Perturbation Analysis of Neural Collapse

Tom Tirer, Haoxiang Huang, Jonathan Niles-Weed

Training deep neural networks for classification often includes minimizing the training loss beyond the zero training error point. In this phase of training, a "neural collapse" be…