Dichotomy of Early and Late Phase Implicit Biases Can Provably Induce Grokking

arXiv:2311.18817v2 Announce Type: replace
Abstract: Recent work by Power et al. (2022) highlighted a surprising "grokking" phenomenon in learning arithmetic tasks: a neural net first "memorizes" the training set, resulting in perfect training accuracy but near-random test accuracy, and after training for sufficiently longer, it suddenly transitions to perfect test accuracy. This paper studies the grokking phenomenon in theoretical setups and shows that it can be induced by a dichotomy of early and late phase implicit biases. Specifically, when training …

abstract accuracy arxiv biases cs.ai cs.lg near neural net power random set tasks test training transitions type work