On Gradient-Weight Alignment
Keywords: generalization, gradients, alignment, influence, memorization
TL;DR: Predicting validation performance without validation data using gradient-weight alignment.
Abstract: Evaluating the performance of deep networks against unseen validation data is a crucial step to measure generalization performance.
However, ostensibly neither the training nor validation and test data are ever sufficiently extensive to replicate real-world application.
This works advocates for a change of perspective for evaluating performance of deep networks.
Instead of evaluating against unseen validation data, we propose to rather capture when the model starts to prioritize learning unnecessary or even detrimental specifics of training data instead of general patterns.
While this has been challenging to theoretically derive, we propose *gradient-weight alignment* as an empirical metric to determine performance on unseen data from training information alone.
Our performance measure is efficient and widely applicable, closely tracking validation accuracy during training.
It connects model performance to individual training samples, enabling its use not only for assessing generalization and as an early stopping criterion, but also for offering insights into training dynamics.
Primary Area: applications to computer vision, audio, language, and other modalities
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: 12479
Loading