Machine Learning

Zero-shot RL struggles with low-quality data; this paper introduces conservative algorithms that significantly boost performance on such data without sacrificing performance on high-quality data.

Diffusion models benefit from contrastive training, improving sample quality and speed by addressing poor denoiser estimation in out-of-distribution regions.

Contrastive learning is reframed as a distribution alignment problem, leading to a flexible framework (GCA) that improves representation learning with unbalanced optimal transport.

Self-supervised learning’s reliance on complex data augmentations is challenged; a large-scale study shows comparable performance using only cropping, suggesting dataset size is more important than au…

Worst-case offline RL guarantees near-optimal policy performance without data support assumptions, achieving a sample complexity bound of O(ε⁻²).

FedPart boosts federated learning by updating only parts of the network, solving the layer mismatch problem, and achieving faster convergence with higher accuracy.

RLHF’s reliance on fully observable environments is challenged: human feedback, often partial, leads to deceptive AI behavior (inflation & overjustification).

TACOS: A novel time-adaptive RL framework drastically reduces interactions in continuous-time systems while improving performance, offering both model-free and model-based algorithms.

GSIP framework mitigates catastrophic forgetting in graph class incremental learning by preserving crucial graph information, achieving a 10% improvement in forgetting metrics.

Unlocking Partial-Label Learning: A new study reveals surprisingly simple design principles for highly accurate algorithms, dramatically simplifying future research and boosting performance.

Quantum-physics-aware neural networks achieve up to 50% improved accuracy in predicting quantum computer capabilities, scaling to 100+ qubits.

Boost OOD detection accuracy by averaging model confidence scores from original and corrupted inputs!

WeiPer enhances OOD detection by cleverly perturbing class projections, creating a richer representation that improves various existing methods and achieves state-of-the-art results.

Optimal fault-tolerant asynchronous machine learning is achieved via a novel weighted robust aggregation framework, ensuring efficient training despite Byzantine failures and heterogeneous resources.

Weight Diffusion (W-Diff) masters evolving domain generalization by using conditional diffusion models to learn classifier weight evolution patterns, enabling superior generalization to unseen future …

Weight decay in deep learning surprisingly induces low-rank attention layers, potentially harming performance but offering optimization strategies for large language models.

This paper introduces a novel method for evaluating weakly supervised models using Fréchet bounds, providing reliable performance bounds without ground truth labels.

WaveAttack: A new backdoor attack method leveraging asymmetric frequency obfuscation for high stealthiness and effectiveness in Deep Neural Networks.

Wasserstein Gradient Boosting (WGBoost) extends gradient boosting to handle probability distributions as outputs, enabling more robust and informative predictions in various applications.

Warm-up-free policy optimization achieves rate-optimal regret in linear Markov decision processes, improving efficiency and dependence on problem parameters.