A prism hierarchy of learning regimes in large linear autoencoders

In three linesTheoretical study of extreme learning regimes in large weight-tied linear autoencoders. Authors identify five fundamental regimes (large-data, small-data, mean-field, narrow-latent, free) associated with faces of a triangular prism. Explicit expressions derived for loss evolution under gradient flow.

Summary generated by Claude — human-verified