neural collapse
A phenomenon observed in the training of neural networks, where representations converge to a shared point in the latent space at the final layers, affecting generalization.
- Consistent Supervised-Unsupervised Alignment for Generalized Category Discovery
- Flatness is Necessary, Neural Collapse is Not: Rethinking Generalization via Grokking
- Neural Collapse in Cumulative Link Models for Ordinal Regression: An Analysis with Unconstrained Feature Model
- Neural Collapse is Globally Optimal in Deep Regularized ResNets and Transformers
- Neural Collapse under Gradient Flow on Shallow ReLU Networks for Orthogonally Separable Data
- The Dual Nature of Plasticity Loss in Deep Continual Learning: Dissection and Mitigation
- The Persistence of Neural Collapse Despite Low-Rank Bias