DocumentCode :
6946
Title :
The Information Geometry of Mirror Descent
Author :
Raskutti, Garvesh ; Mukherjee, Sayan
Author_Institution :
Depts. of Stat. & Comput. Sci., Univ. of Wisconsin-Madison, Madison, WI, USA
Volume :
61
Issue :
3
fYear :
2015
fDate :
Mar-15
Firstpage :
1451
Lastpage :
1457
Abstract :
We prove the equivalence of two online learning algorithms: 1) mirror descent and 2) natural gradient descent. Both mirror descent and natural gradient descent are generalizations of online gradient descent when the parameter of interest lies on a non-Euclidean manifold. Natural gradient descent selects the steepest descent along a Riemannian manifold by multiplying the standard gradient by the inverse of the metric tensor. Mirror descent induces non-Euclidean structure by solving iterative optimization problems using different proximity functions. In this paper, we prove that mirror descent induced by Bregman divergence proximity functions is equivalent to the natural gradient descent algorithm on the dual Riemannian manifold. We use techniques from convex analysis and connections between Riemannian manifolds, Bregman divergences, and convexity to prove this result. This equivalence between natural gradient descent and mirror descent, implies that: 1) mirror descent is the steepest descent direction along the Riemannian manifold corresponding to the choice of Bregman divergence and 2) mirror descent with log-likelihood loss applied to parameter estimation in exponential families asymptotically achieves the classical Cramér-Rao lower bound.
Keywords :
gradient methods; learning (artificial intelligence); Bregman divergence proximity function; Cramer-Rao lower bound; Riemannian manifold; convex analysis; information geometry; iterative optimization problems; log-likelihood loss; metric tensor; mirror descent algorithm; natural gradient descent algorithm; non-Euclidean manifold; online learning algorithm; parameter estimation; Geometry; Manifolds; Measurement; Mirrors; Standards; Tensile stress; Vectors; Differential geometry; Information geometry; Mirror descent; Natural gradient; Online learning; differential geometry; information geometry; mirror descent; natural gradient;
fLanguage :
English
Journal_Title :
Information Theory, IEEE Transactions on
Publisher :
ieee
ISSN :
0018-9448
Type :
jour
DOI :
10.1109/TIT.2015.2388583
Filename :
7004065
Link To Document :
بازگشت