AI Theory

This study solves a key challenge in neural network learning, deriving a closed-form expression for the Bayes-optimal test error of extensive-width networks with quadratic activation functions from qu…

Randomness-efficient algorithms are developed for online decision making, requiring only 2log n random bits and achieving near-optimal competitiveness for metrical task systems.

This paper reveals the optimal mistake bounds for online multiclass classification under bandit feedback, showing the cost of limited feedback is at most O(k) times higher than full information, where…

This paper introduces BSR, a novel banded square root matrix factorization for differentially private model training. Unlike existing methods, BSR avoids computationally expensive optimization, enabli…

Despite their brittleness, continuous attractors remain functionally robust analog memory models due to persistent slow manifolds surviving bifurcations, enabling accurate approximation and generaliza…

B-cosification: cheaply transform any pre-trained deep neural network into an inherently interpretable model.

This paper revolutionizes AI alignment by applying social choice theory axioms to RLHF, exposing flaws in existing methods and proposing novel, axiomatically guaranteed reward learning rules.

Deep Neural Collapse (DNC) explained via Average Gradient Outer Product (AGOP).

AI models retraining with model-annotated data incorporating human strategic responses can lead to unexpected outcomes, potentially reducing the proportion of agents with positive labels over time, wh…

This study presents a novel single-step outlier rectification method using optimal transport with a concave cost function, surpassing the limitations of conventional two-stage approaches by jointly op…

MC-EIF automates efficient statistical estimation for high-dimensional models, integrating seamlessly with existing differentiable probabilistic programming systems and achieving optimal convergence r…

More sophisticated autobidders surprisingly worsen online auction efficiency; a fine-grained analysis reveals that less powerful, uniform bidders lead to better market outcomes.

New metrics audit label privatization, revealing differentially private schemes often outperform heuristic methods in the privacy-utility tradeoff.

Auditing local explanations is surprisingly hard: proving explanation trustworthiness requires far more data than previously thought, especially in high dimensions, challenging current AI explainabil…

Boosting machine learning model accuracy in privacy-preserving applications, this research introduces novel noise calibration methods directly targeting desired attack risk levels, bypassing conventio…

Node Injection-based Fairness Attack (NIFA) reveals GNNs’ vulnerability to realistic fairness attacks by injecting a small percentage of nodes, significantly undermining fairness even in fairness-awar…

High-degree representations significantly boost the expressiveness of E(3)-equivariant GNNs, overcoming limitations of lower-degree models on symmetric structures, as demonstrated theoretically and em…

Graph Neural Networks (GNNs) learn optimal approximation algorithms for combinatorial optimization problems, achieving high-quality solutions for Max-Cut, Min-Vertex-Cover, and Max-3-SAT, while also p…

Random-order stream data necessitates efficient top eigenvector approximation; this paper presents novel algorithms with improved space complexity, achieving near-optimal bounds.

Latent Mutual Information (LMI) approximation accurately estimates mutual information in high-dimensional data using low-dimensional learned representations, solving a critical problem in various scie…