Generalization Gradient Descent

ICLR 2025 Conference Submission11378 Authors

27 Sept 2024 (modified: 26 Nov 2024)ICLR 2025 Conference SubmissionEveryoneRevisionsBibTeXCC BY 4.0
Keywords: out-of-distribution, generalization problem, gradient descent
TL;DR: n/a
Abstract: We propose a new framework for evaluating the relationship between features and generalization via a theoretical analysis of the out-of-distribution (OOD) generalization problem, in which we simultaneously use two mathematical methods: a generalization ratio that quantitatively characterizes the degree of generalization, and a generalization decision process (GDP) that formalizes the relationship of loss between seen and unseen domains. By combining the concepts of informativeness and variation in the generalization ratio, we intuitively associate them with OOD problems to derive the generalization inequality. We then introduce it to the GDP to select the best loss from seen domains to gradient descent for backpropagation. In the case where the classifier is defined by fully connected neural network, the entire system is trained with backpropagation. There is no need for any model selection criterion or operating on gradients during training. Experiments demonstrate the potential of the framework through qualitative and quantitative evaluation of the generalization ability.
Supplementary Material: zip
Primary Area: learning theory
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: 11378
Loading