IEEE transactions on neural networks and learning systems
Progressive Feedforward Collapse of ResNet Training.
Sicong Wang, Kuo Gai, Shihua Zhang
Published: 202510.1109/TNNLS.2025.3620798
Abstract
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 (ETF) aligning with th…
Preview only. Read the full abstract at the source