11
65

Neural Collapse with Cross-Entropy Loss

Abstract

We consider the variational problem of cross-entropy loss with nn feature vectors on a unit hypersphere in Rd\mathbb{R}^d. We prove that when dn1d \geq n - 1, the global minimum is given by the simplex equiangular tight frame, which justifies the neural collapse behavior. We also prove that as nn \rightarrow \infty with fixed dd, the minimizing points will distribute uniformly on the hypersphere and show a connection with the frame potential of Benedetto & Fickus.

View on arXiv
Comments on this paper