🏢 University of Tokyo

Wide two-layer neural networks can generalize well from mislabeled adversarial examples because adversarial perturbations surprisingly contain sufficient class-specific features.

Vision Transformers (ViTs) generalize surprisingly well, even when overfitting training data; this work provides the first theoretical explanation by characterizing the optimization dynamics of ViTs a…

A mean-field theory explains Fourier Neural Operator (FNO) behavior, linking expressivity to trainability by identifying ordered and chaotic phases that correspond to vanishing or exploding gradients,…

Linear probing then fine-tuning (LP-FT) significantly improves language model fine-tuning; this paper uses Neural Tangent Kernel (NTK) theory to explain why.

Transformers excel at in-context learning by leveraging minimax-optimal nonparametric learning, achieving near-optimal risk with sufficient pretraining data diversity.

LoTNext framework tackles human mobility prediction’s long-tail problem by using graph and loss adjustments to improve the accuracy of predicting less-visited locations.

Risk-sensitive control is recast as inference using Rényi divergence, yielding new algorithms and revealing equivalences between seemingly disparate methods.

This paper provides novel algorithms and matching lower bounds for multi-armed bandits with expert advice and contextual linear bandits, resolving open questions and advancing theoretical understandin…

LLM agents effectively generate realistic personal mobility patterns using semantically rich data.

Deep learning framework integrating GNNs and neural ODEs precisely estimates non-reciprocal two-body interactions in mixed-species collective motion, accurately replicating both individual and collect…

Improving LLM alignment, this paper introduces soft preference labels & geometric averaging in Direct Preference Optimization, consistently improving performance on standard benchmarks.

Learn to project, solve faster! This paper introduces data-driven projections for solving high-dimensional linear programs, proving theoretical guarantees and demonstrating significant improvements in…

One-shot animatable head avatar reconstruction is achieved using a novel dual-lifting method that generates 3D Gaussians from a single image, enabling real-time expression control and rendering with s…

This paper introduces a novel approach for fast rates in online convex optimization by exploiting the curvature of feasible sets, achieving logarithmic regret bounds under specific conditions.

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…

AI-generated images contaminate online continual learning datasets, hindering performance. A new method, ESRM, leverages entropy and real/synthetic similarity maximization to select high-quality data…

Koodos: a novel Koopman operator-driven framework that tackles Continuous Temporal Domain Generalization (CTDG) by modeling continuous data dynamics and learning model evolution across irregular time …

ADOPT, a novel adaptive gradient method, achieves optimal convergence rates without restrictive assumptions, unlike Adam, significantly improving deep learning optimization.

A new adaptive learning rate for FTRL achieves minimax regret of O(T²/³) in online learning, improving existing best-of-both-worlds algorithms for various hard problems.

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