AI Theory

Learnware systems enable model reuse; this paper proves RKME specifications protect developers’ training data while enabling effective model identification.

Latent Diffusion Transformers (DiTs) achieve almost-linear time training and inference through low-rank gradient approximations and efficient criteria, overcoming high dimensionality challenges.

This paper presents novel, efficient algorithms and formulations for Sparse Canonical Correlation Analysis (SCCA), a method that improves the interpretability of traditional CCA. SCCA is especially us…

This paper reveals the surprising computational hardness of achieving fairness in low-rank approximation while offering efficient approximation algorithms.

Graph representation learning’s structural vulnerabilities are proven and mitigated via noisy aggregation, revealing crucial privacy-utility trade-offs.

DNNs are powerful but lack the clear semantics of PGMs. This paper innovatively constructs infinite tree-structured PGMs that exactly correspond to DNNs, revealing that DNN forward propagation approxi…

Autoregressively trained transformers surprisingly learn algorithms during pretraining, enabling in-context learning; this paper reveals when and why this ‘mesa-optimization’ happens.

Unlocking the scaling secrets of structured state-space models, this research identifies novel scaling rules for improved stability, generalization, and hyperparameter transferability, revolutionizing…

New algorithms achieve near-optimal differentially private U-statistic estimation, significantly improving accuracy over existing methods.

This paper presents novel measures quantifying data easiness for DP subspace estimation, supporting them with improved upper and lower bounds and a practical algorithm.

Adam optimizer achieves near-optimal convergence in non-convex scenarios with unbounded gradients and relaxed noise assumptions, improving its theoretical understanding and practical application.

DGES, a novel framework, efficiently detects & handles deterministic relations in causal discovery, enhancing accuracy and scalability for real-world applications.

This paper develops novel finite-sample bounds for off-policy linear treatment effect estimation with adaptively collected data, proposing online learning algorithms to improve estimation accuracy and…

RNNs trained on dual-agent path integration develop distinct internal representations compared to single-agent models, exhibiting weaker grid cell responses and enhanced border/band cell activity, wit…

SAGD-IV: a novel functional stochastic gradient descent algorithm for stable nonparametric instrumental variable regression, excelling in handling binary outcomes and various loss functions.

A novel semi Forward-Backward Euler scheme provides convergence guarantees for non-geodesically-convex optimization in Wasserstein space, advancing both sampling and optimization.

This paper reveals how a one-layer transformer’s training converges for next-token prediction, showing sub-linear convergence for both layers and shedding light on its surprising generalization abilit…

Near-optimal algorithm for private resource allocation is introduced, achieving improved accuracy and privacy guarantees.

Meta-learning and differential privacy combine to enable accurate, well-calibrated private regression, even with limited data, via the novel DPConvCNP model.

This paper reveals efficient stochastic bandit algorithms for maximizing M-concave functions and proves NP-hardness for adversarial full-information settings.