phenomenon · first seen 26 Oct 2026
effective rank compression of weight matrices
Also: rank compression
The decrease in effective rank of neural network weight matrices during training, linked to phase transitions in learning.
Mentions per week
PapersCodeDiscussion
Earliest sources
- Subspace Crystallization in Grokking: Rank Compression Tracks the Phase TransitionPapers · 26 Oct 2026 · Shlok Mehendale, Paras Chopra
Latest sources
- Subspace Crystallization in Grokking: Rank Compression Tracks the Phase TransitionPapers · 26 Oct 2026 · Shlok Mehendale, Paras Chopra