNN) composed of a single LSTM layer followed by multiple feed-forward layers and a one-class SVM. The first implements probabilistic outlier detection by judging low probability data points as anomalies while the second resembles a more straightfoward application of the one-class SVM. For the SWaT dataset the DNN obtained an f1-score of 80.3% as the one-class SVM achieved a slightly smaller value for the same metric, 79.6%. Furthermore, the authors have concluded that both methods share the same limitations: detecting anomalous actuator behaviour and identifying gradual changes of sensor values.", "In {{cite:242eaf3b85d3ce2c23c91bed6a28ce69ef67fc36}}, D. Li et al. presents a novel anomaly detection approach using Generative Adversarial Networks (GAN). In this work two LSTM Recurrent Neural Networks (LSTM-RNN) are trained in different roles, one is trained as a detector finding anomalies that deviate from the baseline, and the other as a generator capable of creating simulated normal baselines. These two models are then introduced as adversaries in an iterative feedback-loop training process, obtaining an F1-score result of 77% on the SWaT testbed.", "For this dataset, the best results we found in the literature, 88.2% F1-score, was obtained by M. Elnour et al. in {{cite:1c4a69da3c3c694ddef38316e834ea5087b4aefd}}. An isolation forest based approach to anomaly detection, that separates anomalies from normal observations by analysing both normal data and its Principal Component Analysis (PCA) transformed representations." ]

[]

[ "Several secondary studies have been conducted on (security in) microservices systems. In a recent review study {{cite:350c56ff509b304aa0ad8ef21e5c3516a195ac46}}, Waseem et al. observed that there are a few concrete solutions for addressing security concerns when implementing microservices systems in DevOps. Another systematic mapping review on 46 papes by Hannousse and Yahiouche {{cite:1f3817091fa96092f55767b901b28ab3f0c18efb}} revealed that most studies on security in microservices are the solutions proposed in the soft-infrastructure layer. They argued that internal attacks, compared to external attacks, are less explored in the literature. They also indicated more efforts should be allocated on other layers of MSA (e.g., communication and deployment layers) and developing mitigation techniques. Pereira-Vale et al. {{cite:af8df0f873828cc669d008fcb2d90ecade053c98}} carried out a multivocal literature review on 36 academic literature and 34 grey literature. Their review led to a classification of 15 security mechanisms, in which authentication and authorization are the security mechanisms most reported in the literature. In addition, mitigate/stop attacks are expressed in about 2/3 of the security mechanisms. Ponce et al. {{cite:e80a875aa91d95dd842a7487b84054295d671855}} focused on security smells and refactorings and collected and analyzed 58 white and grey literature published from 2014 to 2021. They found ten security smells with their security properties and corresponding refactorings. They also elaborated on how the corresponding refactorings mitigate the effects of the smells.", "Waseem et al. {{cite:b029cba30af50e12133c770b1e5b5ea13688788d}} empirically investigated the issues reported in five open source microservices systems hosted on GitHub. They found that 10.18% of these issues can be attributed to security. Yarygina and Bagge {{cite:2c68845c144e4c6a46e586171c10646edce070a0}} asserted that security in microservices systems is a multi-faceted problem, and a layered security solution can address it. Hence, they categorized microservices security concerns into six layers (hardware, virtualization, cloud, communication, service/application, and orchestration) with solutions to address them (e.g., secure implementation of service discovery and registry components).", "To secure IoT microservices, Pahl et al. {{cite:fe54eca75e33afc82cbd9e5cff3253031ee4b298}} proposed a graph-based access control module in a network of IoT nodes. This module can monitor the communication of IoT microservices systems to create robust security and mitigate the security holes of such systems. Moreover, Pahl and Donini {{cite:077f8b77399d17b372420e7135db5565247cbe28}} provided a method based on “X.509 certificates” for authenticating IoT microservices. One of the main goals of this method is to verify the security properties locally through the distributed IoT nodes. Yu et al. {{cite:9defb9d5cb93b8f3a8d52b4021d78afef36a182d}} focused on security issues of microservices-based fog applications because such systems are composed of numerous microservices with complex communications, which is a challenge in terms of security. They reviewed 66 articles and identified 17 security issues (e.g., kernel exploit or DOS attack) regarding the microservices communication categorized in four groups: containers issues, data issues, permission issues, and network issues.", "By surveying 67 participants, Rezaei Nasab et al. {{cite:29b440d8a3d826196694f91486fbf9852f9500a8}} affirmed that security challenges in microservices systems differ from non-microservices systems. To bridge the security knowledge gap among microservices practitioners, they developed a set of machine and deep learning approaches to automatically recognize security discussions (including security design decisions, challenges, or solutions) from open source microservices systems. Chondamrongkul et al. {{cite:e6c32e0773d7eafa097015cb4cf8baa2c889139b}} also developed an automated approach using ontology reasoning and model checking techniques to identify security threats of microservices architectures through analyzing security characteristics. The identified security threats show how the attack scenarios may happen. Sun et al. {{cite:499805ff7674299791d77407c09470fa050c9192}} designed an API primitive FlowTap that provides security-as-a-service for microservices-based cloud applications. The proposed technique can monitor and protect the network of such systems from internal and external threats.", "In contrast to the works above, our study presents 28 best security practices for microservices systems. These practices were collected from developer discussions occurring during the development of microservices systems. We further validated the usefulness of these practices by seeking feedback from 63 microservices practitioners. Finally, we articulated the positive and negative sides (if any) of the 28 practices." ]

[]

[ "We briefly review related works on traditional FAS methods, and then detail domain generalization-based methods and representation disentanglement-based methods which are two popular research orientations based on deep learning. Finally, we give a sketch of Bayesian deep learning.", "Traditional Face Anti-Spoofing. Early researchers have introduced lots of handcrafted features to achieve FAS task, such as LBP {{cite:be4fb8bd592b0df3b1c559e2a98fa1e7d7a11be2}}, {{cite:83d950d57e8fbfbe24e2c2362c4d897d6690f77f}}, {{cite:dc3cb59a733798e64ac2a7c9f08da3d0d16088d5}}, {{cite:003195ef9f8d112a9459d7231cb4da9537a27345}}, HOG {{cite:14a2150b0e24da9cabf8bda7371c283e78ed3234}}, {{cite:c6713dbae6bd129fa36474af02fab0ec750d2bbb}} and SIFT {{cite:17a3b9301a5a283552cb7cbba09d24701f139d93}}. Since they are too simple to perform well, more liveness cues are explored later, such as eye blinking {{cite:5757fdfcbcb0a0ad9169448937f3ea1feb224665}}, face movement {{cite:ea175750ebffaac65d88749a7dfc5a2332f6d540}}, light changing {{cite:d8cd78148fd85f7ca3e824768e4cc663cc3c1642}} and remote physiological signals (e.g., rPPG {{cite:11e0485a3ec46cf84320c25e6d6353105ad5cb46}}, {{cite:e4f65acfde6173e9da710ffa0ca668995ff6ce94}}, {{cite:20d1e823374d5f5dcec0a2053e2360e7f7905a92}}, {{cite:d04beeb2ae30517ae6c6f03a90e7cc6682b20732}}). However, these methods are always limited by low accuracy or complicated process in video data.", "Domain Generalization-Based Face Anti-Spoofing. Although deep learning facilitates the FAS task, the generalization ability for multiple domains still need to be improved. To this end, researchers have tapped the potential of various techniques. Some approaches measured and constrained the distance of features or domains to obtain domain-agnostic features. For instance, {{cite:7681b7ff68492a6a7e24f4c2099b529130252cf9}} used the MMD distance to make features unrelated with domains. {{cite:f93486c9c8bd9a537e8cb2c6d3fbd36ec1b2591c}} and {{cite:7b18971f7365f3e4d223d639cf2e8ebed7c7dca7}} introduced triplet loss {{cite:1a12a97ada311e28c1182447164a9662c9f76bb1}}, {{cite:0508da1255b54ef0bd58eb186bfa1f2f7a6dbc3e}} and {{cite:b893d31143a6b0e91f23b6ab2cc44df44beb21c4}} even constructed a similarity matrix to constrain the distance between features. Also, many meta-learning-based methods were exploited to find a generalized space among multiple domains {{cite:4174a2d9742fd6f20a516e5b380153ffd57d0108}}, {{cite:0a0309e72d6efffab4510dc26a4c9e5410b5093c}}, {{cite:b863f81434102f7d9a370b81e47b768c9d2eb3e4}}, {{cite:7a640a5ef67b0114385aeec655626706d2905e82}}, {{cite:15b5e42ccf8b995c1432cdff9ea741c29da49a94}}, {{cite:54a05300da147b71c56e297e5d7583690e46a1ff}}. Besides, {{cite:e82a6259a390949c330e1e8821483df95217d817}} and {{cite:559eac8624de2ab7f4e27fb461430fd58937a19e}} utilized adversarial training, while {{cite:ec150c7bc1c260b9ac4bfe6321c5142e800929bf}}, {{cite:546340b141c5cb15566642fac7e21ebd53de95a4}}, {{cite:35d2026c889a6630f0e1d3f2119189f854ad223e}}, {{cite:658bc023204f717c436b773ecbc17568c8f51e56}}, {{cite:70128bdd9e7e71927730dac303d7093c42da7e9a}}, {{cite:81f9b07c389e361e5f4bd07d1b63b82ff3845dc9}} and {{cite:f273bff79a3da80151c477c4e8dad1def449fd98}} proposed some special network structures and loss functions for better generalization. Although the effects were improved from different aspects, all these methods intended to make the feature extraction backbone generalized. Nevertheless, mapping all real faces and attacks of different domains to a shared feature space is difficult and the shared feature space is usually not generalized well. Furthermore, in the view of uncertainty estimation, given some inputs from unknown domains, these models might produce surprisingly bad results because of incapacity to distinguish what they know and what they don't.", "Representation Disentanglement-Based Face Anti-Spoofing. Some other representative methods for generalization realized FAS through detecting spoof traces which were disentangled from input faces. {{cite:91d3a019df218eca92bb258de366f9aba00208bf}} would like to synthesize and identify noise patterns from seen and unseen medium/sensor combinations, and then benefited from adversarial training. {{cite:190ea0f014bf4ab6ede1485814b8631009d2e717}} intended to disentangle spoof traces via arithmetic operations, adversarial training and so on. Especially, motivated by CycleGAN {{cite:7f41f97635fd80aaed68249ab9112851976a64b6}}, {{cite:7198cc93618233e6a2024a477cd4954b5db20671}} finished disentangling liveness features of real faces and attacks by exchanging, reclassification and adversarial training. Generally speaking, these methods successfully disentangled liveness-related features and thus had relatively stronger interpretability.However, the set of attacks is an open set and it is almost impossible to construct the distribution of attacks or disentangle diverse spoof traces accurately.{{figure:3bf45dbb-4c73-4adf-a095-05c342030c89}}", "Bayesian Deep Learning. Bayesian deep learning is one of the commonly used uncertainty estimation methods, which can capture both epistemic uncertainty and aleatoric uncertainty. The network combined with Bayesian deep learning is referred to Bayesian Neural Network (BNN) {{cite:696e2023be16d90b824351e0f4643041f41311ae}}, {{cite:9d2a1c32a738451bcc565bd400aada8c4c6315db}}. Later, several approximation approaches, such as variational inference and MC dropout, were presented to model the uncertainty {{cite:f7269399742953c96c2265e1f8a09c06726d72eb}}, {{cite:b1ab54db8458b11a39b19b3968d0480608fb8f83}}, {{cite:52661e7f25ff2423fee5c64b6008993bce54d42f}}, {{cite:6ef10ec183fb0e19cd5beecea55f876c3ebbeca3}}. Recently, methods incorporated with Bayesian deep learning have been applied to various fields, such as recommender systems {{cite:bd68a0e7ba9291508175a3ea50840924df5cc602}}, object tracking {{cite:9f268544b349144df1a8212fab2e7fd41bd91273}}, health care {{cite:c4ac492820d7411074917abc4962c04faef7a85d}} and salient object detection {{cite:afe601d85a26605bf6882a6b81c8d22de9f8c5f7}}. Whereas, there are rare applications of Bayesian deep learning to face anti-spoofing." ]

[]

[ "One of the most common techniques to find cognates is based on the manual design of rules describing how orthography of a borrowed word should change, once it has been introduced into the other language. koehn2000estimating expand a list of English-German cognate words by applying well-established transformation rules. They also noted that the accuracy of their algorithm increased proportionally with the length of the word since the accidental coexistence of two words with the same spelling with different meanings (we identify them as `false friends') decreases the accuracy.", "Most previous studies on automatic cognate identification do not investigate Indian languages. Most of the Indian languages borrow cognates or “loan words” from Sanskrit. Indian languages like Hindi, Bengali, Sinhala, Oriya and Dravidian languages like Malayalam, Tamil, Telugu, and Kannada borrow many words from Sanskrit. Although recently, kanojia2019cognate perform cognate detection for a few Indian languages, but report results with manual verification of their output. Identification of cognates for improving IR has already been explored for Indian languages {{cite:d6b67976de649b8de3e8947902e4b76b2e4316d2}}. String similarity-based methods are often used as baseline methods for cognate detection and the most commonly used among them is Edit distance based similarity measure. It is used as the baseline in the early cognate detection papers {{cite:ab6d2b1a47a4363237de15e56682487591d47506}}. Essentially, it computes the number of operations required to transform from source to target cognate.", "Research in automatic cognate detection using phonetic aspects involves computation of similarity by decomposing phonetically transcribed words {{cite:9202322fb5813d438e784fb9ada123f6c1c3be37}}, acoustic models {{cite:6f093282ed6fac7b3897c25d2c00b3885352c7cc}}, phonetic encodings {{cite:0d632e315f30da42e7329a1893dcee9eeefb42fc}}, aligned segments of transcribed phonemes {{cite:373d9a74b039a0f5c464a472dcfbf4231a3da482}}. We study rama2016siamese's research, which employs a Siamese convolutional neural network to learn the phonetic features jointly with language relatedness for cognate identification, which was achieved through phoneme encodings. Although it performs well on the accuracy, it shows poor results with MRR. 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:df6047031e495b90b2be946596d029a3f368fecc}}, {{cite:4885745578fe2c6a6f9a3d290776540a271a6d37}} or HMM {{cite:68c66f0481af32658ad6aac43524f52491b01984}}. We also consider the method of ciobanu2014automatic, which employs dynamic programming based methods for sequence alignment. Among cognate sets common overlap set measures like set intersection, Jaccard {{cite:b0bab1b81f3012bf2ecce378df5fb57086fb7abc}}, XDice {{cite:2a5b127b20375c1fc54699e86e2989a16f13b871}} or TF-IDF {{cite:84d735853d7df17be197e86d4caf764af5f990af}} could be used to measure similarities and validate the members of the set." ]

[]

[ "With the advancement of computer vision and natural language processing, many works exploit to combine radiology images and free-text for automatically generating reports to assist radiologists in the clinic {{cite:f036ee926698e0c9215af941504a9d9bb77a7fd5}}. Inspired by image captioning, {{cite:4c7147767bf2c1cd9b3038e9d9accf7b961eb37b}} adopts the CNN-RNN framework to describe the detected diseases based on visual features on a chest x-ray dataset. The work is restricted to the categories of predefined diseases. The Co-Att {{cite:d8e4d087dd0d419db6854671363a4ecfa2e019e2}} and {{cite:2c92495dd4bf3f6ab908ac2302d77df14071eeac}}, {{cite:c6b701c9b46d278a023b1f0f1783c284cf0717d2}} propose different attention mechanisms and hierarchical LSTM to generate radiology reports. However, most reports generated by such works tend to describe normal observations, which indicates that such methods have difficulties in capturing subtle changes in the image. Our proposed method is similar to the TieNet {{cite:20f1b1ed63efcd77fc70cf552785a6b83961bd1e}} which proposes an attention encoded text embedding and a saliency weighted global average pooling to boost the image classification and report generation. However, the method uses the report embeddings as a part of LSTM's input which is not available in the inference stage, resulting in embedding bias in the report generation.Other types of work explore injecting extra prior knowledge into the generation model to improve the quality of the generated radiology reports. Following the writing practice of radiologists, the HRGR {{cite:c5112c98d183e0f292d92cb8d2f6d7df66511596}} compiles a manually extracted template database to generate radiology reports by reinforcement learning. The KERP{{cite:c1eb6fe45b293a1fa2f9514aeaf80339df4ac37a}}, MKG {{cite:5cbf8f377c481fcaef87c117d10faeec0ad8139a}} and PPKED {{cite:239707b6ad79cc0a855d7c161f1ba0940397df52}} propose to combine pre-constructed knowledge graph for radiology report generation. Although these methods achieved remarkable results, building the template database and knowledge graph is still laborious, making it hard to transfer those approaches directly to other datasets. In this paper, we propose a knowledge-enhancing generator without manual labor to build a knowledge base.", "As revealed by recent studies {{cite:d7d6e3d505ad21457300330d63f346d0a25b0057}}, {{cite:32d02d1ab5a7ed233e64f8c76027bfb06952b6cf}}, {{cite:797b2494503dd2d6424a50dfcb6252d24d87c048}}, the memory mechanism can provide prior knowledge to boost the generation model. The R2Gen {{cite:56e20b676f332b134c87a4ee49946b5c6b62aefa}} proposes a relational memory in the decoder to learn the order of words or sentences. Nevertheless, the order relies on the radiologist's personal style. The KVMN {{cite:d3eb707810578b79d63a05a2b989c02a05042b11}} proposes a key-value structured memory that automatically encodes prior knowledge into the memory for question and answer tasks (QA). Our work is different from KVMN in three aspects. First, KVMN deals with a different task (QA) which is a classification problem. As a result, their framework cannot be directly applied to medical report generation. Second, the knowledge source of KVMN is pre-defined by a manually prepared database, while our knowledge base is learned from scratch during training. Third, we propose a multi-head updating mechanism that has more expressive ability than the KVMN model.{{figure:c39337d9-5da0-4e2e-bced-5461e114e229}}{{figure:f916c9d9-98ac-4c5d-99dc-af410b8d1257}}" ]

[]

[ "As a large quantity of labeled data is not always available for training deep learning models, transfer learning becomes quite important for many of NLP problems.With transfer learning, widely available unlabeled text corpora containing rich semantic and syntactic information can be leveraged for learning language models, such as BERT {{cite:0b36449e8a0a41ce448b21105522407a8782ff2e}}, GPT {{cite:25515bcdafd4af074fda8e87e2d450054dd84f70}}, and T5 {{cite:f08e21879fa0362838961d01c7400e01fd5e54fd}}.Then, these language models are fine-tuned on downstream tasks, which is the dominant transfer learning method adopted in NLP at the moment.The second way of using transfer learning in NLP is to further pre-train pre-trained language models in domain data before fine-tuning on downstream tasks {{cite:df35b19ca1b3723e94051ac50db772c70c3fde10}}, {{cite:419d86eb6b24126a4451e11fa38cb97c03c7af4d}}.The third approach, which is the method we investigate in our work, is to transfer models fine-tuned on an intermediate task for a target task {{cite:b28a871fc484debdb0ed518604e309fbbde2fc89}}.", "A recent work {{cite:b28a871fc484debdb0ed518604e309fbbde2fc89}} investigated when and why intermediate-task training is beneficial for a given target task.They experimented with 11 intermediate tasks and 10 target tasks, and find that intermediate tasks requiring high-level inference and reasoning abilities tend to work best, such as natural language inference tasks and QA tasks.Another recent work {{cite:134dc7f0c3ebfa295bc834dcb64cdc99511bebe2}} has explored transferability across three types of tasks, namely text classification/regression, question answering and sequence labeling.They found that transfer learning is more beneficial for low-data source tasks and also found that data size, task and domain similarity, and task complexity all can affect transferability.{{table:8f3d7234-d54e-4bf4-a7f9-ab6fe3db3612}}" ]

[]

[ "Both monocular depth estimation (MDE) and MHE are geometry-related regression tasks; the former motivates the development of the latter to some extent. In this section, related works on MDE are firstly investigated and then MHE is introduced.", "Monocular Depth Estimation. Early works on MDE utilized hand-crafted visual features and probabilistic graphical models (PGMs) to encode depth-specific visual cues, including object size and texture density, based on a strong geometric assumption {{cite:fdb6303ea4cb42753125d6f5eef3c3bb5d0d2ab1}}, {{cite:88c529041fa8c3ce148bb7b95553bbd9c394f234}}, {{cite:505d054858b50a7bde54b55daae153c6884a7ae4}}. Recently, deep learning-based methods have dominated this field because of their powerful feature representation capacity. There are roughly two types of deep learning-based MDE, supervised methods {{cite:b2eb1240d6efec5b1250a5e2411c266ce62de57b}}, {{cite:01417b4b021f75677ad9e88da33cb96e45d9759f}}, {{cite:0acf5d584007dfe6247ce1bc057c76e14f662068}}, {{cite:6164be5416b5b6e28689d2b463e68cedf6521b8d}} and self-supervised (unsupervised) methods {{cite:08e8262dd43dcf57a4e59c2463e67f46e04d7ec6}}, {{cite:be7dafe5296b0da34da903e8e771058152bd5f60}}, {{cite:94deef39efce0b8cd7164cf135e1f6c188486351}}, {{cite:a51e58b0e601eb14a160611eab1c4e2fc94f2201}}, {{cite:dd63eed5855de4430a7889bb48bdc191c2e119b2}}, {{cite:dd63eed5855de4430a7889bb48bdc191c2e119b2}}, {{cite:080f3b5c3f68fb5f49445acb8ee049b80111c735}}, {{cite:d66280462388acb9ead91a18b30cfd5736232c75}}, {{cite:ed11ac52c60588a64ab4cf131af41714a868cf56}}, {{cite:ea55ba828d59d41c1508bb450559852ec1320f8a}}, {{cite:99ee6d6b4ed4e806b98447f3a0569408147ecef4}}, {{cite:7ac7f6055e192249c113fe027bd36c62ca129617}}, {{cite:80737cd5a46f85cdd7c7d65d6be582556525eb5c}}.", "In this work, we mainly introduce the supervised methods, which take a single image as input and generate a pixel-wise depth prediction map, following the standard supervised-learning workflow with the need for manual-annotated depth labels. These methods achieve state-of-the-art performance by making breakthroughs in innovative architecture designs, effective incorporation of geometric and semantic constraints, and novel objective functions. Some enlightening works {{cite:378a81453b48aed573bffaa010c61a17b3d9bcd2}}, {{cite:939adddafb1916b59003df1867e0b595a5f2c868}}, {{cite:b2eb1240d6efec5b1250a5e2411c266ce62de57b}} have applied deep convolutional neural network (CNN) architectures to MDE, directly estimating depths from single monocular images in an end-to-end trainable manner, and achieving impressive performance. To model the semantic and geometric structure of objects within a scene, some work {{cite:8f6a9c42e33309966453d505d867f8cbc48674c5}}, {{cite:1cb0737bb9d4427c54e1774e207078fa3805e01d}}, {{cite:30be3cc6eefd53f38ee838cd3dee2fd3a0bc0935}} has introduced semantic segmentation into MDEs as an auxiliary task, which can guide depth estimation at the object level.", "Taking into account the imbalanced depth distribution that restricts model performance, Jiao et al. {{cite:0acf5d584007dfe6247ce1bc057c76e14f662068}} presented an attention-based distance-related loss to concern more distant depth regions. Lee et al. {{cite:851f6b2033b07e86538dcbaab25d6009277f8f82}} combined multiple loss terms adaptively to train a monocular depth estimator from a constructed loss function space containing many loss terms. To balance the coverage speed of these losses, a loss-aware adaptive rebalancing algorithm was further designed in the course of training. The work most closely related to ours is {{cite:82a349f6f5dc7329baf48238c82890cdf95619e0}}, in which a robust training objective is designed to train deep learning-based MDE models using multiple mixing datasets. For the first time, they propose to evaluate MDE models in a zero-shot cross-dataset transfer setting. More recently, vision Transformer-based deep models {{cite:c78b31a3ba2b9491cf59c6c6ee45775164f823bd}}, {{cite:096dd39ff74e88afea39c147eed4aa71914d34d0}} have also been proposed, to take advantage of the powerful representation learning ability of the Transformer backbone.", "Monocular Height Estimation.Motivated by the success of deep learning-based MDE, researchers have attempted to directly predict the height of objects, i.e., the digital surface model (DSM), within single aerial images from an overhead view {{cite:b64908948ddb6d16b1206e1f309336d74f0f1144}}, {{cite:260550c43b71e6abf5f480459f2c9eb2591eac5a}}, {{cite:1c90aab7f75658718670fff826c6745d8113ccb5}}, {{cite:10815239df7ff47f685aaf03df9c95064acb1140}}, {{cite:9c879040407ba4d4f68955f80d2b33f3d255474b}}, {{cite:a6c04085592c7cf97175d5cec4b785845c396b5b}}, {{cite:da46e5d0a48b65cf08428c41d3288751ae3fdc82}}, {{cite:4374af50a5f8519b521f8928cc2bc4ce4e8346f7}}, {{cite:bb581232d270a1b9a371334c4c3a2a73af9fe2a1}}, {{cite:bb581232d270a1b9a371334c4c3a2a73af9fe2a1}}, {{cite:72aa2c0c3fdd12f88e174487d50f4c76ff016ae6}}, {{cite:404f2fc2d29a9636b8af85d9f6751cf723dd35ae}}, {{cite:b1f68e2279ae2aaf5748575a2baf129dc1fba8f7}}, {{cite:219de2ba82f7cb8e3eb5884a9991993a91eaf272}}.Srivastava et al. {{cite:b64908948ddb6d16b1206e1f309336d74f0f1144}} proposed a multi-task CNN architecture for joint height estimation and semantic segmentation in monocular aerial images in 2017 for the first time.Mou and Zhu {{cite:260550c43b71e6abf5f480459f2c9eb2591eac5a}} published a concurrent work that proposed a fully convolutional-deconvolutional network for MHE and demonstrated its usefulness for instance segmentation of buildings. Paoletti et al. {{cite:1c90aab7f75658718670fff826c6745d8113ccb5}}, {{cite:10815239df7ff47f685aaf03df9c95064acb1140}} performed the image-to-image translation from monocular optical images to the corresponding depth maps within three cities, using the technique of generative adversarial network (GAN). Amini and Arefi {{cite:9c879040407ba4d4f68955f80d2b33f3d255474b}} presented a CNN-based method to identify collapsed buildings after an earthquake, based on pre-event and post-event satellite images as well as airborne LiDAR data. Based on the CNN architecture, they further designed a post-processing approach to merge multiple predicted height image patches into a seamless continuous height map {{cite:a6c04085592c7cf97175d5cec4b785845c396b5b}}. Liu et al. {{cite:da46e5d0a48b65cf08428c41d3288751ae3fdc82}} proposed a joint framework called height-embedding context reassembly network (HECR-Net) to simultaneously predict semantic labels and height maps from single aerial images, by distilling height-aware embeddings implicitly.", "Leveraging the optical flow prediction technique, Chridtie et al. {{cite:4374af50a5f8519b521f8928cc2bc4ce4e8346f7}} developed an encoding strategy of the universal geocentric pose of objects within static monocular aerial images, and trained a deep network to compute the dense representation; these attributes were exploited to rectify oblique images to dramatically improve the accuracy of height prediction of multiple images taken from different oblique viewpoints.Mahmud et al. {{cite:bb581232d270a1b9a371334c4c3a2a73af9fe2a1}} proposed a boundary-aware multi-task deep-learning based architecture for fast 3D building modeling from single overhead images, by jointly learning a modified signed distance function, a dense height map, and scene semantics from building boundaries in order to model the buildings within the scenes.Madhuanand et al. {{cite:72aa2c0c3fdd12f88e174487d50f4c76ff016ae6}} aimed to estimate depth from a single Unmanned Aerial Vehicle (UAV) aerial image, by designing a self-supervised learning approach named Self-supervised Monocular Depth Estimation that does not need any information other than images.", "Although these works have contributed to the development of MHE in the past few years, most of them stayed within a particular comfort zone. There is an urgent need to study some more crucial issues that restrict the practical application of MHE in the open real world, such as the exploration in zero-shot and few-shot knowledge transfer in a cross-dataset setting, and scale-adaptive MHE model design. Aiming to address these problems, this paper conducts the corresponding research and exploration." ]

[]

[ "Several studies have questioned whether DP is valid for correlated data. {{cite:25f4c2316d336cd99f776076f5e16a0ae8788beb}} was the first to raise the issue of privacy degradation when DP is applied over a dataset with correlated tuples. To this end, existing solutions that try to handle the correlation between tuples in the datasets can be categorized into two types, by considering: 1) the dependency between different tuples (i.e., individual-individual correlations), and ii) the dependency among single individual’s data at different time-series (streams) entries (i.e., temporal correlations).", "First, to handle the individual-individual correlations (or vertical correlations) between tuples, Group DP {{cite:5be4196fd159c6696a95e7cd5e1d3cd9e719f4e5}} is one of the first studies, which proposes adding noise proportional to the size of the largest correlated tuples in the dataset. Their method adds a tremendous amount of noise (i.e., {{formula:907d7759-a46a-4af1-816c-d31e1c79d566}} noise to a dataset with {{formula:4cefef9b-9acc-4348-a78c-738dab6a3028}} dependent tuples), thus destroying the data utility. As a generalization of DP, {{cite:2ee06cf145b81c73b67bc07d19842dfbfc69c34c}} proposes another general and customizable method called Pufferfish to handle the dependent tuples by adjusting the Laplace scale, however, the main challenge of Pufferfish is the lack of suitable mechanisms to achieve the expected privacy guarantees. Following this general approach of Pufferfish, the baseline approach proposed by {{cite:06c6fb55ff321aea231197e95a676ed0362bc582}} tries to handle the correlation by multiplying the original sensitivity of the query with the number of correlated records {{formula:b30c212b-ee08-41e8-bf3c-1b5737a972a1}} (i.e., query sensitivity = {{formula:4199798a-2cf8-402e-87fe-a05e6feaa172}} {{formula:55a71d0c-0161-4de8-ad30-5302385c48a8}} query original sensitivity). Bayesian DP {{cite:20d50b97ae931b345a5c955acf19a5e4fc76f78f}} uses a modification of Pufferfish, but it only focuses on modeling the tuples correlation by the Gaussian Markov Random Fields. All the following studies such as {{cite:bf656fe693009533e08e790d16860b2d3f7b38f5}}, {{cite:111969984aedbf6867232b3c8e3d5299760b00a0}}, {{cite:1be239c6ac5cf076cf49c752e76b239ee87e8798}} are trying to adjust the sensitivity by introducing dependence coefficients according to the number of correlated data, considering the pairwise correlation between dataset tuples as in {{cite:bf656fe693009533e08e790d16860b2d3f7b38f5}} or using heuristic analysis (empirically-computed query sensitivity) as in {{cite:1be239c6ac5cf076cf49c752e76b239ee87e8798}}.", "Following the second setting to handle the temporal correlations, {{cite:d491ee9ecd1e5121de13346115fdbcbf74af545a}}, {{cite:6dec021c6b7d44fe3dda1bd93ece4b0c556afa27}} propose sharing statistics and counts of a data stream considering horizontal correlations. In {{cite:d491ee9ecd1e5121de13346115fdbcbf74af545a}}, they propose two algorithms for the Wasserstein mechanism and the Markov Quilt mechanism when the correlations can be modeled by Bayesian Network. {{cite:40f0652bbba3fc7ac1089de405e5394c9b5c929e}} also considers the temporal correlation which can be modeled by a Markov Chain.", "In Section REF , we compare our model (in terms of privacy) with the existing similar approaches from the two aforementioned categories {{cite:bf656fe693009533e08e790d16860b2d3f7b38f5}}, {{cite:1be239c6ac5cf076cf49c752e76b239ee87e8798}}, {{cite:d491ee9ecd1e5121de13346115fdbcbf74af545a}}, {{cite:5be4196fd159c6696a95e7cd5e1d3cd9e719f4e5}}. Since Hidden Markov would not work to model statistical genomic dataset, we are not comparing our model with the mechanisms proposing hidden Markov-based models {{cite:20d50b97ae931b345a5c955acf19a5e4fc76f78f}}, {{cite:d491ee9ecd1e5121de13346115fdbcbf74af545a}}, {{cite:6dec021c6b7d44fe3dda1bd93ece4b0c556afa27}}.{{figure:04a2fc85-67be-4676-9ead-8cc25b6e8862}}" ]

[]

[ "Various publications, such as {{cite:22ea4ec21334ef8ff996a92905255ad750aa25a6}}, {{cite:7e415dac0002d9b8162c0e1bd61fbe7298038ed1}}, {{cite:24330ccd6d5f7c107c734fbc35c315f67fe2ef7c}}, {{cite:579f3a06829b9f4d632019783e08ddead78c0739}}, {{cite:2aa04ca809993242d3e8031fcec6b940812cc24a}}, {{cite:b55d7970cbad3b587ec2766e29c15dee33cdfec2}}, suggested different frameworks to theoretically study various multi-task learning settings. However, all of these works consider learning settings in which the learning algorithm is provided with a sequence of learning problems, and therefore, are unable to characterize the case where only one source task is available.", "In contrast, we propose a theoretical framework in which the learning algorithm is provided with one classification task and is required to learn a representation of the data that can be adapted to solve new unseen tasks using few samples. This kind of modeling is aligned with the common practice of transfer learning using deep neural networks.", "While in unsupervised domain adaptation (DA) {{cite:d4b6885b8a72fcc5d3f33ebf96e019abddb28dee}}, {{cite:fe0d7d4a85a6a7f2c3c334157e6d595e430a4f5f}} one typically has access to a single source and a single target task, one does not have access to labels from the former, making the problem ill-posed by nature and shaping the algorithms to behave differently than those that are provided with labeled samples from the target task. In particular, although in transfer learning we typically train the feature map on the source dataset, independently of the target task, in DA this would significantly limit the algorithm's ability to solve the target task." ]

[]

[ "The barren plateaus phenomenon of QNNs is first reported in random ansatzes {{cite:8e324e4c75529e4589cb7620af682c742993f7c1}}. Specifically, for {{formula:8d1e1c10-b32a-4a1c-90e6-f4380b56da3a}} -qubit random quantum circuits with global 2-design distributed unitaries, the expectation of the derivative to the objective function is zero, and the variance of the derivative vanishes to zero with the rate exponential in the number of qubits. Given the fact that random circuits form approximate 2-designs {{cite:9141f76dd9e51cedc991ea3f5be7002ebbdae0f0}}, QNNs with random ansatzes are essentially untrainable.Subsequently, Refs. {{cite:e2c1723a8656ffbef68d454e96796f11697c8128}}, {{cite:fae9e4f96aff868b1b0fe71a3e16b84f15e6f332}} relax the global 2-design assumption to the local case and prove the trainability of shallow alternating layered circuits with local observables. Their results imply that in the low-depth {{formula:77e209ee-0d16-4581-a20c-2ff5b1de8dd2}} case, the norm of the gradient could be {{formula:97500b6a-0ff7-46a3-b37d-80bde819db06}} , which is better than previous exponential vanishing results. Similar results hold for the quantum convolutional neural network (QCNN) {{cite:7ed82f0137aa1ac0da61e06ea8a822d4d4a6aa52}} which adopts a tree tensor structure {{cite:f8fa13919884ba30914dbd0e5e4aafef82a4cd55}} with {{formula:4ee0259f-659e-4cc0-b792-0ec3f7915913}} depth.Besides, most existing works show a negative relationship between trainability and circuit complexity. Specifically, states generated from QNNs, which satisfy the volume law {{cite:35a5c35b464277716f19bd0b95845d97ead5ee91}}, could lead to the barren plateaus landscape.QNNs with high expressibility {{cite:77971bd3dbc40505a17beded72762b0f82ad5ded}}, {{cite:b193959ef3c75bfad6e28a844a102e7bc5eb2b02}} could also suffer from the vanishing gradient problem.For the depolarizing channel case, noisy QNNs with linear depth could have an exponentially small gradient {{cite:fd1f3a032b7d2391e1f4e21bdd303b94eae6c90e}}.The vanishing gradient phenomenon is inherently related to the concentration nature of measure for quantum states {{cite:e5446a60dd2d6fc87e9191fdd18e387a2bcbe433}}, which cannot be solved using gradient free {{cite:6746dbd72df5440a3e548ed4b661ddff8bf34e26}} or higher order {{cite:a503bb3b421afbb85f61ed63926582cad975cb92}} methods. We summarize existing theoretical results about barren plateaus problem in Table REF .", "Recently, some techniques have been proposed to address the barren plateaus problem, including the special initialization strategy {{cite:972dd0490d63666182cd31497cf911e26f52e778}} and the layerwise training method {{cite:88f0864d8fe434d85d9f1a34fb2ba48d9bca435b}}. We remark that these techniques rely on the assumption of low-depth quantum circuits. Specifically, Ref. {{cite:972dd0490d63666182cd31497cf911e26f52e778}} initializes parameters such that theinitial quantum circuit is equivalent to an identity matrix ({{formula:594ba47c-514c-41ea-8bf1-9bc00c03740f}} ). Ref. {{cite:88f0864d8fe434d85d9f1a34fb2ba48d9bca435b}} trains subsets of parameters for each layer, so that a low-depth circuit is optimized during the training.", "Notice that almost all the theoretical analysis on the trainability of QNNs is based on the assumption that the learning circuits satisfy the unitary 2-designs or the Haar distributions. However, QNNs in most scenarios are tuned in the parameter space consisting of rotation angles followed by CNOT or CZ gates, instead of the entire unitary spaceAlthough the assumption of a uniform distribution in the parameter space could generate approximate 2-designs for certain cases {{cite:1f6b8f2530ac19c525a59a3c821ca1740a96e864}} with linear depths, the corresponding analysis requires that all gates are tunable, which does not hold for circuits employing CNOT or CZ gates..Thus, earlier claims that deep QNNs are doomed to fail are not on solid ground. In the work, we provide a first analysis of the gradient behaviour of the quantum neural network within the parameter space. Most importantly, we theoretically prove the existence of deep QNNs whose gradient norm is independent of the circuit depth {{formula:d84eebab-ce5e-41b8-88d6-7ccfb8ad2f08}} .", "The rest of this paper is organized as follows. Preliminaries of QNNs are introduced in Section . The theoretical results are presented in Section . We show numerical results in Section , which includes finding the ground state of the Ising model (Section REF ) and the binary classification (Section REF ). We summarize our conclusion in Section .{{figure:6233530c-338d-4f49-bbd7-b2f08a91d5bc}}" ]

[]

[ "With the advancement of computer vision and natural language processing, many works have been exploited to combine radiology images and clinical information for automatically generating reports to assist radiologists. Inspired by image captioning, {{cite:5726d2d74c67c4d207f63aba809a561d41b2b6f0}} adopts the CNN-RNN framework to describe the detected diseases based on visual features on a chest X-ray dataset. The work is restricted to the categories of pre-defined diseases. The Co-Att {{cite:db9b9b344f4a809f6087a9172f4e667ea7a94bfe}} and {{cite:d538e54b608b2446a9f72b222f7a4a8bacc321c9}}, {{cite:2934369c80cdefd12bc99db7ea28e71237d6760f}} propose different attention mechanisms and hierarchical LSTM to generate radiology reports. However, most reports generated by such works tend to describe normal observations, which indicates that such methods have difficulties capturing subtle changes in the image. {{cite:de81f46a9489b27f0c7b39533a7a13ed85cb4432}} proposes a contrastive learning mechanism to capture and depict the changes by comparing the input radiology image with known normal radiology images. However, the performance of this method is restricted by the quality of the normal input images. {{cite:e6d0432573fda892fd4ec42ec647c80491ab53ac}} introduces a curriculum learning in medical report generation to help existing models better utilize the limited data and alleviate data bias. {{cite:c5c75a7a723e828b9fd166d3d2d8de36be13c800}} proposes the AlignTransformer framework to alleviate the data bias problem and model the long sequence for radiology report generation. {{cite:81d0cd2e91a64122baeede8e186b8a89a8e791e5}} proposes a two-stage radiology generation model by incorporating high-level concepts. Nevertheless, it needs extra labour to extract high-level context to guide the first stage generation.", "Other types of work are also explored, which inject extra prior knowledge into the generation model to improve the quality of the generated radiology reports. Following the writing habit of radiologists, the HRGR {{cite:3cea2416554a6472e8ce3cf80ef2bc7b50a5dd6c}} compiles a manually extracted template database to generate radiology reports by reinforcement learning. The KERP{{cite:201929ff82010a7401ed3eae039a41dc32fbd34e}}, MKG {{cite:7e017d63fd66e48b4af0473ef4830357a7870c91}} and PPKED {{cite:b836727ee6254f1c12251ff33cc09cc4d5d065b2}} propose to combine pre-constructed knowledge graph for radiology report generation. Although these methods achieved remarkable performance improvement, building the template database and knowledge graph is still laborious, making it hard to transfer those approaches directly to other datasets. The R2Gen {{cite:307606d477db4ab3d4af6688b72a22f1681941ec}} proposes a relational memory in the decoder to learn the order of words or sentences. However, the order is not essential as different radiologists have different writing habits. This work proposes a knowledge-enhancing generator without manual labour to build a knowledge base. The procedure involves no manual labour and thus can be much more efficient. The relations of knowledge also avoid the bias caused by orders of the word.", "Our proposed method is similar to the TieNet {{cite:7c9fa8050cc614255ecad01f3f4a32e2d615446e}}, which proposes an attention encoded text embedding and a saliency weighted global average pooling to boost the image classification and report generation. However, TieNet uses report embeddings as a part of LSTM's input which is not available in the inference stage, resulting in an embedding bias problem in the report generation." ]

[]

[ "Turbulent fluid dynamics are ubiquitous across the universe, and turbulence simulation is a crucial step in problems across disciplines.These include forecasting problems, like predicting the spread of a wildfire {{cite:b2d383441adf9a9df79f96ac9af2abf35263a89b}} or the evolution of a hurricane {{cite:44e1014f0f8adec1a55e3f31a5021f47cc5201ce}},engineering challenges, like designing more aerodynamic machinery {{cite:322b4028d213b60429c9c93e9f8396cc2199dc52}},and scientific questions, like understanding the physics of insect flight {{cite:245ea9d04901fae1223a0882e6ac41c0dd18c797}} or dynamics of celestial objects {{cite:2c6b46f4a1ff4eda2403c579369bd531ddb8d7ca}}.Across applications, a variety of statistical measures are relevant for quantifying physical variables under turbulence (Velocity Autocorrelation, variable histograms, Cooling, Energy spectra).The primary challenge for simulating turbulence is accurately capturing small- and large-scale flow dynamics, and their interplay, simultaneously.", "Traditional approaches to computational fluid dynamics integrate the governing PDEs describing the fluid flow, given initial and boundary conditions.State-of-the-art turbulence simulators, such as Athena++, are built upon efficient algorithms that prioritize the preservation of high-order statistics.Such simulators are used frequently in scientific applications to numerically compare model predictions with experimental observations, but can be extremely–often prohibitively–expensive to run.Thus, there is significant practical interest in developing more efficient simulators.", "Deep learning models for predicting physical dynamics have traditionally been used for modeling low-dimensional dynamical systems with graph-like {{cite:53901d09e4ea35c2efe58d3383c6e17f2f088bed}}, {{cite:6775d4f3ba45613dfde6961cbbca9259db62cb3c}} and grid-like specifications {{cite:7ff1e083adb6222bf2bd76f92816afcd0d18c92d}}, {{cite:8606ad86a7ce2695567decff05f8f612a742a785}}, but there is a growing body of work for fine-grained modeling of complex dynamics, as in particle-based liquids and soft-body physics {{cite:7f97735a1d8a939e8e291a517be597bdd2ae47ae}}, {{cite:fd8c5c985d10fa0ec4f6f58191cfb2126d5a4619}}, {{cite:16bfcd936fdcc71fa410a303fe8933e38ea474b9}}.Learned models are often used in computer graphics to speed up {{cite:f1fc14ab37e84b98f55cbd16b5e87ff72550bf97}}, {{cite:444d682ec8d3adc7dc22e66cfe1200a2a0a0a3d6}}, {{cite:b95c57f6066c657ed9fb64cfe79f34f974350fec}}, {{cite:006b369218045342e16189cd26b4f3c54ff93fa7}} or super-resolve {{cite:efe3a93b7798a90008123982a70f8e9c1a273862}} simulations.A number of papers predict the steady-state flow around 2D airfoil wings, using CNN {{cite:136130534fadb07d5dfda528cfdd07d333e2e3f1}}, {{cite:14be5e4fba0018168d8f8dc0e892555fda995f3c}} or GCN {{cite:034ce4a4c16b63522ce5156e03466fd3f32f9720}} architectures, or the transient dynamics of various low Reynolds number flows using mesh representations {{cite:9570244f7a93b0d72a7db6b972cbf31f1954c96d}}.", "In recent years, a number of models have been introduced specifically for learning turbulent dynamics.Many of these use neural networks to model some part of an otherwise classical, theoretically motivated numerical integration setup {{cite:8352e327d23d38e29369ac4ef851583781d7b6da}}, {{cite:be1b4703a7abbf2b18a5cb5fd287531c8bfe30de}}, {{cite:a7bb876e17696cb0043ce4150b0322542b3b5284}}, {{cite:485b1aafa9e4290b4eb892eb3d92bffa5ced7a8c}}, {{cite:6bf220604d656a9fc08920d9a350ab2b86a477ad}}, {{cite:329d9bd1970af2dd250eecd1c2d838ae89de4528}}, PDE formulation {{cite:7ff1e083adb6222bf2bd76f92816afcd0d18c92d}}, {{cite:dd784d93c65e08d4651258a629f9514c8a8a7c79}}, or embedding space {{cite:56438c8633376464b0d73a77e59f59ff3966a73c}}.Other models incorporate some information about turbulence simulation into the architecture, but like ours, are learned end-to-end in a supervised way.Wang et al. {{cite:e6513498fece9d4b81ce57f3f222362eb9b6006e}} implemented a fully learned, neural network “Turbulent Flow Network” (TF-Net) that applies physically inspired decomposition with learned parameters and processes the components with U-Nets for fluid simulation.Li and colleagues introduced “Fourier Neural Operators” for turbulence simulation {{cite:a372e64bf60c06bb8ddc2b345ec2f58e9d8fd623}}, which apply neural networks that convolution parameters in the Fourier domain. This model is part of a broader family of work {{cite:4ac03e420659f9f1aa0ab4c9d31cb52800fc75eb}}, {{cite:f0392949436ccfa254a3d525f506fb6c23537466}} which combines basis transforms (Fourier, Multipole, etc) with neural operators that process the transformed input.We implement models from these papers as baselines to compare to our learned simulator." ]

[]

[ "Our work occurs within BO, a framework for global optimization of expensive-to-evaluate black-box functions that originated with the work of {{cite:c432ae60c2e8edaf415bfb934189929449acb951}} and {{cite:9dec7f85daceb5a13ea45316329b678ae6bfad18}}, and has recently become popular due to its remarkable performance in hyperparameter tuning of machine learning algorithms {{cite:893e965827f653e9c2e690986f4eca16880292c4}}, {{cite:a274bd6727e7df03f612a195dc9d678f71fdbab7}}, {{cite:97a950b6b6c9134241ed9735ef293cb7e6d1cf63}}. We refer the reader to {{cite:3987d3ec2c798e26eeb38b4e2e01d2a3940365fd}} and {{cite:971b3f5906f60b3d6c49241be5a8e938eb1b942c}} for modern introductions to BO.", "Our approach can be catalogued as a grey-box BO method since it does not treat the objective function entirely as a black box, and instead exploits known structure to improve sampling efficiency. Other examples of grey-box BO approaches include multi-fidelity BO {{cite:0b486aafa47d16f8daf3e8bd94ad8497f39e207b}}, {{cite:97a950b6b6c9134241ed9735ef293cb7e6d1cf63}}, which leverages cheaper approximations of the objective function; BO of objective functions that are the integral of an expensive-to-evaluate integrand {{cite:f876ec7fc2c509bf0c675158294aefaf23b99f6e}}, {{cite:37ec0e49ff3bea5d600382586737c3b4c875bd29}}, {{cite:c91d34629f98124e05400be1083a0d36d6a5797e}}; BO of objective functions that are a sum of squared errors {{cite:f32d1670b482d1b14bfc2e7915251949e7bd62b5}}; and, more generally, BO of objective functions that are the composition of an expensive-to-evaluate inner function and a known inexpensive-to-evaluate outer function {{cite:eb663bcc0c7c5040a452dfe2fafc0dd24ea5f928}}, of which our work can be seen as a significant generalization. We refer the reader to {{cite:67e63ad74dda3e91751ddfa16eeae8ddac2118a2}} for a tutorial on grey-box BO.", "Our work is also closely related to {{cite:ab0d7c8a555df3c1c9c1a2f7c629141763d58614}}, which proposes a method for efficient sequential experimental design of nested computer codes, also using GPs. This work focuses on the case where there are only two node functions, and one takes as input the output of the other. In contrast with our work, the goal of the proposed method is to learn the output code as accurately as possible within a limited evaluation budget, but optimization is not pursued.", "Optimization of composite (a.k.a nested) functions has also been considered in the gradient-based optimization literature {{cite:d7d72c756042949d50860ca0dc47e7ad4125f27a}}, {{cite:6768079c7d5fdc55b41bca190e9933246d01030f}}, {{cite:ae3db8d13764cf1d8d410d3e45c9638a1088e009}}, {{cite:5ca166ebeb3dcb90d8d230cfa01ce27acf72f897}}. In contrast with ours, these works assume that evaluations are inexpensive, and often also some form of convexity, along with availability of gradients. However, this literature is similar in spirit to ours in that information from inner functions improves efficiency.", "Function networks arise in many application areas.While these applications have not, to our knowledge, been previously formulated as specific instances of the general function network model we propose, their literatures are nonetheless relevant to our work.", "In engineering and aerospace design, function networks arise in multidisciplinary optimization {{cite:ebe9024193cf02956eee76086611900c48c3a925}}, {{cite:1216ad008034b9fd13fdcac4eded8159951bdf77}}, {{cite:326f8318c687efa2f13fd95cc70801d57a42354b}}, where simulators focusing on different physical laws are coupled into a function network. For example, a simulation of airflow over a wing may output the forces on the wing to another simulation of mechanical stress on the wing structure.In drug discovery and materials design, function networks arise from thesequence-structure-function {{cite:3ea9ae7eab4ae62386f632a80649fe216134159d}}and composition-structure-property{{cite:499d402dfaa214872efcaaf7406c9047c761ea19}}paradigms.Here, decision variables (composition, e.g., what fraction of what raw materials are used) determine the structure of the material(how the atoms combine together), which in turn determines how the material behaves (properties).Function networks arise in the design of queuing networks {{cite:44473e33a10d7bdf1c13a28acd6600f9df4e6051}}, {{cite:1be15d98d550dfddf2b24524ece99ba4e1686d77}}. This includes manufacturing systems {{cite:8a7146a7e0c9fd2c81f6d6f870274fb7ad792a40}}, where the partially-processed output of one workstation is input to the next workstation, the design of service systems {{cite:0128a39f0375bacf7f083bc87fb301e1447e9bdc}} such as hospitals and airport security checkpoints, and choosing traffic signal timings for a city's street network {{cite:9f5f4cd30d3d12414baff8923d9f509f38aea6bc}}.Finally, function networks arise in reinforcement learning {{cite:2a2561f8ce1f1d5d8bcef8258db1ead55b5fb190}} and Markov decision processes{{cite:e9e24a8e27079aaa8125731430ef7dd486d17b2a}},where the transition kernel transforms the state variable at the start of a time range to another state variable at the end of this range. This outputted state variable is the input to the next time range.§REF shows an example." ]

[]

[ "Work on restricted-space classes of parameterized problems was initiated by Cai et al. {{cite:aca97198dba956c4f73e7116f8c2f2727f14167c}}, who defined the classes {{formula:85a4e06b-ecc2-493c-8928-d07db135e95a}} and {{formula:5f5b964e-eeb5-45ea-96df-8dbc89e074d7}} , which we call {{formula:a942933c-acfb-4034-a986-03c0803f9aa6}} and {{formula:503f6e8d-937c-4e0a-84d7-b87a6174580b}} respectively. Among other things, they showed that {{formula:3fd6d2bf-d514-4739-8cfd-fe2491146569}} under the usual solution-size parameterization is in {{formula:468821b2-8a24-4268-b6a0-69429d0dd12d}} . Continuing this line of work, Flum and Grohe {{cite:ec3e5617b6ce65f883d7b650f2f6970eb36c20c0}} showed that the parameterized model-checking problem of first-order formulas on graphs of bounded degree is in {{formula:12612268-f567-4851-a205-3781d5c3a622}} . As a consequence, many standard parameterized graph problems are in {{formula:1b5a7015-329f-4c2f-ba4a-944b4f8998de}} when restricted to bounded-degree graphs. Then, Elberfeld et al. {{cite:e267aa19d6350ba823c2ab63f24563c625061a3d}} introduced two new restricted-space classes and gave completeness results for those classes, showing that there are problems fixed-parameter tractable (FPT) problems which are outside {{formula:2c7b2dfc-996e-46c6-aa08-3ea76146849c}} under the assumption that {{formula:8454c558-c63b-4cb1-91b3-83c3e195ac77}} . In particular, they identified {{formula:b6a52700-e50e-4ca4-81c9-316b32165454}} as one such problem. In contrast, {{formula:f9be00a9-7a5a-4769-8a89-db09518c0e09}} restricted to tournaments can be shown to be in {{formula:6e93dd4a-2c21-4c19-9851-f04d920649b6}} (Corollary REF ). Later work in this setting includes the paper of Chen and Müller {{cite:f2dbd82575c2d429728eeba1feef3fd099d5cfed}}, who studied additional complexity-theoretic aspects of restricted-space classes and showed that {{formula:6abbfe64-2b2f-4ea5-9e83-bed4eebc965c}} is in {{formula:7334aa6f-c9ff-4d94-911e-c2b666165964}} .", "{{formula:e4221053-c878-4c1b-b180-c04911719561}} , a generalization of {{formula:d96758e1-c24a-4d7d-afe0-36a504ba4320}} to hypergraphs was shown by Fafanie and Kratsch {{cite:138ef47cf8eb08e5d22554740d7fd4e52f1f2f30}} to be kernelizable in logarithmic space, which puts it in {{formula:e8b50033-b2fb-42a7-be08-0c78a2e32b96}} (see Lemma REF ). As a consequence they showed that various graph deletion problems where the target classes are characterized by finite forbidden sets are also in {{formula:0f527f45-0875-40c6-882d-9c577fc8c869}} . For related results, see {{cite:b8890f6c9b9dfa8a706d04cec4a038c500c104bc}}, {{cite:57444b8b60abf9bc913cc3fb37d2919fd090632a}} which give constant-time parallel kernelization algorithms for {{formula:6bd4f25f-09cb-474b-8651-427e8605c6cb}} ." ]

[]

[ "Generally speaking, when it comes to measuring BP, commonly, either a cuff-based approach is utilized, which provides upper arm BP measurements, or one resorts to cuff-less (possibly invasive) solutions. Upper arm BP monitoring can provide users with an indirect and non-continuous BP measurement technique by using an inflatable cuff and stethoscope. Such cuff-based methods suffer from several drawbacks including: (i) Being inconvenient and unhealthy especially in public places; (ii) Require proper training prior to utilization; (iii) Not being ideal for self-use and long-term monitoring of BP, and; (iv) Being incapable of providing continuous BP measurements {{cite:f6417bad2da9a0b7cba38805386da4b6cce19fb5}}. Cuff-less BP monitoring {{cite:44313380d8465eec95d58535ac6cda984ce374bb}}, on the other hand, eliminates the common uncomfortable factors associated with the former category and has the potential to continuously provide BP estimates without using any inflatable cuff.", "Recently, there has been a great surge of interest towards the goal of performing continuous BP monitoring via Physiologically Inspired Models {{cite:053e251fb462e3ce23a97dd825bd91fbe99bb0c5}}-{{cite:31a9f3c02f3644c6f6898cd270dfbe93d92a34e0}} in particular pulse transit time (PTT) and pulse arrival time (PAT). The PAT, sum of PTT and pre-ejection period, is considered as the main marker of BP for development of cuff-less BP estimation algorithms due to its simple measurement procedure. The PAT {{cite:cf08fed6d5bb82e21e223002328073be8b13037f}}, {{cite:88366a74d30292029be723f3f8d0e3204f5d28e0}} is defined as the time required for a heart beat to transfer to a body peripheral and has a tight relationship (correlation) to the BP. The existing correlation between PAT and BP, although well established, is highly non-linear depending on several uncertain factors varying across different individuals and over time {{cite:b60355b3b1b58a70b2b2e4e0c3d15e316db2a217}}, {{cite:c18425c66ca8e57baf8b825355a7bc97f2c319eb}}. Therefore, there has been different attempts {{cite:2d0ea7de22a28f10690fef27bfed69e801ba7a4f}}-{{cite:157b5287c748f5304e676918514781c603b20dfa}} to construct calibration models to account for such variations. In calibration-based models, the focus is on extracting meaningful features to be fed to processing and learning models. Such models, however, are applicable only for use in short intervals such as exercise tests. Existing methods for cuff-less and continuous BP estimation can be classified into the following two main categories:", "(i) Hand-crafted Regression-based Models: Models belonging to this category are developed by extracting hand-crafted features and exploiting various conventional BSP and ML algorithms such as decision trees (DT), support vector regression (SVR), shallow neural networks, and Bayesian linear regression (BLR) to name but a few. Typically, PPG and ECG signals are used jointly {{cite:f6417bad2da9a0b7cba38805386da4b6cce19fb5}}, {{cite:5654a1c8d0aeb9a357820cb3750e8088e14bbdd2}} to extract PAT features to construct a regression model for estimating the BP. In Reference {{cite:5853ea1c248816070eb337cbba16e3755ed0a299}}, for instance, a linear regression model (i.e., the SVR) is coupled with a radial basis function (RBF) kernel, and a single hidden layer neural network with a linear output to estimate the BP.", "The key drawback of the aforementioned conventional methods is that such models rely heavily on extraction of hand-crafted features, such as PAT, and directly map the given input into the target value while ignoring the critical temporal dependencies in the BP dynamics. This could be considered as the root of long-term inaccuracy of such models, and results in lack of robustness due to strong dependencies on the calibration parameters and the choice of hand-crafted features to describe the signal for subsequent regression {{cite:5a9b88ac23e0fb586f91c4ed9da454c37b7634a9}}. Lack of robustness translates into accuracy decay over time requiring frequent calibrations, especially for multi-day continuous BP prediction tasks. Robustness is essential for use of developed models in clinical settings, therefore, lack of robustness of existing solutions has limited their practical utilization.", "(ii) Deep Learning-based Models: While research works on hand-crafted and regression-based models are extensive, deep-learning based BP estimation {{cite:1df63d90c8931363387e143e0e813ed7186bfb42}}, {{cite:a8be8ccecf09208a365a01bf762c515572bf8601}} is still in its infancy. In deep-learning models, commonly, hand-crafted features (e.g., extracted PAT features) are fed to neural network models such as long short term memory (LSTM) models, recurrent neural networks (RNN), convolutional neural networks (CNN), or bidirectional RNN (BRNN). For instance, Reference {{cite:1df63d90c8931363387e143e0e813ed7186bfb42}} proposed to formulate BP prediction as a sequence learning problem, and proposed a deep RNN model, which is targeted for multi-day continuous BP prediction. The RNN model works with a set of seven representative hand-crafted features extracted from ECG and PPG signals. Another recent example is Reference {{cite:200d11f0e244bcaed2b86bc62f18075c3f7c7958}}, where the authors proposed a waveform-based Artificial neural network (ANN)-LSTM model. The model consists of a hierarchical structure where the lower hierarchy level uses ANNs to extract the required features from the ECG/PPG waveforms, while the upper hierarchy level uses stacked LSTM layers to learn the time domain variations of the features extracted in the lower hierarchy level.", "One can identify two key drawbacks associated with the limited deep learning-based models developed recently for BP estimation: (a) In most of the studies presented so far, before extracting the deep features, representative hand-crafted features of input signals are first selected/extracted, which are then used to train a deep neural network. In other words, such methods ignore the real potential of deep learning in utilizing the intrinsic features (deep features) of the input signals, and; (b) Lack of a benchmark dataset for evaluation and comparison of developed deep learning-based BP estimation algorithms. In other words, datasets used in recent literature are not unified and properly defined, which makes evaluations and comparison difficult. Commonly a subset of MIMIC-I or MIMIC-III databases from PhysioNet is used without providing details on the training, validation, and test sets rendering reproducibility and fair comparisons impossible. In some cases, results evaluated based on two different datasets are compared, which can not be considered as a fair base for evaluation purposes. Furthermore, commonly a limited number of subjects within specific age/gender group is used making it difficult to evaluate the generality of the obtained results across different characteristics.", "To alleviate the aforementioned identified shortcomings of existing continuous BP estimation solutions, next we present the proposed {{formula:3315c3d7-771d-4f32-9419-394ca434baa8}} architecture." ]

[]

[ "For cloud removal from satellite images, many different approaches have been proposed over the years. They can be broadly classified into the following categories:Traditional approaches which involve detection of clouds from images and using image inpainting techniques to fill the cloudy areas with the help of cloud-free images taken over different time intervals. In {{cite:28206411806035873b071abdf7be5d2dbfa94b52}}, a similar approach for removal of clouds from satellite images is presented. It involves usage of a threshold value to detect cloudy pixels from satellite images and replacing those pixels with the data from another images of the same location. Due to variations of patterns observed in clouds, selection of one definite threshold is a challenge.", "Generative Adversarial Network(GAN) based approaches: In {{cite:93474d0f962b078da7158e4f547f2fcfe5319902}}, authors have proposed a cyclic generative adversarial network, that eliminates the need of a paired training dataset and Synthetic Aperture Radar (SAR) images. They introduced a set of generator and discriminator which compete with each other until the generator can produce realistic cloud free images. Due to lack of paired cloudy-cloud free images they didn't have any quantitative results.", "Combination of multiple models based on deep learning techniques: In {{cite:df4a29877f0d3bd36a18e60810b45c68052371c9}}, authors have introduced two new paired datasets which they have created using real-world images available from European Space Agency satellite Sentinel-2 and a novel Spatiotemporal Generator Networks (STGAN) which utilizes RGB and IR data to determine relationships across multiple images over a location.", "In {{cite:86701073b5b843573355507af31cdfd68bafdbc2}}, SpaceEye, a first of its kind model which utilizes both multi-modal and multi-temporal satellite images. SpaceEye takes multi spectral and SAR images as input, and uses a combination of a coarse and refinement network to patch the missing cloud pixels in an image. The predictions generated by this model are validated using a discriminator network which further improves quality of SpaceEye. Authors also stated the applications of SpaceEye by conducting case-studies.", "In {{cite:e1caeb8317f9cd655855e1a4d8292900f7db4a0c}}, authors have been able to generate the entire image instead of only cloudy pixels by using a hierarchical combination of 3 neural networks - conditional Generative Adversarial Network (cGAN), Convolutional Long Short-Term Memory (ConvLSTM) and U-shaped Convolutional Neural Network. This method has shown encouraging results when compared with state-of-the-art techniques." ]

[]

[ "Self-supervised learning relates to the idea of extracting supervisory signals from the data.Methods including relative patch prediction {{cite:1cb2e9a4c0b8d218ddc7ead9b91dd4d6385c97c6}}, jigsaw puzzle {{cite:57f1a385342b262508172f1e3a35b06296295a70}}, rotation prediction {{cite:10a89fda350a6d83d7856ab7018b96cb42d92e6f}}, and colorization {{cite:dd4a6aee8f025262f6a4d2cd2c6c0f97d075a9ed}} propose manually crafted pretext tasks, that when optimized, can make a deep neural network learn useful representations without the need for human annotation.Most of these methods work with the same principle.They corrupt the input with stochastic random transformations and challenge the network to predict some property of the corrupted input.", "One such pretext task is instance discrimination {{cite:2ab9d903c16c37807e19cefda27e4045d5f74a91}}, {{cite:8e67fca68e4c060e5f7f426b3f44a3d2af83e9c9}}.It describes a classification task in which each image is treated as a unique class and, therefore, stochastic transformations of the same image, often called views, should belong to the same class.{{cite:2ab9d903c16c37807e19cefda27e4045d5f74a91}} proposed to optimize this task by learning a linear classifier where the number of output classes matches the number of observations in the dataset.Following, {{cite:8e67fca68e4c060e5f7f426b3f44a3d2af83e9c9}} proposed to use a Noise-Contrastive Estimation (NCE) approximation of the non-parametric softmax classifier that could scale to large datasets.Currently, popular contrastive learning {{cite:4a9021fe42f776f6dd3491009cb1cb0571988ab2}} methods rely on an NCE based loss function called InfoNCE {{cite:f11ee3ba25e0a4ca6d0c97987ecbcb730ab92b7b}}, {{cite:dfe23035cffc13c551fe966930b58f18d1073286}}.Recent work describes optimizing the InfoNCE loss through the lens of maximizing the MI between representations of the same image {{cite:860a9ef337a91aa6595c24ea5c7bbdb6cf3016ec}}, {{cite:a4352844f32059f480a0cf74c8aa867171be1ec9}}, {{cite:f46dfe70daadf47c90c2e8ddb12ed91d526b270a}}.InfoNCE characterizes an {{formula:4a886076-33c2-4493-a310-68248a43230f}} classification task where a pair of positive examples needs to be identified among the set of {{formula:9f3455b8-4098-4643-a325-50763d6ef650}} negatives.In practice, the success of InfoNCE requires a large number of negative examples.Nonetheless, since negatives are usually randomly sampled from the dataset, it often leads to a false-negative problem where representations from images of the same class are treated as negatives {{cite:e1537499c58e34080c612f275aab02ed3406aee1}}.{{cite:3819be5d9de975bb81e9e8135618c86990109c9b}} and {{cite:592e6f55603f6446717369e19c799b0bbebb8fc5}} presented MoCo, a contrastive learning framework that employs an additional momentum encoder to provide consistent instance representations for the InfoNCE loss.{{cite:e1cca318ade85ae7b28b47d879f87bc8bc284d4a}} presented SimCLR, a Siamese-based {{cite:49f5821f43db319ad07567d59536c368b643501a}} contrastive learning method trained with InfoNCE that relies on large batch sizes to draw a high number of negative samples.BYOL {{cite:fc311fd3882a16083f0cb8a0d1aab8c67bcbb69e}} proposes a framework that does not require negative samples and learns visual representations by approximating augmented views of the same data point using an {{formula:d9c06d5a-e5cb-4bd3-8b3d-2d039cddc7cf}} loss in the latent space.Unlike contrastive learning with instance discrimination that treats each image as an individual class, we seek to learn prototype vectors that act as anchors in the embedding space.We use these anchors as energy beacons for the images.Our goal is to use the energy distributions induced by the similarity of the images w.r.t. the prototypes to find representations that share similarities in the representation space.", "Recent work proposed clustering-based methods for deep unsupervised representation learning {{cite:79eb5eff09e0f1b326625510e53756e3c2038bea}}, {{cite:8f10d3a256310689d3ea22936d9b8a7218107294}}, {{cite:39cb4b3b97e27d7a9012b3433c224c2c16d57325}}, {{cite:18581de0a3a93f37827a15e718977ec667423c5a}}, {{cite:d64c264f255e1ff20d5b6a87a2a7842e64b49e5e}}.DeepCluster {{cite:a2b79c578b58e10d1a6cded3d6086beb826ff710}} learns representations by predicting cluster assignments.Once per epoch, the algorithm clusters the entire dataset using {{formula:7ecefbd6-155e-4afa-b262-6596ee5aa6eb}} -Means and uses the pseudo-labels as supervised signals in an additional classification task.One of the limitations of this approach is that the classification layer needs to be reinitialized once per clusterization.DeeperCluster {{cite:39cb4b3b97e27d7a9012b3433c224c2c16d57325}} builds on top of caron2018deep work and presents an algorithm to combine hierarchical clustering with unsupervised feature learning using the rotation prediction pretext task {{cite:10a89fda350a6d83d7856ab7018b96cb42d92e6f}}.Similarly, Prototypical Contrastive Learning (PCL) {{cite:18581de0a3a93f37827a15e718977ec667423c5a}}, formulates a self-supervised visual representation learning framework as an Expectation-Maximization algorithm.", "Our method utilizes a conceptually distinct methodology.Unlike caron2018deep and li2020prototypical works, our method does not require a pre-clusterization step of the entire corpus, which vastly reduces memory and computing power requirements.Moreover, since we do not use {{formula:8da2f298-a480-4ddd-ade8-37b9cb578181}} -Means clustering as a proxy to learn an additional task, we do not need to reinitialize any layers during optimization, nor is our method susceptible to limitations and assumptions implied by the {{formula:c277960b-e5d8-41c3-bd15-d407b46f4942}} -Means algorithm.Instead, we propose to learn the prototypes end-to-end by consistently enforcing different views of the same image to be assigned to the same prototypes.Lastly, our method does not rely on handcrafted pretext tasks.", "Similar to our work, {{cite:d64c264f255e1ff20d5b6a87a2a7842e64b49e5e}} proposes an online clustering method to learn visual representations by contrasting cluster assignments on a second set of latent variables.To avoid collapsing modes where all observations are assigned to a few classes, they use the Sinkhorn-Knopp algorithm {{cite:b8d6e3a146fc339ee46b1e396e774fcbd2c6af87}} to solve the cluster assignment problem over the latent variables to guide the encoder during learning.Unlike caron2020unsupervised work, our method learns cluster assignments end-to-end via gradient descent and avoids trivial solutions by enforcing a regularization term over the cluster assignments of views in a given batch.", "{{cite:384040f6241ebf29f315236ae8e6f8798c3b0a07}} presented a two-step algorithm for unsupervised classification and proposed the SCAN-Loss (Semantic Clustering by Adopting Nearest neighbors) as part of the learning pipeline.{{cite:384040f6241ebf29f315236ae8e6f8798c3b0a07}} decouple feature learning and clustering by learning two encoders in distinct optimization steps.The algorithm extracts a set of nearest neighbors from each observation and uses them as priors to learn a second network for semantic clustering.The nearest neighbors are mined from representations of an encoder trained using a self-supervised pretext task.Our implementation builds on top of the SCAN-Loss, but unlike {{cite:384040f6241ebf29f315236ae8e6f8798c3b0a07}}, we employ a Siamese network architecture to learn representations via cluster assignments, end-to-end, without the necessity of optimizing for a second self-supervised task or mining of nearest neighbors.Our objective is to compare consistent assignments to a set of clusters for each view instead of comparing the underlying representations nor their true clusters.By contrasting high-level representations of the data, we obtain reliable data representations that perform well in downstream tasks." ]

[]

[ "Conventional decompilation method. Decompilation originated from the need for software porting. The relevant research can be traced back to the 1960s, the decompiler D-Neliac {{cite:e03f1837670e594c67a1df17954e41308721b615}} built by the US Naval Electronics Laboratory for the Neliac program language {{cite:49cb28cc292008563836d1b4ba50fc43d122d30c}}. In the following half-century of research and development, many decompilers were developed, such as Hex-Rays {{cite:5bf216d7847fd07bd7fecca0a19bbf7fc1722f4f}}, Phoenix {{cite:530028edf55c14bb658204631e742df6d4bd4487}}, RetDec {{cite:c196e4031be4567748ffaee3d01d03ee647b4d7f}}, and Ghidra {{cite:b8d7fae95b66e03970d80e9ab8d49c497b0b925f}}. Hex-Rays is currently the most mature commercial decompiler and is still considered the de-facto industry standard in the software security industry. RetDec is a redirectable decompiler based on the LLVM architecture developed by the Czech security company Avast, aiming to be the first “universal” decompiler to support multiple architectures and program languages. Ghidra is developed by the National Bureau of Security Research, which supports multiple processor instruction sets and executable formats.", "Other works, such as TIE {{cite:a05dadd3f554fb6f19af4ff35464562c95170cc8}}, DIRE {{cite:d923d8c583d36d176d26a000d5d688b966d7775f}}, DREAM {{cite:f2d7cbf27dd97d6093aa4f32ae035c3d829116ee}}, DREAM++ {{cite:bf00c518f3f42abe2f496b8f802c6fc241346d36}}, target at improving the readability of decompiled code by refactoring variable types {{cite:a05dadd3f554fb6f19af4ff35464562c95170cc8}} or names {{cite:d923d8c583d36d176d26a000d5d688b966d7775f}}, as well as eliminating goto statements {{cite:f2d7cbf27dd97d6093aa4f32ae035c3d829116ee}}, {{cite:bf00c518f3f42abe2f496b8f802c6fc241346d36}}.Although these works have made significant improvements to decompilation, they all need to detect known control flow structures based on manually written rules and patterns, and these rules are difficult to develop, error-prone, usually only capture part of the known CFG, and require long development cycles. Different from these traditional rule-based decompilers, the goal of SEAM is based on deep neural networks to automatically learn and extract rules from code data, which is learned from the principles of NMT, trying to break through the problems of poor scalability, complex development, and long cycle of those tools due to over-reliance on manual rule definition.", "Neural networks for program analysis.The artificial intelligence (AI) technology represented by deep learning has been widely used in program analysis such as function recognition {{cite:13f445402571b2fcd65426cee611f4cee03017b0}}, {{cite:4713e4fd758c52cc774b8b06b41ee3e1d35b140f}}, {{cite:bc917a98fd476a7dcd87d4bf4257c408b310a2c9}}, {{cite:e865dfe2c897c48af0b9233717d0e27ebfa7cd38}}, {{cite:cbeb649d24ec22e8c6411dea9ecffa6bf5a48a67}} and identifier renaming {{cite:4034b5314ba77728dc4a4aa32e9550bf4c24fe01}}, {{cite:f481bc49220ec662fe67df1d461d52b5557d0742}}, {{cite:d923d8c583d36d176d26a000d5d688b966d7775f}}, improving the degree of automation and intelligence effectively of the software development or analysis process.Function identification is the cornerstone of reverse engineering and binary program analysis. The premise of decompilers to provide helpful output is to know the location of the function {{cite:bc917a98fd476a7dcd87d4bf4257c408b310a2c9}}. Rosenblum et al. {{cite:13f445402571b2fcd65426cee611f4cee03017b0}} first converted the function recognition problem in binary programs into a supervised machine learning classification problem. Bao et al. {{cite:bc917a98fd476a7dcd87d4bf4257c408b310a2c9}} proposed a machine-learning and weighted prefix tree-based automatic function recognition system BYTEWEIGHT. Shin et al. {{cite:e865dfe2c897c48af0b9233717d0e27ebfa7cd38}} introduced RNN based on BYTEWEIGHT, proved that RNNs have a specific advantage in binary function recognition.Identifier renaming aims to improve the readability of decompiled code. Jaffe et al. {{cite:4034b5314ba77728dc4a4aa32e9550bf4c24fe01}} designed a lexical n-gram-based NMT model to automatically assign meaningful names to variables, which is based on vocabulary information obtained by tokenized code. DIRE {{cite:d923d8c583d36d176d26a000d5d688b966d7775f}} further integrates the structural information of the AST corresponding to the tokenized code.Different from their work, we recovery semantic information based on functional characteristics of functions.", "Neural networks for program decompilation.Nowadays, Artificial intelligence technologies such as deep learning are widely used in NMT, image recognition, and other fields, relying on its powerful learning and expression capabilities to build deep learning-based decompilation methods is a popular research direction {{cite:c2d1a8e264d56961036432fb13a8fb01e117a1db}}, {{cite:37bf3935c213bdb074d65891b9b467833515f221}}, {{cite:39aeb756bc00e2ad1d8539f3316bcc310fec07ea}}, {{cite:5bac2dcbdb5fe1210592666871277206ac01f7c0}}. Katz et al. {{cite:c2d1a8e264d56961036432fb13a8fb01e117a1db}} first proposed RNN–based method for decompiling binary code snippets, which demonstrated the feasibility of using neural machine translation for decompilation. In order to improve the semantic and grammatical correctness of decompiled code, Katz et al. {{cite:37bf3935c213bdb074d65891b9b467833515f221}} further proposed a decompilation architecture based on LSTM called TraFix, considering the difference between natural language and programming language. Several approaches, such as preprocess assembly language (TraFix input) and the post-order traversal of C language (TraFix output), have been taken to shorten this difference. Fu et al. {{cite:39aeb756bc00e2ad1d8539f3316bcc310fec07ea}} proposed an end-to-end code decompilation framework based on several neural networks, different models used for different statement types. Unlike these three exiting works, Our neural translation model SEAM is solely based on the self-attention mechanism that can effectively generate high-level PL." ]

[]

[ "Existing incentives allocation methods usually are divided into two stages: (1) the response model estimates the response score; (2) the response score is served as an input of the optimization model, which tries to maximize the MPP under the budget constraint.", "Response Model Estimation. {{cite:889d050fc0655e66be688a9d781e34bad79d6498}}, {{cite:9db9bee989079741ec07c13d67aff217b038a226}} introduce a dynamic marketing allocation budget. Through strong fitting ability, neural networks can achieve a high prediction accuracy, widely used in many scenarios {{cite:42ca25488ebadca945d51a559ffb46dc8596c425}}, {{cite:743f1029bb694c7377c989b6178342753add7f31}}, {{cite:23c0831daaffab7a0e958be37c5e77f502190ccf}}. {{cite:2b619cc444eb0024f858769c15d91649763a4fcb}}, {{cite:1cb744baa341b83a3af26ca10cc440204178ec36}} use machine learning techniques to estimate future demand for new products and provide recommended prices for Airbnb hosts. However, as a black-box model, there are some gaps between the prediction and decision-making of deep neural networks {{cite:6a73d8378da5dfdc15516b6a2e4fe8cfd0606ab8}}. Therefore,  {{cite:f5d27835b9625b5bd828a609d8c65501b0c200c6}}, {{cite:298e3919a8d51a76b6b3ab103acea14d404f20b0}} propose a semi-black box model that extends the logarithmic demand curve through neural network and graph learning to solve this problem.", "Allocation Optimization.{{cite:691285423b7c05549e0556ab8dc8ef067a715e05}} proposes a fast approximation using semi-definite programming relaxation. {{cite:2b619cc444eb0024f858769c15d91649763a4fcb}} optimizes pricing decisions by using demand predictions from the regression trees as inputs of the price optimization model.{{cite:2dc9c072ea97e2324c942dbbdb321410e9ec19ab}}, {{cite:3f3e7a080d0ce9803bd7307e0f830f08baa85b4b}} through markov decision process value function and the connection with continuous sub-module function to solve the allocation optimization problem.{{figure:01feaf78-eb8f-4788-88e5-c1ea0b71497d}}" ]

[]

[ "Change detection has a long history, being one of the early problems tackled in remote sensing image understanding {{cite:28f307897190d5e60e3d49634912243228b5fd30}}. It is done using coregistered image pairs or sequences, and consists of identifying areas in the images that have experienced significant modifications between the acquisitions. Many of the state-of-the-art ideas in pattern recognition have been used for change detection in the past, from pixel-level comparison of images, to superpixel segmentation, object-level image analysis, and image descriptors {{cite:44b58fba63700bc9ac56b1e6021764753dfe7368}}. In this paper we treat change detection as a two class semantic segmentation problem, in which a label is predicted for each pixel in the input images. With the rise of machine learning algorithms for semantic segmentation, notably convolutional neural networks, many algorithms have attempted to learn to perform change detection. Most algorithms circumvented the problem of the scarcity of training data through transfer learning by using pretrained networks to generate pixel descriptors {{cite:b53417576c2001e1b845a841590d9f63d0db7bde}}, {{cite:7a83298030c988c569d966025a64aaff88f95272}}, {{cite:3248a2522b08cff05adfbb684efc5dd84e9f002a}}. Fully convolutional networks trained end-to-end to perform change detection have recently been proposed by several authors independently, usually using Siamese architectures {{cite:c71203f50e0016d7d884aa87cd0d727bb9293fab}}, {{cite:860b57aee96bcf32b7d70a166a9079ea660145a9}}, {{cite:c0a68badc0bb04aefb64f666fc4dea166021ed31}}, {{cite:50795c09b7435956a70845542d863c78986ab20c}}, {{cite:3721143017cc4d78cd783a225a4a47ac74f16110}}. Unsupervised {{cite:16f48d9402fee08fc0b84423fdea8b22ee2b9773}}, {{cite:277cc015e0852f76cae0eeb72ef76fc4da6cf3ec}}, {{cite:8411bb6eb3f5835aa32e7c92d8647ccb9e3ccae5}}, {{cite:81d6de33fbe97eef4173af0032d9806b4f2431d9}} and semi-supervised {{cite:461dec8b749a82116d6f4429fc34ed0a0519ae89}} alternatives have also been proposed to cope with the scarcity of accurately labelled data.", "Semantic segmentation algorithms attempt to understand an input image and predict to which class among a known set of classes each pixel in an input image belongs. Change detection is modelled in this paper and many others as a semantic segmentation problem which takes as input two or more images. Long et al. proposed the first fully convolutional network for semantic segmentation, which achieved excellent performance and inference speed {{cite:d010f79d32a49bfc695b64c72edd22490d145530}}. Since then, several improvements have been proposed for CNNs and FCNs. Ioffe and Szegedy have proposed batch normalization layers, which normalize activations and help avoid the vanishing/exploding gradient problem while training deep networks {{cite:22e4939daee0bfdbca06691d4b1f242962e79bd5}}. Ronneberger et al. proposed the usage of skip connections that transfer details and boundary information from earlier to later layers in the network, which improves the accuracy around the edges between semantic regions {{cite:7ced32ce3b23a84f7f6e9381ab83eb8798ba0a28}}. He et al. proposed the idea of residual connections, which have improved the performance of CNNs and FCNs and made it easier to train deep networks {{cite:6b4efc0260139e0e4be32cf6319cbb05208369cb}}.", "Noisy labels for supervised learning is a topic that has already been widely explored {{cite:2b033bae9333877df5df0744db0c4433a3b639da}}, {{cite:f83811ea14c21aa76af5548a61801191e0d7b8d3}}. In many cases, label noise is random (by this we mean, following the literature terminology, independent of the data and not correlated) and is modelled mathematically as such {{cite:31960afa30b52102dc571013507cda5dfeffae17}}, {{cite:d679398de79b9249187deb86d430843aa68f64fb}}, {{cite:7dd8c64eddb8eb3d87137d2ba829d8323b30c43e}}. Rolnick et al. showed that supervised learning algorithms are robust to random label noise, and proposed strategies to further minimize the effect label noise has on training, such as increasing the training batch sizes {{cite:7dd8c64eddb8eb3d87137d2ba829d8323b30c43e}}. In the case presented in this paper, the assumption that the label noise is random does not hold. Incorrect change detection labels are usually around edges between regions or grouped together, which leads the network to learn to overestimate detected changes as seen in Fig. REF (e). Ignoring part of the training dataset, known as data cleansing (or cleaning), has already been proposed in different contexts {{cite:0eacc6842975810066acc5304d627c25af398820}}, {{cite:3ea701ff2e3fe17282d0c2f1387c144ed658d86c}}, {{cite:0b0a6c08d39b08b796361d7967cf18e342d78bf8}}, {{cite:cdc7d22bc7deb94f9aaf2781172e701c3a22f92e}}.", "Weakly supervised learning is the name given to the group of machine learning algorithms that aim to perform different or more complex tasks than normally allowed by the training data at hand. Weakly supervised algorithms have recently gained popularity because they provide an alternative when data acquisition is too expensive. The problem of learning to perform semantic segmentation using only bounding box data or image level labels is closely related to the task discussed in this paper, since most methods propose the creation of an approximate semantic segmentation ground truth for training and dealing with its imperfections accordingly. Dai et al. proposed the BoxSup algorithm {{cite:3716b674dc50676fdadbcf10c2e517987e187b6e}} where region proposal algorithms are used to generate region candidates in each bounding box, before a semantic segmentation network is trained using these annotations and finally used to iteratively select better region proposal candidates. Khoreva et al. proposed improvements to the BoxSup algorithm that include using ad hoc heuristics and an ignore class during training {{cite:995ad03ff1b84afb3c2d9300d326dfa96085f4ef}}. They obtained best results using region proposal algorithms to create semantic segmentation training data directly from bounding boxes. Lu et al. modelled this problem as a simultaneous learning and denoising task through a convex optimization problem {{cite:ebbc2e0434477f2fbad1ce1b61a3ee2c855c1e93}}. Ahn and Kwak proposed combining class activation maps, random walk and a learned network that predicts if pixels belong to the same region to perform semantic segmentation from image level labels {{cite:b40ae0a67541af4f74b2735ab68c223e51bb8eb6}}. Zhou et al. proposed the class activation mapping technique {{cite:30c171fc1aa41df338beea55dc6738357b976fe9}}, which allows the networks to localize what regions in the image contribute to the prediction of each class, which can be harnessed for generating pixel-level predictions from image-level labels.", "Post-processing methods that use information from guide images to filter other images, such as semantic segmentation results, have also been proposed {{cite:92251623987b5c12a082bcdb4ee8a3030727238b}}, {{cite:99d36dcf8fd901b4165f7d869528f764df1b027d}}, {{cite:14d4966544a7e562ac672fc7b4d670303ecbf242}}. A notable example is the Dense CRF algorithm proposed by Krähenbühl and Koltun, in which an efficient solver is proposed for fully connected conditional random fields with Gaussian edge potentials {{cite:050e37715bc08554800e5239c97771c6dd6a2de5}}. The idea of using a guide image for processing another is also the base of the Guided Image Filtering algorithm proposed by He et al.  {{cite:970045c1371c24d3de72f87ceb00de8e873c505a}}, where a linear model that transforms a guide image into the best approximation of the filtered image is calculated, thus transferring details from the guide image to the filtered image. The use of joint filtering is popular in the field of computational photography, and has been used for several applications {{cite:92251623987b5c12a082bcdb4ee8a3030727238b}}, {{cite:99d36dcf8fd901b4165f7d869528f764df1b027d}}, {{cite:14d4966544a7e562ac672fc7b4d670303ecbf242}}. One of the building blocks of the filtering method we propose in this paper is the anisotropic diffusion, proposed by Perona and Malik {{cite:3c2418e34b20800a017d860fbbf6f9b33ff58864}}, an edge preserving filtering algorithm in which the filtering of an image is modelled as a heat equation with a different diffusion coefficient at each edge between neighbouring pixels depending on the local geometry and contrast. However, to the best of our knowledge, this algorithm has not yet been used for guided filtering." ]

[]

[ "Crowd-Sourced Annotations.The inter-and intra-observer variations among experts due to intrinsic differences in human annotators widely occur in computer vision tasks. Many approaches {{cite:4861a27172343900bbb5ebe4fd56842970027075}}, {{cite:35fde4f667b4b2f09e258f63e7a0072d843ef47d}}, {{cite:abb9c6fd942669a69a36f6d76554804e3731c7bd}}, {{cite:bec7af86a3b8041433beb386b548d884d102cd79}}, {{cite:fe0a2d1af3060091caa814ac09f2f35aba81d98b}}, {{cite:edc6513bd664747845e0db77cbe1fc8f1130e4db}}, {{cite:aec6943e5e7476269f4aed5dcca7bc5503e70ff2}} have been proposed to deal with the inconsistent label problems from the crowd-sourced annotations.For example, {{cite:c1782efa7e9758c1c3d1ed08f96cee21195fa2bd}} proposes an EM algorithm and a crowd layer to deal with the crowd-sourcing problems.To solve the noise and inconsistency in some limited-sized datasets from multiple annotators, {{cite:35fde4f667b4b2f09e258f63e7a0072d843ef47d}} proposes a crowd-sourced representation learning framework that generates a large number of training instances from only small-sized labeled sets.However, most existing works focus on classification tasks. How to deal with inconsistent annotations with numerical labels for regression tasks is still underexplored.", "Cross-Dataset Adaptation.Since data from source and target domains have different distributions, the key challenge for domain adaptation {{cite:23390b0d42d4710e65f72b0bb4be14d4735c09b5}}, {{cite:b2ee0dee24e7b97bf4b5730c36f99786ff6c11e8}}, {{cite:2203c0538fe56fca2cf6462904f6370db65200cf}}, {{cite:01444846bda36acda2ad79fc4da595673e269c4f}}, {{cite:e5bb3d8963a40403e14723d99212f0bea32ec8ae}}, {{cite:bd34b6965120c7db333bc00e7cad3e551f586034}}, {{cite:0610eec538095c8d8d89871707f763a33df7ab72}} is to address the domain shift between them. Fine-tuning {{cite:cf12e6fc84330fe644436f84a98c39b8b51f2ca7}}, {{cite:de4a6df6d7623486952fdb41f69d40dabe5dd2f3}}, which diminishes the shift between the two domains, is a typical way for domain adaptation with transfer learning {{cite:f8100b39c5aaf9ca1b38ef0a47597d3650a10043}}, {{cite:17cfa350064a14ad1cc9310d0989df1b11207c58}}.For instance, {{cite:de4a6df6d7623486952fdb41f69d40dabe5dd2f3}} proposes an adaptive fine-tuning approach to find the optimal strategy per image for the target data. Transfer learning with triplet loss {{cite:17cfa350064a14ad1cc9310d0989df1b11207c58}} is employed for bridging the gap between different domains.Adversarial learning is another essential way for domain adaptation {{cite:23390b0d42d4710e65f72b0bb4be14d4735c09b5}}, {{cite:b2ee0dee24e7b97bf4b5730c36f99786ff6c11e8}}, {{cite:65c719cebfc42c9e7fcd5979036288880abbedb3}}, {{cite:6de3d55f7984ee052fdcec634f4ff235ce1ffa1d}}, {{cite:0cfddef625227ed4513514a977b5e407c50daeb8}}.For example, {{cite:b2ee0dee24e7b97bf4b5730c36f99786ff6c11e8}} presents a partial adversarial domain adaptation approach to alleviate negative transfer by reducing the weights of outliers and promote positive transfer.A multi-adversarial domain adaptation approach {{cite:23390b0d42d4710e65f72b0bb4be14d4735c09b5}} is proposed to capture multi-mode structures to ensure fine-grained alignment of distributions. The domain-symmetric network {{cite:65c719cebfc42c9e7fcd5979036288880abbedb3}} is presented on a symmetric design of source and target classifiers with adversarial training. However, such domain adaptation approaches cannot directly address the inconsistency annotation problem raised from multi-sourced data." ]

[]

[ "The field of vision-based lane detection has grown considerably in the last decade. The popularity of camera sensors has allowed lane detection in 2D to gain significant momentum. Traditional methods typically design hand-crafted features, adopt mathematical optimized algorithms and geometry or scene context to learn lane structures well {{cite:3adeef9144436c50fac184d5878f0027cad7df94}}, {{cite:70e25bd3220c0f1c3e865a4dcd1868fb78f576d3}}, {{cite:6d5710cc6542582163c2aac12ddfe869774e8fcd}}. Deep methods built by convolutions have exploded in recent years, making significant progress and applicable systems for real applications {{cite:3ba46f8c2241cfd66362e799b50807adec4fedc3}}, {{cite:f3d62eb072b726de656285b8d250e6687688c962}}. Two-stage methods which extract segmentation or proposal plus post-processing ruled the filed for several years {{cite:187507b3dc1031473e133eb840f9d3ac346f293a}}, {{cite:175b832dec7b5db4066195d0a5d57bafc55acd6b}}, {{cite:33b7049d4174fe3274681f9a6be90c3200401136}}, {{cite:956700bce58e9e228a76a582f3e4ee00f6873d65}}, {{cite:88ea538e37d8919ca4b6ae121e021eadf700c259}}, {{cite:2ef148869a6085b7e9eb4f2fe9a48a27aa1a72bf}}, {{cite:aca4ead11bb585f8c800f4ae2f119d98c333fa66}}, {{cite:10c61e1454bcdfff8ae8a76c4622698e52236ee1}}, {{cite:86737f54293b695c4bb8f7d8028ac895cafefcfb}}, {{cite:771d8d009ad7c8ea6db4b61f5a83ff2af3ea050d}}, {{cite:ca2bb88afb09de0cae03e188cad3cd48ebbdc005}}, {{cite:e3d4099cbc6f1641cdf6ff76ae6307b26b12c833}}, {{cite:a9b60e66aae301f0a480a3de28e68b28c374ba81}}, {{cite:b1bdd888786010de0bbdffdc44f38d086cfff867}}. To streamline the pipeline into an end-to-end fashion, single-stage methods {{cite:468dffa8beaaf72e5db12bd9f826b6084d61bd20}}, {{cite:8e3b6299e635a2af961a150a53b8f604def7c684}} directly estimate coefficients of prior mathematical curves have shown both higher efficacy and efficiency. However, those 2D detectors are built with specific planar world assumptions, resulting in a limited representation of realistic and complex 3D road undulations.", "The 3D lane detection task has attracted research interest because it doesn't rely on the plane assumption to predict the lanes {{cite:87bdffc9854579636a4303725a16cd452640ecc2}}, {{cite:7350331e649798b23b0ada8bdc6797a0fa59227e}}, {{cite:f6e0ac24dceca1e9936941198d0388ee6b695c39}}, {{cite:b4b435aaf46800a8daec9133f240d504bce00853}}, {{cite:f9826ee09cb0d375e031d7cfa818a85258a99326}}, {{cite:c33a040e0a447d3aa17c459af7354cd085d219dc}}, {{cite:1fbb53064364bbcb5a21f44289a02b83ec09d994}}. However, detecting a 3D lane from a single RGB image is non-trivial. The results can be easily affected by appearance variation due to changes in lighting condition, texture, background, and camera pose. One solution is to utilize the extra information from other sensors such as LiDAR or stereo cameras {{cite:87bdffc9854579636a4303725a16cd452640ecc2}}, {{cite:7350331e649798b23b0ada8bdc6797a0fa59227e}}, {{cite:f6e0ac24dceca1e9936941198d0388ee6b695c39}}. Still, the cost of these devices is too expensive to be applied to them on consumer-grade products widely {{cite:f9826ee09cb0d375e031d7cfa818a85258a99326}}, while arising multi-sensor calibration problems during driving.Unlike multi-sensor methods, monocular methods {{cite:b4b435aaf46800a8daec9133f240d504bce00853}}, {{cite:f9826ee09cb0d375e031d7cfa818a85258a99326}}, {{cite:c33a040e0a447d3aa17c459af7354cd085d219dc}} only require a single image and camera pose to transform image/features from perspective-view to top-view for accurate 3D lane detection. Previous methods rely on the ground truth camera pose for evaluation and cannot work well when the camera pose changes dynamically when driving at rough terrain or accelerating {{cite:e63dddc288a7e9f3f316b0565b8d93bdd061f920}}). In contrast, our method provides a full-vision-based 3D lane detector that is not dependent on the ground truth camera pose and other sensors. It can flexibly adapt to the changing driving environment with an affordable device cost." ]

[]

[ "There have been a number of prior work that tackled HE acceleration using CPU, GPU, FPGA, and ASIC.", "CPU acceleration: {{cite:e62ea5f45f2cd1736e4b929219d643e178f3969f}} spawns multiple threads to parallelize HE ops over multi-core CPUs.{{cite:8505987cc85de99e21af1724718e6d9ddba39cf8}}, {{cite:f5bf28d4ec8c75df837fb4811ba26da347a714c8}} leverage the short-SIMD support.{{cite:29b3a5e5c47853c5de185831d29acd3d99740510}} exploits the algorithmic analysis of {{cite:cd05fec898f12cb1963fb169f79e93b2fdf024e5}} for efficient bootstrapping implementation.However, other platforms outperform CPU while processing HE ops.", "GPU acceleration:GPU is a prime candidate for accelerating HE ops as it is equipped with a massive number of integer units and abundant memory bandwidth.However, a majority of prior works did not handle bootstrapping {{cite:2e0a7e0b0da632ab23d8dfc33d9995c57bd91275}}, {{cite:730184af7deda56bbac585bf15054d028b0471d8}}, {{cite:655526a41829fc82d46ef331f5d9544bad70b1c6}}, {{cite:8505987cc85de99e21af1724718e6d9ddba39cf8}}.{{cite:a57dc5427ea7e4cea8262c4df9fe2ed274bdcd18}} is the first work that supports CKKS bootstrapping on GPU.It optimized HE computation by fusing GPU kernels,saving off-chip access and henceleading to 242{{formula:310035e5-d35c-4bb0-a401-dadcf808b186}} faster bootstrapping over CPU.However, the deficiency of on-chip storage forces some kernels to stay unfused {{cite:42a412396e7e11acacb65babf23897411605bfd7}}.BTS allows all temporal data to reside on-chip, minimizing the off-chip access.", "FPGA/ASIC acceleration:A different set of works accelerates HE using FPGA or ASIC, but most of them do not consider bootstrapping {{cite:36de4b8d92f835576782b153d094e291b824db19}}, {{cite:5911767f2bfeea5d43485160ffd66ac4b9bbe286}}, {{cite:db370d77744f0d60bf286a45800d7eb1204c5b89}}, {{cite:d6a9445569a3bca89823231105a023b111b8adf5}}.HEAX {{cite:36de4b8d92f835576782b153d094e291b824db19}} dedicated hardware for CKKS mult on FPGA, reaching 200{{formula:edc30400-7dd9-4826-ac66-f4b706f40262}} performance gain over CPU.However, its design is rather fixed to a limited set of parameters, not considering bootstrapping.F1 {{cite:3b9c9bba32a4feab0c3d691d4192a4e495df37c2}} is the first ASIC design to support bootstrapping.It is a programmable accelerator supporting multiple FHE schemes, including CKKS and BGV.F1 achieves impressive performance on various FHE applications as it provides tailored high-throughput computation units and stores all {{formula:04605b80-ba1b-4d75-aa68-d2ddde621f3d}} s on-chip, minimizing the off-chip access.However, F1 targets the parameter sets with low degree {{formula:03d0969f-65f8-4486-8e9f-d60f50b70336}} ; thus, it supports only non-packed (single-slot) bootstrapping, the throughput of which is greatly exacerbated compared to BTS." ]

[]

[ "Previous work on KR couplings has focused on existence and approximation questions in relation to universal approximation {{cite:28a26fa91ebf9fa0201c250e471bbe3c1243765a}}, {{cite:61c4b7c42d9a1687e92c46cb21a98d51df6eec64}}, {{cite:8aa842a23164c6a512be7077a29eb1d931923630}}. This universal approximation property of KR maps has been a frequent motivation from a theoretical viewpoint of normalizing flows; see Sec. 3 in {{cite:9414db03112744b5b2d6eb72fadb8dd6c0aef840}}. KR maps have been used for various learning and inference problems {{cite:9414db03112744b5b2d6eb72fadb8dd6c0aef840}}, {{cite:5fc88b9ea2cb61fd097db9ead85cee37e07d2bb2}}, {{cite:c3990f6d784c58b2c65e7e5899c68ffe98adfdcb}}, {{cite:2bba785ebdbc6569deb82b60fa9489d2cdd47f5f}}.", "Kong and Chaudhuri {{cite:e4bb41f4030c283b3a4757fbc6720d413d6b2b63}} study the expressiveness of basic types of normalizing flows, such as planar flows, Sylvester flows, and Householder flows. In the one-dimensional setting, they show that such flows are universal approximators.However, when the distributions lives in a {{formula:021a96f0-6b6f-4ad2-be38-1cce851f839d}} -dimensional space with {{formula:5598c530-2aac-4b2e-860d-fc623a475b28}} , the authors provide a partially negative answer to the universal approximation power of these flows. For example, they exhibit cases where Sylvester flows cannot recover the target distributions. Their results can be seen as complementary to ours as we give examples of arbitrary slow statistical rates and we develop the consistency theory of KR-type flows {{cite:5fc88b9ea2cb61fd097db9ead85cee37e07d2bb2}}.", "Jaini et al. {{cite:5e0a392ade7f97ad5b14c18b339ebb85a0f13b2a}} investigate the properties of the increasing triangular map required to push a tractable source density with known tails onto a desired target density.Then they consider the general {{formula:c7b828ee-7558-4b4b-9d9d-f730df1cb38a}} -dimensional case and show similarly that imposing smoothness condition on the increasing triangular map will result in a target density with the same tail properties as the source. Such results suggest that without any assumption on the target distribution, the transport map might be too irregular to be estimated. These results echo our assumptions on the target to obtain fast rates and complement ours by focusing on the tail behavior while we focus on the consistency and the rates." ]

[]

[ "After the rapid success of deep neural networks, the community has recently started to acknowledge that we have very limited understanding of how these architectures work and how they are able to achieve such remarkable results. Some of the first visualization tools that were proposed were saliency maps {{cite:1ba68ff093ed800d2aa049f1c4d55852dba6f823}}, which indicate the areas of the input image that are discriminative and most important with respect to the given class by using the intensity of the pixels. That is, given an input image, the saliency method ranks its pixels based on their influence on the final class score by using derivatives and backpropagation. A similar idea is the Grad-CAM method {{cite:5edbe71821b93e06a74159267ced2a3bcc1bff59}}, which employs the gradients of any target concept to produce a localization map that highlights the important regions in the image that predict the concept. Grad-CAM can also be combined with existing pixel-space visualizations (Guided Grad-CAM) to achieve visualizations that are both high-resolution and class-discriminative.", "Another way to study the importance of the pixels in the input image is to perturb the image by occluding patches of the image and see how the classification score drops {{cite:b4c0bc1995b2d15705b41a79481e709ae3e801db}}. Other authors follow a similar idea by directly deleting some parts of the image in order to find the part of the image that makes the final class score drop the most {{cite:5dc71e39068d66ee6dd0a755cbedc4c5a111e0eb}}. A bit differently, Koh and Liang use influence functions to gain understanding on which training points are more relevant to the final classification without having to retrain the network {{cite:26b785ead9495c0ddb3a63a912280e3bfd20fb81}}. Their method essentially shows how the model parameters change as we upweight a training point by an infinitesimal amount.", "In {{cite:b4c0bc1995b2d15705b41a79481e709ae3e801db}}, the authors try to understand CNNs backwards by creating a Deconvolutional Network. A DeConvNet can be thought as a ConvNet model with the same components but in reverse, so that it maps features to pixels. A similar idea is followed in {{cite:dd6ffdf186c50c3bbf15085fa7d69ad0dbebe8f8}}, where image representations are studied by inverting them with an up-convolutional neural network. Other work focuses on developing human-friendly concepts to help understand the machine. In {{cite:7412f21032170280c236176e91fb70cde2fb667c}}, Concept Activation Vectors are introduced to provide an interpretation of the internal state of a network in human-friendly concepts, and Deep Dream and Lucid are Google projects that intend to humanize what the hidden layers and neurons see in the input images {{cite:85e715baa5ceae4243083ef8fdf0b7b0f7cf1c95}}. Their paper develops human-like visualizations for understanding what each neuron is focusing on. In the conclusion, the authors point out that one of the issues that still stands out in network interpretability is finding which units are most meaningful for understanding neural net activations, which is what we study in this paper. Other studies also engage in trying to bridge human concepts and neural networks; for example, {{cite:7b40d3b500b2e7219d3421479ffa372000b32469}} investigates how transferable are features in deep neural networks by differentiating between general and specific features learnt by the architecture. It was asked in {{cite:407606af755dcdc4eaa3cffa6272a51e73868845}} if CNNs learn class hierarchy, and {{cite:dbc043eafcf3181fadb0e86b067fdeda63d690c8}} contains a study of the semantic concepts learnt by the units, such as colors, scenes, and textures.", "Other papers bring up criticism to some of the methods we just described. In {{cite:80cac0431a4fa1e19228aed28c92ed2d5e9d99f1}}, it is argued that saliency methods lack reliability when the explanation is sensitive to factors that do not contribute to the model prediction, and in {{cite:2ba745fe4b84dd80688a989383a4701f141e91db}} it is shown that DeConvNets and Guided Backpropagation do not produce the theoretically correct explanations for a linear model, and so even less for a multi-layer network with millions of parameters. Finally, in {{cite:7d4d42089eb252d6a0abb42f5a8ea6bce0fec373}} and {{cite:d51545217e7aa63e94c3c18f7547ad6e6595bf87}}, the authors propose that neurons do not encode single concepts and that they are in fact multifaceted, with some concepts being encoded by a group of neurons rather than by a sole neuron by itself.", "As summarized in this section, most papers in the literature focus on qualitatively studying the specific features and concepts that are being learnt in the network, rather than quantifying the importance of each hidden neuron towards the final class score. Our method bridges this gap by focusing on quantifying the relevance of each neuron, beyond providing qualitative information based on visualization methods. An exception in the literature is the recent paper {{cite:1ee77dae0f1ab0384eb667de23ef32aafee4f597}}, where the most relevant units for each class are defined by computing which cause the most accuracy loss when removed individually. However, this approach is too computationally costly." ]

[]