Skip to Main content Skip to Navigation
New interface
Journal articles

Proximal Methods for Hierarchical Sparse Coding

Rodolphe Jenatton 1, 2, 3 Julien Mairal 1, 2 Guillaume Obozinski 1, 2, 3 Francis Bach 1, 2, 3 
1 WILLOW - Models of visual object recognition and scene understanding
DI-ENS - Département d'informatique - ENS Paris, Inria Paris-Rocquencourt, CNRS - Centre National de la Recherche Scientifique : UMR8548
3 SIERRA - Statistical Machine Learning and Parsimony
DI-ENS - Département d'informatique - ENS Paris, Inria Paris-Rocquencourt, CNRS - Centre National de la Recherche Scientifique : UMR8548
Abstract : Sparse coding consists in representing signals as sparse linear combinations of atoms selected from a dictionary. We consider an extension of this framework where the atoms are further assumed to be embedded in a tree. This is achieved using a recently introduced tree-structured sparse regularization norm, which has proven useful in several applications. This norm leads to regularized problems that are difficult to optimize, and we propose in this paper efficient algorithms for solving them. More precisely, we show that the proximal operator associated with this norm is computable exactly via a dual approach that can be viewed as the composition of elementary proximal operators. Our procedure has a complexity linear, or close to linear, in the number of atoms, and allows the use of accelerated gradient techniques to solve the tree-structured sparse approximation problem at the same computational cost as traditional ones using the L1-norm. Our method is efficient and scales gracefully to millions of variables, which we illustrate in two types of applications: first, we consider fixed hierarchical dictionaries of wavelets to denoise natural images. Then, we apply our optimization tools in the context of dictionary learning, where learned dictionary elements naturally organize in a prespecified arborescent structure, leading to a better performance in reconstruction of natural image patches. When applied to text documents, our method learns hierarchies of topics, thus providing a competitive alternative to probabilistic topic models.
Document type :
Journal articles
Complete list of metadata

Cited literature [74 references]  Display  Hide  Download
Contributor : Rodolphe Jenatton Connect in order to contact the contributor
Submitted on : Tuesday, July 5, 2011 - 10:40:48 AM
Last modification on : Thursday, March 17, 2022 - 10:08:43 AM
Long-term archiving on: : Thursday, October 6, 2011 - 2:21:30 AM


Files produced by the author(s)


  • HAL Id : inria-00516723, version 4
  • ARXIV : 1009.2139



Rodolphe Jenatton, Julien Mairal, Guillaume Obozinski, Francis Bach. Proximal Methods for Hierarchical Sparse Coding. Journal of Machine Learning Research, 2011, 12, pp.2297-2334. ⟨inria-00516723v4⟩



Record views


Files downloads