AI Theory

Generative models often produce invalid outputs; this work shows that ensuring validity is easier than expected when using log-loss and carefully selecting model classes and data distributions.

Researchers dissected attention paths in Transformers using statistical mechanics, revealing a task-relevant kernel combination mechanism boosting generalization performance.

Cross-environment Intra-class Alignment (CIA) and its label-free variant, CIA-LRA, significantly improve node-level OOD generalization on graphs by aligning representations and eliminating spurious fe…

This paper introduces graphical criteria and an algorithm for disentangling causal factors from heterogeneous data in non-Markovian settings, advancing causal representation learning.

Researchers surpass the 1/e barrier in submodular maximization with novel combinatorial algorithms!

New sub-optimality bounds for gradient descent leverage directional smoothness, a localized gradient variation measure, achieving tighter convergence guarantees and adapting to optimization paths.

D-PBEMO: A novel framework for preference-based multi-objective optimization using clustering-based stochastic dueling bandits to directly leverage human feedback, improving efficiency and managing co…

Dimension-free private mean estimation is achieved for anisotropic data, breaking the curse of dimensionality in privacy-preserving high-dimensional analysis.

This work delivers dimension-free deterministic equivalents for random feature regression, revealing sharp excess error rates and scaling laws.

Diffusion models are certifiably robust classifiers due to their inherent O(1) Lipschitzness, a property further enhanced by generalizing to noisy data, achieving over 80% certified robustness on CIFA…

DiffHammer unveils weaknesses in diffusion-based adversarial defenses by introducing a novel attack bypassing existing evaluation limitations, leading to more robust security solutions.

Tighter differential privacy (RDP) guarantees for DP-SGD with fixed-size minibatches are achieved, improving private deep learning model training.

Differentially private set representations achieve optimal privacy-utility tradeoffs with exponentially smaller error than prior histogram methods.

This paper presents DP-Nash-VI, a novel algorithm ensuring trajectory-wise privacy in multi-agent reinforcement learning, achieving near-optimal regret bounds under both joint and local differential p…

This paper presents new, nearly optimal differentially private algorithms for handling sparse gradients, significantly improving efficiency and scalability in large embedding models.

This paper introduces a novel differentially private graph diffusion framework ensuring edge-level privacy, significantly improving utility-privacy trade-offs for personalized PageRank computation.

First differentially private algorithm for testing equivalence between continuous distributions, enabling privacy-preserving comparisons of sensitive data.

Differentially private scalable kernel learning is achieved via a novel DP K-means Nyström method, enabling efficient and accurate model training for general kernels while safeguarding privacy.

This research introduces a novel plug-and-play module that efficiently integrates prior partial order constraints into differentiable structure learning, significantly improving structure recovery qua…

Diffeomorphic interpolation boosts topological optimization by transforming sparse gradients into smooth vector fields, enabling efficient large-scale point cloud optimization and black-box autoencode…