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

Download:

Current browse context:

cs.LG

Change to browse by:

References & Citations

DBLP - CS Bibliography

Bookmark

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

Computer Science > Machine Learning

Title: A Probabilistic Representation of DNNs: Bridging Mutual Information and Generalization

Abstract: Recently, Mutual Information (MI) has attracted attention in bounding the generalization error of Deep Neural Networks (DNNs). However, it is intractable to accurately estimate the MI in DNNs, thus most previous works have to relax the MI bound, which in turn weakens the information theoretic explanation for generalization. To address the limitation, this paper introduces a probabilistic representation of DNNs for accurately estimating the MI. Leveraging the proposed MI estimator, we validate the information theoretic explanation for generalization, and derive a tighter generalization bound than the state-of-the-art relaxations.
Comments: To appear in the ICML 2021 Workshop on Theoretic Foundation, Criticism, and Application Trend of Explainable AI
Subjects: Machine Learning (cs.LG); Machine Learning (stat.ML)
Cite as: arXiv:2106.10262 [cs.LG]
  (or arXiv:2106.10262v1 [cs.LG] for this version)

Submission history

From: Xinjie Lan [view email]
[v1] Fri, 18 Jun 2021 17:51:48 GMT (1434kb,D)

Link back to: arXiv, form interface, contact.