AI Theory

New framework directly controls neural network sensitivity by precisely parameterizing overall bi-Lipschitzness, offering improved robustness and generalization.

BLOCC, a novel first-order algorithm, efficiently solves bilevel optimization problems with coupled constraints, offering improved scalability and convergence for machine learning applications.

Estimating heterogeneous treatment effects (CATE) under unmeasured confounding is revolutionized by a novel non-parametric direct learning approach using instrumental variables, offering efficient and…

A novel neural network efficiently answers arbitrary Most Probable Explanation (MPE) queries in large probabilistic models, eliminating the need for slow inference algorithms.

Huber loss minimization ensures accurate and robust mean estimation under user-level differential privacy, especially for imbalanced datasets and heavy-tailed distributions.

Researchers introduce a generalized neural tangent kernel for analyzing surrogate gradient learning in neural networks with non-differentiable activation functions, providing a strong theoretical foun…

Revolutionizing probabilistic inference, PLIA₁ uses tensor operations and FFT to scale integer arithmetic, achieving orders-of-magnitude speedup in inference and learning times.

This study provides a unified theory for kernel ridge regression’s learning curve, improving existing bounds and validating the Gaussian Equivalence Property under minimal assumptions.

This paper introduces a compositional framework for algebraic circuits, deriving novel tractability conditions for compositional inference queries and unifying existing results.

A new combinatorial algorithm dramatically speeds up semi-discrete optimal transport calculations, offering an efficient solution for large datasets and higher dimensions.

Debunking a common myth, this paper proves that AUPRC is not superior to AUROC for imbalanced datasets, and in fact, can worsen algorithmic bias.

Solved a long-standing open problem: Factorized ADABOOST.MH now has a proven convergence rate!

Researchers discovered four distinct compute-optimal phases for training neural networks, offering new predictions for resource-efficient large model training.

2D-OOB: a novel framework for jointly attributing data values to individual features, enabling fine-grained outlier detection and improved model performance.