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

Download:

Current browse context:

math.ST

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

Mathematics > Statistics Theory

Title: The loss landscape of deep linear neural networks: a second-order analysis

Abstract: We study the optimization landscape of deep linear neural networks with the square loss. It is known that, under weak assumptions, there are no spurious local minima and no local maxima. However, the existence and diversity of non-strict saddle points, which can play a role in first-order algorithms' dynamics, have only been lightly studied. We go a step further with a full analysis of the optimization landscape at order 2. We characterize, among all critical points, which are global minimizers, strict saddle points, and non-strict saddle points. We enumerate all the associated critical values. The characterization is simple, involves conditions on the ranks of partial matrix products, and sheds some light on global convergence or implicit regularization that have been proved or observed when optimizing linear neural networks. In passing, we provide an explicit parameterization of the set of all global minimizers and exhibit large sets of strict and non-strict saddle points.
Subjects: Statistics Theory (math.ST); Machine Learning (cs.LG)
Cite as: arXiv:2107.13289 [math.ST]
  (or arXiv:2107.13289v2 [math.ST] for this version)

Submission history

From: El Mehdi Achour [view email]
[v1] Wed, 28 Jul 2021 11:33:18 GMT (223kb,D)
[v2] Fri, 11 Mar 2022 08:52:06 GMT (267kb,D)

Link back to: arXiv, form interface, contact.