Machine Learning

New algorithms conquer adversarial low-rank MDPs, improving regret bounds for unknown transitions and bandit feedback.

GMDI: a novel Bayesian domain adaptation algorithm significantly improves adaptation by dynamically modeling domain indices using Gaussian Mixture Models, outperforming state-of-the-art methods.

BACON: a novel Bayesian adaptive calibration and optimal design algorithm maximizes information gain for data-efficient computer model calibration, significantly outperforming existing methods in synt…

BEEBO: a novel acquisition function for Bayesian Optimization, offering superior explore-exploit balance and handling large batches efficiently, even with noisy data.

This paper introduces ‘bandits with ranking feedback,’ a novel bandit variation providing ranked feedback instead of numerical rewards. It proves instance-dependent cases require superlogarithmic reg…

MAXMINLCB, a novel game-theoretic algorithm, efficiently solves bandit problems with preference feedback over continuous domains, providing anytime-valid, rate-optimal regret guarantees.

The Confidence-Rated Bandits with Abstentions (CBA) algorithm significantly improves reward bounds for prediction with expert advice by strategically leveraging an abstention action.

BAN: a novel backdoor defense using adversarial neuron noise for efficient detection and mitigation.

Longer context in RL boosts generalization but slows down learning; this paper reveals the crucial tradeoff and offers theoretical insights.

B-ary Tree Push-Pull (BTPP) achieves linear speedup for distributed learning on heterogeneous data, significantly outperforming state-of-the-art methods with minimal communication.

AUF-MICNS: A novel sequential method efficiently solves the avoiding undesired future problem by dynamically updating influence relations in non-stationary environments while minimizing action costs.

PASPO: a novel autoregressive policy optimization method for constrained allocation tasks guarantees constraint satisfaction and outperforms existing methods.

New method maximizes AUC under positive distribution shift using only positive and unlabeled training data, and unlabeled test data; improving imbalanced classification.

Attraos: a novel long-term time series forecasting model leveraging chaos theory, significantly outperforms existing methods by utilizing attractor dynamics for efficient and accurate prediction.

This paper rigorously analyzes alpha-divergence variational inference, proving its convergence and providing convergence rates, thereby advancing the theoretical foundations of this increasingly impor…

Reinforcement learning agents achieve emergent cultural accumulation by balancing social and independent learning, outperforming single-lifetime agents.

AROMA: Attentive Reduced Order Model with Attention enhances PDE modeling with local neural fields, offering efficient processing of diverse geometries and superior performance in simulating 1D and 2D…

Evidential deep learning’s uncertainty quantification is unreliable; this paper reveals its limitations, proposes model uncertainty incorporation for improved performance.

Cross-Attention-only Time Series Transformer (CATS) outperforms existing models by removing self-attention, improving long-term forecasting accuracy, and reducing computational cost.

Deep MIL algorithms’ instance-level learnability is theoretically proven, revealing crucial conditions for success and highlighting gaps in existing models.