Neural Collapse with Cross-Entropy Loss

Abstract
We consider the variational problem of cross-entropy loss with feature vectors on a unit hypersphere in . We prove that when , the global minimum is given by the simplex equiangular tight frame, which justifies the neural collapse behavior. We also prove that as with fixed , the minimizing points will distribute uniformly on the hypersphere and show a connection with the frame potential of Benedetto & Fickus.
View on arXivComments on this paper