Is Memorization Actually Necessary for Generalization?
Keywords: Memorization
TL;DR: Memorization approximation methods have high errors, and memorization has an insignificant impact on accuracy
Abstract: Memorization is the ability of deep models to associate training data with seemingly random labels. Even though memorization may not align with a model's ability to generalize, recent work by~\citet{feldman2020longtail} has demonstrated that memorization is in fact \textit{necessary} for generalization. However, upon closer inspection, we find that their methodology has three limitations. First, the definition of memorization is imprecise, leading to contradictory results. Second, their proposed algorithm used for \textit{approximating} the leave-one-out test (the gold standard for calculating memorization scores) suffers from a high approximation error. Three, the authors induce a distribution shift when calculating marginal utility, leading to flawed results. Having accounted for these errors, we re-evaluate the role of memorization on generalization. To do so, we track how memorization changes at different levels of generalization (test accuracy). We control model generalization by training 19 different combinations of models, datasets, and training optimizations. We find that memorization and generalization are \textit{strongly} negatively correlated (Pearson -0.997): As one decreases, the other increases. This shows that memorization is not necessary for generalization, as otherwise, the correlation would have been positive. In light of these findings, future researchers are encouraged to design techniques that can accurately approximate memorization scores.
Primary Area: alignment, fairness, safety, privacy, and societal considerations
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.
Reciprocal Reviewing: I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6.
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: 7677
Loading