Stochastic Gradient Descent (SGD) is the workhorse beneath the deep learning revolution. However, SGD is known to reduce its convergence speed due to the plateau phenomenon. Stochastic Natural Gradient Descent (SNGD) was proposed by Amari to resolve that problem by taking benefit of the geometry of the space. Nevertheless, the convergence of SNGD is not guaranteed.
The aim of this article is to modify SNGD to obtain a convergent variant, that we name Convergent SNGD (CSNGD), and test it in a specific toy optimization problem. In particular, we concentrate on the problem of learning a discrete probability distribution.
Based on variable metric convergence results presented by Sunehag et al. , we prove the convergence of CSNGD. Furthermore, we provide experimental results showing that it significantly improves over SGD. We claim that the approach developed in this paper could be extensible to more complex optimization problems making it a promising research line.
IOS Press, Inc.
6751 Tepper Drive
Clifton, VA 20124
Tel.: +1 703 830 6300
Fax: +1 703 830 2300 firstname.lastname@example.org
(Corporate matters and books only) IOS Press c/o Accucoms US, Inc.
For North America Sales and Customer Service
West Point Commons
Lansdale PA 19446
Tel.: +1 866 855 8967
Fax: +1 215 660 5042 email@example.com