The Implicit Bias of Depth: From Neural Collapse to Softmax Codes

In three linesTheoretical study showing that network depth induces implicit low-rank bias, promoting alternatives to neural collapse. Analysis of unregularized deep UFM (unconstrained feature model) training dynamics reveals how depth favors softmax codes over classical structured geometries.

Summary generated by Claude — human-verified