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

Download:

Current browse context:

math-ph

Change to browse by:

References & Citations

Bookmark

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

Mathematical Physics

Title: On Random Matrices Arising in Deep Neural Networks. Gaussian Case

Authors: Leonid Pastur
Abstract: The paper deals with distribution of singular values of product of random matrices arising in the analysis of deep neural networks. The matrices resemble the product analogs of the sample covariance matrices, however, an important difference is that the population covariance matrices, which are assumed to be non-random in the standard setting of statistics and random matrix theory, are now random, moreover, are certain functions of random data matrices. The problem has been considered in recent work [21] by using the techniques of free probability theory. Since, however, free probability theory deals with population matrices which are independent of the data matrices, its applicability in this case requires an additional justification. We present this justification by using a version of the standard techniques of random matrix theory under the assumption that the entries of data matrices are independent Gaussian random variables. In the subsequent paper [18] we extend our results to the case where the entries of data matrices are just independent identically distributed random variables with several finite moments. This, in particular, extends the property of the so-called macroscopic universality on the considered random matrices.
Comments: 28 pages
Subjects: Mathematical Physics (math-ph); Machine Learning (stat.ML)
MSC classes: Primary 15B52, Secondary 92B20
Cite as: arXiv:2001.06188 [math-ph]
  (or arXiv:2001.06188v2 [math-ph] for this version)

Submission history

From: Leonid Pastur [view email]
[v1] Fri, 17 Jan 2020 08:30:57 GMT (30kb)
[v2] Tue, 7 Apr 2020 15:11:44 GMT (31kb)

Link back to: arXiv, form interface, contact.