Learning in deep networks is proposed to consist of a fast memorization phase followed by a slow compression phase that drives generalization, with mutual information between hidden layers and inputs as a progress measure.
Grokking and the Geometry of Circuit Formation,
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
hep-th 1years
2025 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
A Two-Phase Perspective on Deep Learning Dynamics
Learning in deep networks is proposed to consist of a fast memorization phase followed by a slow compression phase that drives generalization, with mutual information between hidden layers and inputs as a progress measure.