Machine Learning (ML) Latest News and Trending articles from all top sources only on Techminis

A naukri.com initiative

New

Home

ML News

Arxiv

Image Credit: Arxiv

Variational Supervised Contrastive Learning

Contrastive learning has been effective in shaping representation spaces by pulling similar samples together and pushing dissimilar ones apart.
Variational Supervised Contrastive Learning (VarCon) addresses limitations by reformulating supervised contrastive learning as variational inference over latent class variables.
VarCon maximizes a posterior-weighted evidence lower bound for efficient class-aware matching and control over intra-class dispersion in the embedding space.
Experiments show that VarCon achieves state-of-the-art performance in contrastive learning, clearer decision boundaries, semantic organization, and superior performance in few-shot learning and robustness across various augmentation strategies.

Read Full Article

5 Likes

Arxiv

Image Credit: Arxiv

Evidential Spectrum-Aware Contrastive Learning for OOD Detection in Dynamic Graphs

Recent focus on Out-of-distribution (OOD) detection in dynamic graphs for security-sensitive applications.
Challenges include high bias and variance due to single-point estimation and score homogenization from lack of OOD training data.
Introduction of EviSEC, an OOD detector using Evidential Spectrum-aware Contrastive Learning on dynamic graphs.
Utilizes evidential deep learning and spectrum-aware augmentation to improve OOD detection performance.

Read Full Article

4 Likes

Arxiv

131

Image Credit: Arxiv

Federated In-Context Learning: Iterative Refinement for Improved Answer Quality

Federated In-Context Learning (Fed-ICL) is proposed as a framework to improve answer quality in question-answering tasks without transmitting model parameters.
Fed-ICL enhances In-Context Learning (ICL) through iterative refinement via multi-round interactions between clients and a central server.
It addresses challenges of scarce high-quality examples by leveraging examples stored on client devices.
Extensive experiments on standard QA benchmarks show that Fed-ICL achieves strong performance with low communication costs.

Read Full Article

7 Likes

Arxiv

101

Image Credit: Arxiv

Extending Epistemic Uncertainty Beyond Parameters Would Assist in Designing Reliable LLMs

Current approaches to ensure reliability in large language models (LLMs) are limited to rejecting outputs with high uncertainty to prevent misinformation.
A new strategy of extending epistemic uncertainty beyond parameters is proposed to systematically manage different sources of uncertainty in LLM deployments.
Advocacy for adopting Bayesian Modeling of Experiments as a framework to address uncertainty in LLM deployments for more reliability and transparency.
The proposed framework enables active resolution of uncertainty in LLMs, facilitating contextually appropriate actions such as seeking clarification or refining inputs.

Read Full Article

6 Likes

Discover more

Arxiv

Image Credit: Arxiv

When Style Breaks Safety: Defending Language Models Against Superficial Style Alignment

Large language models (LLMs) can be prompted with specific styles, even in jailbreak queries, but the safety impact of these style patterns is unclear.
In a study evaluating 32 LLMs across seven jailbreak benchmarks, it was found that malicious queries with style patterns increased the attack success rate for almost all models.
ASR inflation correlated with the length of style patterns and the attention LLMs placed on them.
The study revealed that fine-tuning LLMs with specific styles made them more vulnerable to jailbreaks of those same styles. A defense strategy called SafeStyle was proposed to mitigate these risks and consistently outperformed baselines in maintaining LLM safety.

Read Full Article

Arxiv

247

Image Credit: Arxiv

Circumventing Backdoor Space via Weight Symmetry

Deep neural networks are vulnerable to backdoor attacks, but existing defenses require labeled data to purify compromised models, limiting their application beyond supervised learning settings.
A new defense technique called Two-stage Symmetry Connectivity (TSC) has been proposed to address backdoor attacks independently of data format and with only a small fraction of clean samples required.
TSC leverages permutation invariance and quadratic mode connectivity in neural networks to increase the loss on poisoned samples while maintaining bounded clean accuracy.
Experiments show that TSC achieves robust performance comparable to state-of-the-art methods in supervised learning and can also generalize to self-supervised learning frameworks like SimCLR and CLIP.

Read Full Article

14 Likes

Arxiv

245

Image Credit: Arxiv

Chasing Moving Targets with Online Self-Play Reinforcement Learning for Safer Language Models

Researchers propose Self-RedTeam, an online self-play reinforcement learning algorithm for safer language models.
The algorithm involves co-evolution of an attacker and defender agent through continuous interaction in a two-player zero-sum game.
Self-RedTeam enables dynamic co-adaptation and aims to converge to a Nash Equilibrium for reliable safety responses.
Empirical results show that Self-RedTeam uncovers more diverse attacks and achieves higher robustness on safety benchmarks compared to traditional static defender approaches.

Read Full Article

14 Likes

Arxiv

131

Image Credit: Arxiv

Premise Selection for a Lean Hammer

Neural methods are being used to enhance automated reasoning in proof assistants, but integrating these advancements into practical verification workflows remains a challenge.
LeanHammer is introduced as the first domain-general hammer for Lean, leveraging a novel neural premise selection system for a hammer in dependent type theory.
LeanHammer dynamically adapts to user-specific contexts and combines symbolic proof search and reconstruction, aiming to improve productivity in the Lean proof assistant.
Comprehensive evaluations show that LeanHammer with its premise selector can solve 21% more goals compared to existing premise selectors, bridging the gap between neural retrieval and symbolic reasoning in formal verification.

Read Full Article

7 Likes

Arxiv

292

Image Credit: Arxiv

Explicit Preference Optimization: No Need for an Implicit Reward Model

Large language models often fine-tune responses to human preferences using reinforcement learning from human feedback (RLHF).
Direct preference optimization (DPO) and related methods eliminate the need for a separate reward training step by inducing an implicit reward through a reparameterization trick.
While DPO-based objectives have shown success, they can suffer from sub-optimal regularization and counter-intuitive interpolation behaviors due to reparameterizations.
A new explicit preference optimization framework called EXPO has been introduced, requiring no reparameterization and offering regularization factors that avoid pitfalls of DPO variants.

Read Full Article

17 Likes

Arxiv

367

Image Credit: Arxiv

Mind the Gap: Removing the Discretization Gap in Differentiable Logic Gate Networks

Logic gate networks (LGNs) aim to provide efficient solutions for image classification by learning a network of logic gates.
Training LGNs to solve even simple problems like CIFAR-10 can take days to weeks, with almost half of the network remaining unused, leading to a discretization gap.
Researchers have introduced Gumbel noise with a straight-through estimator during training to speed up training, improve neuron utilization, and reduce the discretization gap in LGNs.
This approach results in training networks 4.5 times faster, reducing the discretization gap by 98%, and eliminating the number of unused gates by 100%.

Read Full Article

22 Likes

Arxiv

364

Image Credit: Arxiv

Graph-of-Causal Evolution: Challenging Chain-of-Model for Reasoning

The research introduces a new model called Graph of Causal Evolution (GoCE) to address limitations in the existing Chain-of-Model (CoM) approach.
GoCE utilizes a differentiable and sparse causal adjacency matrix to maintain long-range dependencies and overcome global context flow obstructions between subchains.
Through interventions consistency loss testing and self-evolution gate mechanisms, GoCE achieves a balance between causal structure learning and transformer architecture updating.
Experimental results demonstrate that GoCE outperforms CoM in capturing long-range causal dependencies and enhancing self-evolution capabilities, providing insights for future causal learning research.

Read Full Article

21 Likes

Arxiv

349

Image Credit: Arxiv

Reinforcement Learning via Implicit Imitation Guidance

Researchers propose a method for reinforcement learning that leverages prior data for guiding exploration instead of using explicit imitation learning objectives.
The approach, Data-Guided Noise (DGN), adds noise to the policy based on the prior demonstrations to improve sample efficiency.
DGN achieves significant improvements in reinforcement learning from offline data methods, showing 2-3x enhancement across seven simulated continuous control tasks.
This method aims to overcome the limitations of traditional imitation learning objectives and focuses on exploration guided by prior data for better long-term performance.

Read Full Article

21 Likes

Arxiv

225

Image Credit: Arxiv

Addressing Correlated Latent Exogenous Variables in Debiased Recommender Systems

Recommendation systems face challenges in unbiased learning due to selection bias, leading to distorted user preferences and inaccurate recommendations.
Various debiasing methods have been developed to address this issue, including error imputation, inverse propensity scoring, and doubly robust techniques.
A new learning algorithm based on likelihood maximization has been proposed to address correlated latent exogenous variables in recommender systems, moving away from assuming independence of exogenous variables.
The proposed method handles latent exogenous variables by modeling the data generation process under normality assumptions and using a Monte Carlo algorithm to estimate the likelihood function, showing effectiveness in experiments with synthetic and real-world datasets.

Read Full Article

13 Likes

Arxiv

244

Image Credit: Arxiv

Improving Memory Efficiency for Training KANs via Meta Learning

KANs offer a novel framework for function approximation by replacing traditional neural network weights with learnable univariate functions.
To improve memory efficiency and training costs associated with KANs, a smaller meta-learner named MetaKANs is proposed to generate weights for KANs.
By training KANs and MetaKANs together in an end-to-end differentiable manner, MetaKANs achieve comparable or superior performance with fewer trainable parameters.
Experiments on various tasks show that MetaKANs can enhance parameter efficiency and reduce memory usage, providing a more scalable and cost-effective training method for KANs.

Read Full Article

14 Likes

Arxiv

157

Image Credit: Arxiv

ChemAgent: Enhancing LLMs for Chemistry and Materials Science through Tree-Search Based Tool Learning

A new approach called ChemAgent has been proposed to enhance Large Language Models (LLMs) for chemistry and materials science tasks.
ChemAgent integrates 137 external chemical tools and a dataset curation pipeline to address challenges faced by LLMs, such as outdated pretraining knowledge and lack of specialized chemical expertise.
The approach utilizes a Hierarchical Evolutionary Monte Carlo Tree Search (HE-MCTS) framework for tool planning and execution optimization, enabling step-level fine-tuning of the policy model.
Experimental evaluations show that ChemAgent significantly improves performance in Chemistry QA and discovery tasks, providing a robust solution for integrating specialized tools with LLMs in advanced chemical applications.

Read Full Article

9 Likes

For uninterrupted reading, download the app