Spectral Analysis

A theory reframes deep learning via Neural Tangent Kernel and population-risk dynamics, arguing benign overfitting and grokking are explained by spectral flow, and touting a practical one-line Adam tweak that could speed training by up to 5× in some setups.

Empirical evidence that diverse neural networks converge to universal low-dimensional weight subspaces, enabling potential model reuse and more efficient training across tasks.