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: Generalization Properties and Implicit Regularization for Multiple Passes SGM

Abstract: We study the generalization properties of stochastic gradient methods for learning with convex loss functions and linearly parameterized functions. We show that, in the absence of penalizations or constraints, the stability and approximation properties of the algorithm can be controlled by tuning either the step-size or the number of passes over the data. In this view, these parameters can be seen to control a form of implicit regularization. Numerical results complement the theoretical findings.
Comments: 26 pages, 4 figures. To appear in ICML 2016
Subjects: Machine Learning (cs.LG); Machine Learning (stat.ML)
Cite as: arXiv:1605.08375 [cs.LG]
  (or arXiv:1605.08375v1 [cs.LG] for this version)

Submission history

From: Junhong Lin [view email]
[v1] Thu, 26 May 2016 17:37:51 GMT (54kb,D)

Link back to: arXiv, form interface, contact.