AI Theory

This paper introduces Expectation Alignment (EAL), a novel framework and interactive algorithm to address reward misspecification in AI, aligning AI behavior with user expectations.

Exogenous Matching learns optimal proposals for efficient counterfactual estimation by transforming variance minimization into conditional distribution learning, outperforming existing methods.

This paper provides the first exact minimax-optimal mechanisms for locally differentially private sampling, applicable across all f-divergences.

Exact Gauss-Newton optimization in deep reversible networks surprisingly reveals poor generalization, despite faster training, challenging existing deep learning optimization theories.

New framework uses rough path theory to enable gradient-based training of SSNNs driven by rough signals, allowing for noise in spike timing and network dynamics.

Chess AI Leela Zero surprisingly uses learned look-ahead, internally representing future optimal moves, significantly improving its strategic decision-making.

Pretrained neural networks surprisingly capture fundamental aspects of human cognition, enabling generalization in image-based learning tasks, as demonstrated by aligning neural network representation…

This study develops a novel two-stage framework for accurately predicting conditional average treatment effects using both observational data and weak instrumental variables, overcoming limitations of…

New consistent estimators precisely track generalization error during robust regression’s iterative model training, enabling optimal stopping iteration for minimized error.

This paper provides novel entrywise error bounds for low-rank kernel matrix approximations, showing how many data points are needed to get statistically consistent results for low-rank approximations.

This paper presents near-optimal algorithms for entropy identity testing, significantly improving Bayesian network testing efficiency.

This paper presents a novel approach to deriving theoretically grounded disentanglement metrics by linking logical definitions to quantitative measures, offering strong theoretical guarantees and easi…

Boosting fair machine learning’s robustness against noisy labels, this work introduces a novel label-spreading method, achieving state-of-the-art worst-group accuracy.

MoE-BiEntIRL: A novel explainable inverse reinforcement learning method enhances GNN robustness against diverse social media attacks by reconstructing attacker policies and generating more robust trai…

GEBM: a novel graph-based energy model for robust GNN uncertainty estimation.

Generative models learn hidden capabilities suddenly during training, which can be explained and predicted using a novel ‘concept space’ framework that analyzes learning dynamics and concept signal.

Homogenized SGD reveals heavy-tailed neural network parameters, offering quantifiable bounds on tail-index and showcasing the interplay between optimization hyperparameters and model generalization.

Elliptical Attention enhances transformers by using a Mahalanobis distance metric, stretching the feature space to focus on contextually relevant information, thus improving robustness and reducing re…

This research introduces LFHSIC, a novel, linear-time independence test that significantly outperforms existing methods, especially for high-dimensional data, by learning optimal Fourier feature pairs…

STREAM-UGS: a novel semi-streaming algorithm for efficient graphlet sampling, enabling fast analysis of massive graphs with limited memory.