Taiji Suzuki

Generalization Bound of Gradient Flow through Training Trajectory and Data-dependent Kernel

From Shortcut to Induction Head: How Data Diversity Shapes Algorithm Selection in Transformers

How Does Label Noise Gradient Descent Improve Generalization in the Low SNR Regime?

Degrees of Freedom for Linear Attention: Distilling Softmax Attention with Optimal Feature Efficiency

State Size Independent Statistical Error Bound for Discrete Diffusion Models

Hessian-guided Perturbed Wasserstein Gradient Flows for Escaping Saddle Points