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

A naukri.com initiative

New

Home

ML News

Arxiv

111

Image Credit: Arxiv

L-VAE: Variational Auto-Encoder with Learnable Beta for Disentangled Representation

Researchers have introduced a new model called Learnable VAE (L-VAE) that learns a disentangled representation along with cost function hyperparameters.
L-VAE extends eta-VAE by dynamically adjusting the hyperparameter eta and learning the weights of loss function terms for better control over disentanglement and reconstruction losses.
The L-VAE model simultaneously learns the weights of loss terms and model parameters, with an added regularization term to avoid bias towards either reconstruction or disentanglement losses.
Experimental results demonstrate that L-VAE achieves a good balance between reconstruction accuracy and disentangled latent dimensions, outperforming or matching other VAE variants on various datasets.

Read Full Article

6 Likes

Arxiv

210

Image Credit: Arxiv

A Matrix Variational Auto-Encoder for Variant Effect Prediction in Pharmacogenes

Researchers propose a transformer-based matrix variational auto-encoder (matVAE) for variant effect prediction in pharmacogenes, dealing with low evolutionary pressure in pharmacogenomics.
The model, matVAE-MSA, outperforms the DeepSequence model in zero-shot prediction on deep mutational scanning (DMS) datasets, requiring fewer parameters and less computation at inference time.
Comparison with a model trained on DMS data, matENC-DMS, shows the latter performs better on supervised prediction tasks.
Incorporating AlphaFold-generated structures into the transformer model boosts performance, paving the way for leveraging DMS datasets to enhance variant effect prediction without significant loss in predictive accuracy.

Read Full Article

12 Likes

Arxiv

226

Image Credit: Arxiv

Medical Data Pecking: A Context-Aware Approach for Automated Quality Evaluation of Structured Medical Data

The paper introduces the Medical Data Pecking approach for evaluating the quality of structured medical data used in Electronic Health Records (EHRs) for research and AI training.
The approach utilizes unit testing and coverage concepts from software engineering to identify data quality concerns and includes the Medical Data Pecking Tool (MDPT).
MDPT was tested on three datasets and successfully identified non-aligned or non-conforming data issues, demonstrating its effectiveness in improving data quality for research purposes.
The approach incorporates external medical knowledge to enhance context-sensitive data quality testing and aims to address challenges in data quality assessment for research purposes.

Read Full Article

13 Likes

Arxiv

348

Image Credit: Arxiv

On Efficient Bayesian Exploration in Model-Based Reinforcement Learning

This work focuses on data-efficient exploration in reinforcement learning by examining information-theoretic approaches to intrinsic motivation.
Exploration bonuses targeting epistemic uncertainty are studied, showing that they signal information gains and converge to zero as the agent learns about the environment.
The analysis provides formal guarantees for these approaches and discusses practical approximations through different models like sparse variational Gaussian Processes and Deep Ensemble models.
The framework called Predictive Trajectory Sampling with Bayesian Exploration (PTS-BE) is introduced, combining model-based planning with information-theoretic bonuses to achieve sample-efficient deep exploration, outperforming other baselines in various environments in the empirical evaluation.

Read Full Article

20 Likes

Discover more

Arxiv

203

Image Credit: Arxiv

OmniDraft: A Cross-vocabulary, Online Adaptive Drafter for On-device Speculative Decoding

OmniDraft is a unified framework designed to address challenges in online deployment settings related to cross-vocabulary mismatch and latency improvements in speculative decoding.
OmniDraft allows a single draft model to work with any target model and dynamically adapt to user data by utilizing an online n-gram cache and hybrid distillation fine-tuning.
This framework is ideal for on-device Large Language Model (LLM) applications focusing on model cost, efficiency, and user customization.
OmniDraft showcases its efficacy through online learning tasks in math reasoning, coding, and text generation, demonstrating compatibility with various target models and providing speed enhancements.

Read Full Article

12 Likes

Arxiv

Image Credit: Arxiv

Guided Generation for Developable Antibodies

A computational framework for optimizing antibody sequences for favorable developability has been introduced.
The framework includes a guided discrete diffusion model trained on natural paired heavy- and light-chain sequences from the Observed Antibody Space (OAS).
Integration of Soft Value-based Decoding in Diffusion (SVDD) Module helps bias sampling towards biophysically viable candidates without compromising naturalness.
The model shows significant enrichment in predicted developability scores over unguided baselines and enables the ML-driven pipeline for designing antibodies.

Read Full Article

1 Like

Arxiv

291

Image Credit: Arxiv

Multi-Agent Reinforcement Learning for Dynamic Pricing in Supply Chains: Benchmarking Strategic Agent Behaviours under Realistically Simulated Market Conditions

A study explores how Multi-Agent Reinforcement Learning (MARL) can enhance dynamic pricing strategies in supply chains by considering strategic interactions among market actors.
The research evaluates three MARL algorithms (MADDPG, MADQN, and QMIX) against static rule-based baselines in a simulated environment using real e-commerce transaction data.
Results indicate that rule-based agents achieve high fairness and price stability but lack competitive dynamics, while MADQN displays aggressive pricing behavior with high volatility and low fairness.
MADDPG offers a balanced approach by supporting market competition, maintaining high fairness, and stable pricing, suggesting that MARL introduces emergent strategic behavior in dynamic pricing scenarios.

Read Full Article

17 Likes

Arxiv

310

Image Credit: Arxiv

Fluid Democracy in Federated Data Aggregation

Researchers have proposed the use of consensus-based protocols to determine a subset of clients with the most useful model weights in federated learning to reduce data transfer costs.
A new fluid democracy protocol named viscous-retained democracy has been introduced, offering better performance than traditional methods like 1p1v (FedAvg) without allowing influence accumulation.
The study also addresses weaknesses of fluid democracy protocols from an adversarial perspective, highlighting vulnerabilities related to topology and number of adversaries. A new algorithm, FedVRD, aims to limit adversarial impact while optimizing cost through delegation topology.

Read Full Article

18 Likes

Arxiv

318

Image Credit: Arxiv

A Forget-and-Grow Strategy for Deep Reinforcement Learning Scaling in Continuous Control

Deep reinforcement learning for continuous control has made significant progress, but current methods often suffer from primacy bias.
A new algorithm called Forget and Grow (FoG) is proposed to address this issue by incorporating mechanisms inspired by neuroscience.
FoG includes Experience Replay Decay (ER Decay) to forget early experiences and Network Expansion to grow neural capacity.
Empirical results on various continuous control benchmarks show that FoG outperforms existing deep RL algorithms like BRO, SimBa, and TD-MPC2.

Read Full Article

19 Likes

Arxiv

Image Credit: Arxiv

Hierarchical Multi-Label Contrastive Learning for Protein-Protein Interaction Prediction Across Organisms

Researchers have proposed a hierarchical contrastive framework, HIPPO, for protein-protein interaction (PPI) prediction involving protein sequences and hierarchical attributes.
HIPPO incorporates hierarchical contrastive loss functions to capture structured relationships among functional classes of proteins and adaptively incorporates domain and family knowledge.
Experiments show that HIPPO outperforms existing methods in PPI prediction, demonstrating robustness in low-data scenarios and strong transferability to other species without retraining.
The proposed framework's hierarchical feature fusion is crucial for capturing conserved interaction determinants, enabling reliable PPI prediction even in less characterized organisms.

Read Full Article

3 Likes

Arxiv

Image Credit: Arxiv

Fast and Simplex: 2-Simplicial Attention in Triton

A new study explores the concept of 2-simplicial attention in Triton as a means to improve token efficiency in models.
The research examines how the 2-simplicial Transformer architecture, which uses trilinear functions through an efficient Triton kernel implementation, can outperform traditional Transformers in tasks related to mathematics, coding, reasoning, and logic within a fixed token budget.
The study suggests that the 2-simplicial Transformer changes the scaling laws for knowledge and reasoning tasks when compared to dot product attention, showcasing better token efficiency.
The findings emphasize the importance of designing architectures that prioritize token efficiency, especially as large language models rely on massive internet-scale datasets.

Read Full Article

2 Likes

Arxiv

Image Credit: Arxiv

Understanding and Improving Length Generalization in Recurrent Models

Recurrent models like state space models and linear attention have gained popularity due to their linear complexity in sequence length.
These models, theoretically able to process long sequences, sometimes struggle to generalize beyond their training context lengths, a phenomenon known as failing to length generalize.
An analysis supports the unexplored states hypothesis, suggesting that models struggle with length generalization when exposed to a limited subset of attainable states during training.
Simple training interventions like initializing states with Gaussian noise or final states of different input sequences have shown to enable length generalization for significantly longer sequences, offering an efficient way to improve performance in long context tasks for recurrent models.

Read Full Article

Arxiv

Image Credit: Arxiv

In-Training Multicalibrated Survival Analysis for Healthcare via Constrained Optimization

Survival analysis in healthcare is crucial for modeling the relationship between an individual's characteristics and the time of an event like death.
Existing survival models are often poorly calibrated for minority subpopulations, leading to potential errors in clinical decisions.
The GRADUATE model proposed in the study tackles this by achieving multicalibration at the subpopulation level through constrained optimization.
Empirical comparisons show GRADUATE's effectiveness in achieving calibration and discrimination balance compared to existing baselines.

Read Full Article

4 Likes

Arxiv

Image Credit: Arxiv

ExPO: Unlocking Hard Reasoning with Self-Explanation-Guided Reinforcement Learning

Recent advances in large language models have been driven by reinforcement learning (RL)-style post-training to improve reasoning by optimizing model outputs based on reward signals.
A new Self-Explanation Policy Optimization (ExPO) framework has been introduced to address limitations in refining model knowledge and enabling exploration beyond current output distributions.
ExPO generates positive samples through conditioning on the ground-truth answer, facilitating efficient exploration and guiding the model to produce better reasoning trajectories.
Experiments demonstrate that ExPO outperforms expert-demonstration-based methods in challenging settings, enhancing learning efficiency and final performance on reasoning benchmarks.

Read Full Article

1 Like

Arxiv

145

Image Credit: Arxiv

LLM-Driven Treatment Effect Estimation Under Inference Time Text Confounding

Estimating treatment effects is crucial in medicine for personalized decision-making but faces challenges due to discrepancies in data available at training and inference times.
An inference time text confounding problem arises where confounders are fully observed during training but only partially available through text at inference, leading to biased estimates.
A novel framework is proposed in this work to address the inference time text confounding, leveraging large language models and a custom doubly robust learner to mitigate biases.
Experiments conducted demonstrate the effectiveness of the framework in real-world applications for estimating treatment effects under inference time text confounding.

Read Full Article

8 Likes

For uninterrupted reading, download the app