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

Download:

Current browse context:

stat.ML

Change to browse by:

References & Citations

Bookmark

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

Statistics > Machine Learning

Title: Generalization Bounds for Noisy Iterative Algorithms Using Properties of Additive Noise Channels

Abstract: Machine learning models trained by different optimization algorithms under different data distributions can exhibit distinct generalization behaviors. In this paper, we analyze the generalization of models trained by noisy iterative algorithms. We derive distribution-dependent generalization bounds by connecting noisy iterative algorithms to additive noise channels found in communication and information theory. Our generalization bounds shed light on several applications, including differentially private stochastic gradient descent (DP-SGD), federated learning, and stochastic gradient Langevin dynamics (SGLD). We demonstrate our bounds through numerical experiments, showing that they can help understand recent empirical observations of the generalization phenomena of neural networks.
Subjects: Machine Learning (stat.ML); Information Theory (cs.IT); Machine Learning (cs.LG)
Cite as: arXiv:2102.02976 [stat.ML]
  (or arXiv:2102.02976v3 [stat.ML] for this version)

Submission history

From: Hao Wang [view email]
[v1] Fri, 5 Feb 2021 03:18:52 GMT (1912kb,D)
[v2] Fri, 29 Oct 2021 18:48:15 GMT (2025kb,D)
[v3] Fri, 31 Dec 2021 21:48:03 GMT (3039kb,D)

Link back to: arXiv, form interface, contact.