Grokking arises from gradual amplification of a Fourier-based circuit in the weights followed by removal of memorizing components.
Dropout: a simple way to prevent neural networks from overfitting
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.LG 1years
2023 1verdicts
ACCEPT 1representative citing papers
citing papers explorer
-
Progress measures for grokking via mechanistic interpretability
Grokking arises from gradual amplification of a Fourier-based circuit in the weights followed by removal of memorizing components.