neural collapse (NC) is a simple and symmetric phenomenon for deep neural
networks (DNNs) at the terminal phase of training, where the last-layer
features collapse to their class means and form a simplex equiangular tight
frame aligning with the classifier vectors. However, the relatio