A Commutator Framework for Selective Spectral Alignment in Deep Neural Networks

Kaj Nyström

We develop a finite-width geometric framework describing how learned feature geometries are organized, transported, and selectively aligned in deep neural networks. Incompatibility…