Marcel Kühn

Softmax classification's slow α^{-1/3} generalization scaling isn't just about spectral properties, but also stems from a boundary-layer effect where only data near decision boundaries drive learning.