What is neural collapse in deep networks?

Neural collapse describes a phenomenon where features and class representations converge to structured geometry.

In late-stage training, features cluster tightly around class means forming simplex ETF geometry, improving classification margin.