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


Current browse context:


Change to browse by:

References & Citations


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

Mathematics > Statistics Theory

Title: Localization of VC Classes: Beyond Local Rademacher Complexities

Abstract: In this paper we introduce an alternative localization approach for binary classification that leads to a novel complexity measure: fixed points of the local empirical entropy. We show that this complexity measure gives a tight control over complexity in the upper bounds. Our results are accompanied by a novel minimax lower bound that involves the same quantity. In particular, we practically answer the question of optimality of ERM under bounded noise for general VC classes.
Comments: 28 pages, accepted version
Subjects: Statistics Theory (math.ST)
Cite as: arXiv:1606.00922 [math.ST]
  (or arXiv:1606.00922v3 [math.ST] for this version)

Submission history

From: Nikita Zhivotovskiy [view email]
[v1] Thu, 2 Jun 2016 22:16:20 GMT (41kb)
[v2] Mon, 1 Aug 2016 09:32:40 GMT (42kb)
[v3] Sun, 17 Dec 2017 21:39:27 GMT (43kb)

Link back to: arXiv, form interface, contact.