A naukri.com initiative
ML News
Arxiv
1w
378
Image Credit: Arxiv
Bandit Social Learning: Exploration under Myopic Behavior
- The study focuses on social learning dynamics influenced by online reviews.
- Agents follow a myopic behavior without exploration in a multi-armed bandit protocol.
- The study derived learning failures for myopic behaviors and provided matching positive results.
- The results emphasize the importance of intentional exploration in bandit algorithms.
Read Full Article
22 Likes
Arxiv
1w
306
Image Credit: Arxiv
Universal Scaling Laws of Absorbing Phase Transitions in Artificial Deep Neural Networks
- Conventional artificial deep neural networks operating near the phase boundary of signal propagation dynamics exhibit universal scaling laws in non-equilibrium statistical mechanics.
- Multilayer perceptrons and convolutional neural networks belong to the mean-field and directed percolation universality classes, respectively.
- Finite-size scaling suggests a potential connection to the depth-width trade-off in deep learning.
- Hyperparameter tuning to the phase boundary is necessary but insufficient for achieving optimal generalization in deep networks, indicating the importance of nonuniversal metric factors.
Read Full Article
18 Likes
Arxiv
1w
331
Image Credit: Arxiv
Comparative Performance Evaluation of Large Language Models for Extracting Molecular Interactions and Pathway Knowledge
- Large language models (LLMs) show potential for automating the extraction of molecular interactions in biological systems.
- The study evaluates the effectiveness of various LLMs in recognizing protein interactions, identifying genes related to radiation-affected pathways, and delineating gene regulatory relationships.
- Larger models demonstrate superior performance, particularly in extracting complex interactions among genes and proteins.
- LLMs face challenges in identifying groups with diverse functions and recognizing highly correlated gene regulatory relationships.
Read Full Article
19 Likes
Arxiv
1w
285
Image Credit: Arxiv
Physics-tailored machine learning reveals unexpected physics in dusty plasmas
- Dusty plasma is a mixture of ions, electrons, and macroscopic charged particles commonly found in space and planetary environments.
- Machine learning models are used to learn the complex forces governing the interactions between particles in a dusty plasma.
- A physics-tailored machine learning approach is demonstrated, accounting for symmetries, non-identical particles, and non-reciprocal forces.
- The model accurately infers particle masses, reveals deviations from theoretical assumptions, and enables the discovery of new physics.
Read Full Article
17 Likes
Arxiv
1w
126
Image Credit: Arxiv
Toward a Theory of Tokenization in LLMs
- Tokenization is considered a necessary initial step for designing performant language models.
- Transformers trained on certain data processes without tokenization fail to learn the right distribution and predict characters according to a unigram model.
- With tokenization, transformers are able to break through this barrier and model the probabilities of sequences drawn from the source near-optimally.
- The use of tokenization in language modeling is justified through the study of transformers on Markovian data.
Read Full Article
7 Likes
Arxiv
1w
344
Image Credit: Arxiv
Potential Field Based Deep Metric Learning
- A new approach to Deep Metric Learning (DML) is proposed.
- The approach represents the influence of each example by a continuous potential field.
- Attractive/repulsive potential fields are used to model interactions among embeddings.
- The proposed method outperforms state-of-the-art baselines on standard DML benchmarks.
Read Full Article
20 Likes
Arxiv
1w
218
Image Credit: Arxiv
ChatEMG: Synthetic Data Generation to Control a Robotic Hand Orthosis for Stroke
- Intent inferral on a hand orthosis for stroke patients is challenging due to the difficulty of data collection. Traditional approaches require a large labeled dataset from the new condition, session, or subject to train intent classifiers. In this paper, the authors propose ChatEMG, an autoregressive generative model that can generate synthetic EMG signals conditioned on prompts. ChatEMG enables them to collect only a small dataset and expand it with synthetic samples conditioned on prompts. Experimental results show that these synthetic samples can improve intent inferral accuracy for different types of classifiers.
- The authors demonstrate that their complete approach can be integrated into a single patient session, including the use of the classifier for functional orthosis-assisted tasks.
- This is the first time an intent classifier trained partially on synthetic data has been deployed for functional control of an orthosis by a stroke survivor.
- Videos, source code, and additional information can be found at https://jxu.ai/chatemg.
Read Full Article
13 Likes
Arxiv
1w
163
Image Credit: Arxiv
Compressing Search with Language Models
- Millions of people turn to Google Search each day for information on various topics.
- A new approach, SLaM Compression, has been introduced to reduce the dimensionality of search data.
- SLaM Compression quantifies search terms using pre-trained language models to create a memory-efficient summary.
- CoSMo, a Constrained Search Model, is presented to estimate real-world events using only search data.
Read Full Article
9 Likes
Arxiv
1w
336
Image Credit: Arxiv
SigmaRL: A Sample-Efficient and Generalizable Multi-Agent Reinforcement Learning Framework for Motion Planning
- This paper introduces SigmaRL, an open-source, decentralized framework designed to enhance sample efficiency and generalization of multi-agent Reinforcement Learning (RL) for motion planning of connected and automated vehicles.
- SigmaRL aims to address the limited generalization capacity of RL agents by proposing five strategies to design information-dense observations, focusing on general features applicable to most traffic scenarios.
- The RL agents trained using SigmaRL's observation design strategies achieved training times of under one hour on a single CPU.
- Evaluation results demonstrate that these RL agents can effectively zero-shot generalize, even in completely unseen traffic scenarios.
Read Full Article
20 Likes
Arxiv
1w
273
Image Credit: Arxiv
RL-STaR: Theoretical Analysis of Reinforcement Learning Frameworks for Self-Taught Reasoner
- The theoretical analysis of reinforcement learning frameworks for self-taught reasoner (STaR) in large language models (LLMs) is presented.
- STaR framework uses reinforcement learning to generate reasoning steps and reduce the dependence on human-labeled data.
- The analysis provides a theoretical understanding of the effectiveness of reinforcement learning on chain-of-thought (CoT) reasoning and STaR.
- The framework explores criteria for pre-trained models, policy improvement, convergence, and the robustness of STaR in improving reasoning in LLMs.
Read Full Article
16 Likes
Arxiv
1w
134
Image Credit: Arxiv
UQ of 2D Slab Burner DNS: Surrogates, Uncertainty Propagation, and Parameter Calibration
- This paper focuses on performing a complete uncertainty quantification analysis of a 2D slab burner direct numerical simulation (DNS).
- The study addresses challenges related to developing data-driven surrogate models, propagating parametric uncertainties, and Bayesian calibration of latent heat and chemical reaction parameters.
- Two surrogate models, Gaussian Process (GP) and Hierarchical Multiscale Surrogate (HMS), were constructed using ensemble simulations generated via Latin Hypercube sampling.
- The study emphasizes the importance of surrogate model selection and parameter calibration in quantifying uncertainty in predictions of fuel regression rates in complex combustion systems.
Read Full Article
8 Likes
Arxiv
1w
411
Image Credit: Arxiv
Marconi: Prefix Caching for the Era of Hybrid LLMs
- Hybrid models that combine the language modeling capabilities of Attention layers with the efficiency of Recurrent layers have gained traction for supporting long contexts in Large Language Model serving.
- Marconi is a system that supports efficient prefix caching with Hybrid LLMs.
- Marconi uses novel admission and eviction policies that assess potential cache entries based on recency, reuse likelihood, and compute savings.
- Marconi achieves significantly higher token hit rates compared to state-of-the-art prefix caching systems.
Read Full Article
24 Likes
Arxiv
1w
58
Image Credit: Arxiv
Prototypical Calibrating Ambiguous Samples for Micro-Action Recognition
- Micro-Action Recognition (MAR) has gained attention for its role in non-verbal communication and emotion analysis.
- A novel approach called Prototypical Calibrating Ambiguous Network (PCAN) is proposed to address the ambiguity in MAR.
- PCAN employs a hierarchical action-tree to identify and categorize ambiguous samples into distinct sets.
- Extensive experiments demonstrate the superior performance of PCAN compared to existing approaches.
Read Full Article
3 Likes
Arxiv
1w
386
Image Credit: Arxiv
Tensor Product Attention Is All You Need
- The paper introduces Tensor Product Attention (TPA), a novel attention mechanism that uses tensor decompositions to represent queries, keys, and values compactly.
- TPA significantly reduces the memory overhead during inference by shrinking the size of the key-value (KV) cache.
- Based on TPA, the Tensor ProducT ATTenTion Transformer (T6) is introduced as a new model architecture for sequence modeling.
- T6 outperforms standard Transformer baselines in language modeling tasks, achieving improved model quality and memory efficiency.
Read Full Article
23 Likes
Arxiv
1w
243
Image Credit: Arxiv
Real-time Verification and Refinement of Language Model Text Generation
- Large language models (LLMs) sometimes generate factually incorrect answers, posing a critical challenge.
- The proposed Streaming-VR approach allows real-time verification and refinement of LLM outputs.
- Streaming-VR checks and corrects tokens as they are being generated, ensuring factual accuracy.
- Comprehensive evaluations show that Streaming-VR is an efficient solution compared to prior methods.
Read Full Article
14 Likes
For uninterrupted reading, download the app