Groups
Category
Neural Collapse describes what happens at the end of training: the penultimate-layer features of each class concentrate tightly around a class mean.
Grokking is when a model suddenly starts to generalize well long after it has already memorized the training set.