A Multicover Approach to Neural Networks Sample Complexity
Keywords: Sample Complexity, Covering Numbers, Neural Networks
TL;DR: We generalize the notion of covering numbers, and use it to derive state of art results for the sample complexity of neural networks.
Abstract: Covering numbers are central to estimating sample complexity. Alas, standard techniques for bounding covering numbers fail in estimating the covering numbers of many classes of neural networks. We introduce a generalization of covers, called {\em multicovers}, which are covers w.r.t. many metrics simultaneously.
Contrary to standard covering numbers, multicovering numbers behave better with the layer-wise structure in neural networks. We utilize this property to recover a recent result of \citet{ADL2019} who defined a new notion called Approximate Description Length (ADL) to establish tight bounds on the sample complexity of networks with weights of bounded Frobenius norm. We also show that ADL and multicovering numbers are closely related.
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.
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: 8108
Loading