Srihari Thyagarajan's picture

Srihari Thyagarajan

Haleshot

·

https://haleshot.github.io/

AI & ML interests

AI, ML, DL, CV, RecSys

Recent Activity

liked a Space 12 days ago

franciszzj/Leffa

reacted to singhsidhukuldeep's post with 🚀 17 days ago

Exciting breakthrough in AI Recommendation Systems! Just read a fascinating paper from Meta AI and UW-Madison researchers on unifying generative and dense retrieval methods for recommendations. The team introduced LIGER (LeveragIng dense retrieval for GEnerative Retrieval), a novel hybrid approach that combines the best of both worlds: Key Technical Innovations: - Integrates semantic ID-based generative retrieval with dense embedding methods - Uses a T5 encoder-decoder architecture with 6 layers, 6 attention heads, and 128-dim embeddings - Processes item attributes through sentence-T5-XXL for text representations - Employs a dual-objective training approach combining cosine similarity and next-token prediction - Implements beam search with size K for candidate generation - Features an RQ-VAE with 3-layer MLP for semantic ID generation Performance Highlights: - Significantly outperforms traditional methods on cold-start recommendations - Achieves state-of-the-art results on major benchmark datasets (Amazon Beauty, Sports, Toys, Steam) - Reduces computational complexity from O(N) to O(tK) where t is semantic ID count - Maintains minimal storage requirements while improving recommendation quality The most impressive part? LIGER effectively solves the cold-start problem that has long plagued recommendation systems while maintaining computational efficiency. This could be a game-changer for e-commerce platforms and content recommendation systems! What are your thoughts on hybrid recommendation approaches?

reacted to singhsidhukuldeep's post with 👍 17 days ago

Exciting breakthrough in AI Recommendation Systems! Just read a fascinating paper from Meta AI and UW-Madison researchers on unifying generative and dense retrieval methods for recommendations. The team introduced LIGER (LeveragIng dense retrieval for GEnerative Retrieval), a novel hybrid approach that combines the best of both worlds: Key Technical Innovations: - Integrates semantic ID-based generative retrieval with dense embedding methods - Uses a T5 encoder-decoder architecture with 6 layers, 6 attention heads, and 128-dim embeddings - Processes item attributes through sentence-T5-XXL for text representations - Employs a dual-objective training approach combining cosine similarity and next-token prediction - Implements beam search with size K for candidate generation - Features an RQ-VAE with 3-layer MLP for semantic ID generation Performance Highlights: - Significantly outperforms traditional methods on cold-start recommendations - Achieves state-of-the-art results on major benchmark datasets (Amazon Beauty, Sports, Toys, Steam) - Reduces computational complexity from O(N) to O(tK) where t is semantic ID count - Maintains minimal storage requirements while improving recommendation quality The most impressive part? LIGER effectively solves the cold-start problem that has long plagued recommendation systems while maintaining computational efficiency. This could be a game-changer for e-commerce platforms and content recommendation systems! What are your thoughts on hybrid recommendation approaches?

View all activity

Organizations

Haleshot's activity

upvoted a collection 18 days ago

PaliGemma 2 Release

Vision-Language Models available in multiple 3B, 10B and 28B variants. • 23 items • Updated 12 days ago • 119

upvoted a collection 23 days ago

BhasaAnuvaad

A Speech Translation Dataset for 13 Indian Languages • 11 items • Updated 27 days ago • 12

upvoted a collection about 1 month ago

UltraVox Audio Language Model Release 🔊

3 items • Updated Nov 15 • 15

upvoted a collection about 2 months ago

SmolLM2

State-of-the-art compact LLMs for on-device applications: 1.7B, 360M, 135M • 15 items • Updated 3 days ago • 195

upvoted an article 2 months ago

Article

Fine-tuning LLMs to 1.58bit: extreme quantization made easy

Sep 18

• 213

upvoted a paper 3 months ago

Depth Pro: Sharp Monocular Metric Depth in Less Than a Second

Paper • 2410.02073 • Published Oct 2 • 41

upvoted an article 3 months ago

Article

Introducing the Open FinLLM Leaderboard

Oct 4

• 66

upvoted 3 collections 3 months ago

Molmo

Artifacts for open multimodal language models. • 5 items • Updated 28 days ago • 289

🎯DART-Math

Difficulty-Aware Rejection Tuning for Mathematical Problem-Solving [NeurIPS 2024] @ https://github.com/hkust-nlp/dart-math • 20 items • Updated Sep 26 • 6

Llama 3.2

This collection hosts the transformers and original repos of the Llama 3.2 and Llama Guard 3 • 15 items • Updated 20 days ago • 548

upvoted an article 3 months ago

Article

ColPali: Efficient Document Retrieval with Vision Language Models 👀

By

•

Jul 5

• 182

upvoted an article 4 months ago

Article

Fine-tune Llama 3 with ORPO

By

•

Apr 22

• 228

upvoted a collection 4 months ago

DeepSeek-Prover

DeepSeek-V1-and-V1.5-Series • 7 items • Updated Aug 16 • 18

upvoted a collection 5 months ago

Parler-TTS: fully open-source high-quality TTS

If you want to find out more about how these models were trained and even fine-tune them yourself, check-out the Parler-TTS repository on GitHub. • 8 items • Updated 24 days ago • 49

upvoted a paper 5 months ago

ToRA: A Tool-Integrated Reasoning Agent for Mathematical Problem Solving

Paper • 2309.17452 • Published Sep 29, 2023 • 3

upvoted 2 collections 5 months ago

AIMO Progress Prize

Models and datasets used in the winning solution to the AIMO 1st Progress Prize • 7 items • Updated Jul 19 • 11

🪐 SmolLM

A series of smol LLMs: 135M, 360M and 1.7B. We release base and Instruct models as well as the training corpus and some WebGPU demos • 12 items • Updated 3 days ago • 204

upvoted an article 5 months ago

Article

SmolLM - blazingly fast and remarkably powerful

Jul 16

• 292

upvoted 2 collections 8 months ago

Distil-Whisper Models

The first version of the Distil-Whisper models released with the Distil-Whisper paper. • 4 items • Updated Mar 21 • 36

OpenELM Pretrained Models

4 items • Updated Oct 4 • 49