After grokking, continued training without weight decay leads to a late generalization collapse, which the authors claim is uniquely tracked by the HTSR alpha exponent.
Marchenko and Leonid Andreevich Pastur
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
citation-role summary
other 1
citation-polarity summary
fields
cs.LG 1years
2025 1verdicts
CONDITIONAL 1roles
other 1polarities
unclear 1representative citing papers
citing papers explorer
-
Grokking and Generalization Collapse: Insights from \texttt{HTSR} theory
After grokking, continued training without weight decay leads to a late generalization collapse, which the authors claim is uniquely tracked by the HTSR alpha exponent.