A rationale from frequency perspective for grokking in training neural network

TL;DR

This paper offers an empirical frequency-based explanation for grokking in neural networks, showing that networks first learn less salient frequencies before generalizing, across synthetic and real datasets.

Contribution

It introduces a novel frequency perspective to explain grokking, highlighting the role of frequency dynamics during training.

Findings

Networks initially learn less salient frequency components.

Frequency dynamics are consistent across synthetic and real datasets.

Provides new insights into the mechanisms behind grokking.

Abstract

Grokking is the phenomenon where neural networks NNs initially fit the training data and later generalize to the test data during training. In this paper, we empirically provide a frequency perspective to explain the emergence of this phenomenon in NNs. The core insight is that the networks initially learn the less salient frequency components present in the test data. We observe this phenomenon across both synthetic and real datasets, offering a novel viewpoint for elucidating the grokking phenomenon by characterizing it through the lens of frequency dynamics during the training process. Our empirical frequency-based analysis sheds new light on understanding the grokking phenomenon and its underlying mechanisms.