Understanding Grokking: Insights from Neural Network Robustness
Keywords: Grokking, Representation Learning, Training Dynamics
Abstract: Recently, an interesting phenomenon called grokking has gained much attention, where generalization occurs long after the models have initially overfitted the training data. We try to understand this seemingly strange phenomenon through the robustness of the neural network. From a robustness perspective, we show that the usually observed decreasing of $l_2$ weight norm of the neural network is theoretically connected to the occurrence of grokking. Therefore, we propose to use perturbation-based methods to enhance robustness and speed up the generalization process. Furthermore, we show that the speed-up of generalization when using our proposed method can be explained by learning the commutative law, a necessary condition when the model groks on the test dataset. In addition, we empirically observe that
$l_2$ norm correlates with grokking on the test data not in a timely way and then propose new metrics based on robustness that correlate better with the grokking phenomenon.
Primary Area: interpretability and explainable AI
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 9713
Loading