DocumentCode :
81552
Title :
Direct Optimization of the Dictionary Learning Problem
Author :
Rakotomamonjy, Alain
Author_Institution :
LITIS EA 4108 UFR Sci., Univ. de Rouen, Rouen, France
Volume :
61
Issue :
22
fYear :
2013
fDate :
Nov.15, 2013
Firstpage :
5495
Lastpage :
5506
Abstract :
A novel way of solving the dictionary learning problem is proposed in this paper. It is based on a so-called direct optimization as it avoids the usual technique which consists in alternatively optimizing the coefficients of a sparse decomposition and in optimizing dictionary atoms. The algorithm we advocate simply performs a joint proximal gradient descent step over the dictionary atoms and the coefficient matrix. After having derived the algorithm, we also provided in-depth discussions on how the stepsizes of the proximal gradient descent have been chosen. In addition, we uncover the connection between our direct approach and the alternating optimization method for dictionary learning. We have shown that it can be applied to a broader class of non-convex optimization problems than the dictionary learning one. As such, we have denoted the algorithm as a one-step block-coordinate proximal gradient descent. The main advantage of our novel algorithm is that, as suggested by our simulation study, it is more efficient than alternating optimization algorithms.
Keywords :
concave programming; gradient methods; learning (artificial intelligence); sparse matrices; block coordinate proximal gradient descent; coefficient matrix; dictionary atoms; dictionary learning problem; direct optimization; nonconvex optimization problem; one step proximal gradient descent; sparse decomposition; Algorithm design and analysis; Approximation methods; Dictionaries; Joints; Materials; Optimization; Sparse matrices; Dictionary learning; non-convex proximal; one-step block-coordinate descent;
fLanguage :
English
Journal_Title :
Signal Processing, IEEE Transactions on
Publisher :
ieee
ISSN :
1053-587X
Type :
jour
DOI :
10.1109/TSP.2013.2278158
Filename :
6578208
Link To Document :
بازگشت