We gratefully acknowledge support from
the Simons Foundation and member institutions.
Full-text links:


Current browse context:


Change to browse by:

References & Citations

DBLP - CS Bibliography


(what is this?)
CiteULike logo BibSonomy logo Mendeley logo del.icio.us logo Digg logo Reddit logo ScienceWISE logo

Computer Science > Machine Learning

Title: Regularizing Neural Networks via Adversarial Model Perturbation

Abstract: Effective regularization techniques are highly desired in deep learning for alleviating overfitting and improving generalization. This work proposes a new regularization scheme, based on the understanding that the flat local minima of the empirical risk cause the model to generalize better. This scheme is referred to as adversarial model perturbation (AMP), where instead of directly minimizing the empirical risk, an alternative "AMP loss" is minimized via SGD. Specifically, the AMP loss is obtained from the empirical risk by applying the "worst" norm-bounded perturbation on each point in the parameter space. Comparing with most existing regularization schemes, AMP has strong theoretical justifications, in that minimizing the AMP loss can be shown theoretically to favour flat local minima of the empirical risk. Extensive experiments on various modern deep architectures establish AMP as a new state of the art among regularization schemes. Our code is available at this https URL
Comments: 16 pages, 13 figures, accepted to CVPR2021
Subjects: Machine Learning (cs.LG); Artificial Intelligence (cs.AI); Machine Learning (stat.ML)
Cite as: arXiv:2010.04925 [cs.LG]
  (or arXiv:2010.04925v4 [cs.LG] for this version)

Submission history

From: Yaowei Zheng [view email]
[v1] Sat, 10 Oct 2020 07:01:00 GMT (6102kb,D)
[v2] Fri, 12 Mar 2021 14:02:46 GMT (6425kb,D)
[v3] Thu, 22 Apr 2021 13:53:27 GMT (6737kb,D)
[v4] Fri, 7 May 2021 04:26:26 GMT (6737kb,D)

Link back to: arXiv, form interface, contact.