A rationale from frequency perspective for grokking in training neural network
This offers a novel viewpoint for understanding grokking, a phenomenon in neural network training, but is incremental as it builds on existing observations without broad practical impact.
The paper tackles the problem of explaining the grokking phenomenon in neural networks, where models initially fit training data and later generalize to test data, by providing an empirical frequency perspective that shows networks learn less salient frequency components first, observed across synthetic and real datasets.
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.