Page 1

Proximal Methods for Sparse Hierarchical Dictionary Learning

Rodolphe Jenatton1

Julien Mairal1

Guillaume Obozinski

Francis Bach

RODOLPHE.JENATTON@INRIA.FR

JULIEN.MAIRAL@INRIA.FR

GUILLAUME.OBOZINSKI@INRIA.FR

FRANCIS.BACH@INRIA.FR

INRIA - WILLOW Project, Laboratoire d’Informatique de l’Ecole Normale Sup´ erieure (INRIA/ENS/CNRS UMR 8548)

23, avenue d’Italie, 75214 Paris. France

Abstract

We propose to combine two approaches for mod-

eling data admitting sparse representations: on

the one hand, dictionary learning has proven ef-

fective for various signal processing tasks. On

the other hand, recent work on structured spar-

sity provides a natural framework for modeling

dependencies between dictionary elements. We

thus consider a tree-structured sparse regulariza-

tion to learn dictionaries embedded in a hierar-

chy.The involved proximal operator is com-

putable exactly via a primal-dual method, allow-

ing the use of accelerated gradient techniques.

Experiments show that for natural image patches,

learned dictionary elements organize themselves

in such a hierarchical structure, leading to an im-

proved performance for restoration tasks. When

applied to text documents, our method learns hi-

erarchies of topics, thus providing a competitive

alternative to probabilistic topic models.

1. Introduction

Learned sparse representations, initially introduced by

Olshausen & Field (1997), have been the focus of much

research in machine learning and signal processing, lead-

ing notably to state-of-the-art algorithms for several prob-

lems in image processing (Elad & Aharon, 2006). Model-

ing signals as a linear combination of a few “basis” vectors

offers more flexibility than decompositions based on prin-

cipal component analysis and its variants. Indeed, sparsity

allows for overcomplete dictionaries, whose number of ba-

sis vectors are greater than the original signal dimension.

1Contributed equally.

Appearing in Proceedings of the 27thInternational Conference

on Machine Learning, Haifa, Israel, 2010. Copyright 2010 by the

author(s)/owner(s).

As far as we know, while much attention has been given

to efficiently solving the corresponding optimization prob-

lem (Lee et al., 2007; Mairal et al., 2010), there are few at-

tempts in the literature to make the model richer by adding

structure between dictionary elements (Bengio et al., 2009;

Kavukcuoglu et al., 2009). We propose to use recent work

on structured sparsity (Zhao et al., 2009; Jenatton et al.,

2009; Kim & Xing, 2009) to embed the dictionary ele-

ments in a hierarchy.

Hierarchies of latent variables, typically used in neural net-

works and deep learning architectures (see Bengio, 2009

and references therein) have emerged as a natural structure

in several applications, notably to model text documents.

Indeed, inthecontextoftopicmodels(Blei et al.,2003), hi-

erarchical models using Bayesian non-parametric methods

have been proposed by Blei et al. (2010). Quite recently,

hierarchies have also been considered in the context of ker-

nel methods (Bach, 2009). Structured sparsity has been

used to regularize dictionary elements by Jenatton et al.

(2010), but to the best of our knowledge, it has never been

used to model dependencies between them.

This paper makes three contributions:

• We propose to use a structured sparse regularization to

learn a dictionary embedded in a tree.

• We show that the proximal operator for a tree-structured

sparse regularization can be computed exactly in a finite

number of operations using a primal-dual approach, with a

complexity linear, or close to linear, in the number of vari-

ables. Accelerated gradient methods (e.g., Nesterov, 2007)

can then be applied to solve tree-structured sparse decom-

position problems, which may be useful in other uses of

tree-structured norms (Kim & Xing, 2009; Bach, 2009).

• Our method establishes a bridge between dictionary

learning for sparse coding and hierarchical topic models

(Blei et al., 2010), which builds upon the interpretation of

topic models as multinomial PCA (Buntine, 2002), and can

learn similar hierarchies of topics. See Section 5 for a dis-

cussion.

Page 2

Proximal Methods for Sparse Hierarchical Dictionary Learning

2. Problem Statement

2.1. Dictionary Learning

Let us consider a set X = [x1,...,xn] ∈ Rm×nof n

signals of dimension m. Dictionary learning is a matrix

factorization problem that aims to represent these signals

as linear combinations of dictionary elements, denoted here

by the columns of a matrix D = [d1,...,dp] ∈ Rm×p.

More precisely, the dictionary D is learned along with a

matrix of decomposition coefficients A = [α1,...,αn] ∈

Rp×n, so that xi≈ Dαifor every signal xi, as measured

by any convex loss, e.g., the square loss in this paper.

While learning simultaneously D and A, one may want

to encode specific prior knowledge about the task at

hand, such as, for example, the positivity of the de-

composition (Lee & Seung, 1999), or the sparsity of A

(Olshausen & Field, 1997; Lee et al., 2007; Mairal et al.,

2010). This leads to penalizing or constraining (D,A) and

results in the following formulation:

min

D∈D,A∈A

1

n

n

?

i=1

?1

2||xi− Dαi||2

2+ λΩ(αi)

?

,

(1)

where A and D denote two convex sets and Ω is a regu-

larization term, usually a norm, whose effect is controlled

by the regularization parameter λ > 0. Note that D is as-

sumed to be bounded to avoid any degenerate solutions of

Eq. (1). For instance, the standard sparse coding formula-

tion takes Ω to be the ℓ1norm, D to be the set of matrices

in Rm×pwhose columns are in the unit ball of the ℓ2norm,

with A = Rp×n(Lee et al., 2007; Mairal et al., 2010).

However, this classical setting treats each dictionary ele-

ment independently from the others, and does not exploit

possible relationships between them. We address this po-

tential limitation of the ℓ1norm by embedding the dictio-

nary in a tree structure, through a hierarchical norm intro-

duced by Zhao et al. (2009) and Bach (2009), which we

now present.

2.2. Hierarchical Sparsity-Inducing Norms

We organize the dictionary elements in a rooted-tree

T composed of p nodes, one for each dictionary ele-

ment dj, j ∈ {1,...,p}. We will identify these indices

j in {1,...,p} and the nodes of T . We want to exploit the

structure of T in the following sense: the decomposition

of any vector x can involve a dictionary element djonly if

the ancestors of djin T are themselves part of the decom-

position. Equivalently, one can say that when a dictionary

element djis not involved in the decomposition of a vec-

tor x then its descendants in T should not be part of the

decomposition. While these two views are equivalent, the

latter leads to an intuitive penalization term.

Figure 1. Left: example of a tree-structured set of groups G

(dashed contours in red), corresponding to a tree T = {1,...,6}

(in black).Right, example of a sparsity pattern: the groups

{2,4},{4} and {6} are set to zero, so that the corresponding

nodes (in gray) that form subtrees of T are removed. The remain-

ing nonzero variables {1,3,5} are such that, if a node is selected,

the same goes for all its ancestors.

To obtain models with the desired property, one considers

for all j in T , the group gj ⊆ {1,...,p} of dictionary

elements that only contains j and all its descendants, and

penalizes the number of such groups that are involved in

the decomposition of x (a group being “involved in the de-

composition” meaning here that at least one of its dictio-

nary element is part of the decomposition). We call G this

set of groups (Figure 1).

While this penalization is non-convex, a convex proxy has

been introduced by Zhao et al. (2009) and was further con-

sidered by Bach (2009) and Kim & Xing (2009) in the con-

text of regression. For any vector α ∈ Rp, let us define

Ω(α)

△=

?

g∈G

wg?α|g?,

where α|g is the vector of size p whose coordinates are

equal to those of α for indices in the set g, and 0 other-

wise2. ?.? stands either for the ℓ∞or ℓ2norm, and (wg)g∈G

denotes some positive weights3. As analyzed by Zhao et al.

(2009), when penalizing by Ω, some of the vectors α|gare

set to zero for some g ∈ G. Therefore, the components

of α corresponding to some entire subtrees of T are set to

zero, which is exactly the desired effect (Figure 1).

Note that even though we have presented for simplicity rea-

sons this hierarchical norm in the context of a single tree

with a single element at each node, it can be extended eas-

ily to the case of forests of trees, and/or trees containing

several dictionary elements at each node. More generally,

this formulation can be extended with the notion of tree-

structured groups, which we now present.

2Note the difference with the notation αg, which is often used

in works on structured sparsity, where αgis a vector of size |g|.

3For a complete definition of Ω for any ℓq norm, a discussion

of the choice of q, and a strategy for choosing the weights wg, see

(Zhao et al., 2009; Kim & Xing, 2009).

Page 3

Proximal Methods for Sparse Hierarchical Dictionary Learning

Definition 1 (Tree-structured set of groups.)

groups G = {g}g∈G is said to be tree-structured in

{1,...,p}, if?

(g∩h ?= ∅) ⇒ (g ⊆ h or h ⊆ g). For such a set of groups,

there exists a (non-unique) total order relation ? such that:

A set of

g∈Gg = {1,...,p} and for all g,h ∈ G,

g ? h ⇒

?g ⊆ h or g ∩ h = ∅?.

Sparse hierarchical norms having been introduced, we now

address the optimization dealing with such norms.

3. Optimization

Optimization for dictionary learning has already been in-

tensively studied, and a typical scheme alternating between

the variables D and A = [α1,...,αn], i.e., minimizing

over one while keeping the other one fixed, yields good re-

sults in general (Lee et al., 2007). The main difficulty of

our problem lies essentially in the optimization of the vec-

tors αi, i ∈ {1,...,n} for D fixed, since n may be large,

and since it requires to deal with the nonsmooth regulariza-

tion term Ω. The optimization of the dictionary D (for A

fixed) is in general easier, as discussed in Section 3.5.

Within the context of regression, several optimization

methods to cope with Ω have already been proposed. A

boosting-like technique with a path-following strategy is

used by Zhao et al. (2009).

reweighted least-squares scheme when ?.? is the ℓ2norm.

The same approach is considered by Bach (2009), but built

upon an active set strategy. In this paper, we propose to

perform the updates of the vectors αibased on a proximal

approach which we now introduce.

Kim & Xing (2009) uses a

3.1. Proximal Operator for the Norm Ω

Proximal methods have drawn increasing attention in the

machine learning community (e.g., Ji & Ye, 2009 and ref-

erences therein), especially because of their convergence

rates (optimal for the class of first-order techniques) and

their ability to deal with large nonsmooth convex problems

(e.g., Nesterov, 2007; Beck & Teboulle, 2009). In a nut-

shell, these methods can be seen as a natural extension of

gradient-based techniques when the objective function to

minimize has a nonsmooth part. In our context, when the

dictionary D is fixed and A = Rp, we minimize for each

signal x the following convex nonsmooth objective func-

tion w.r.t. α ∈ Rp:

f(α) + λΩ(α),

where f(α) stands for the data-fitting term1

At each iteration of the proximal algorithm, f is linearized

around the current estimate ˆ α, and the current value of α

is updated as the solution of the proximal problem:

2?x − Dα?2

2.

min

α∈Rpf(ˆ α)+(α − ˆ α)⊤∇f(ˆ α) + λΩ(α) +L

2?α − ˆ α?2

2.

The quadratic term keeps the update in a neighborhood

where f is close to its linear approximation, and L > 0

is a parameter. This problem can be rewritten as,

min

α∈Rp

1

2?α −?ˆ α −1

L∇f(ˆ α)??2

2+λ

LΩ(α).

Solving efficiently and exactly this problem is crucial to en-

joy the fast convergence rates of proximal methods. In ad-

dition, when the nonsmooth term Ω is not present, the pre-

vious proximal problem exactly leads to the standard gra-

dient update rule. More generally, the proximal operator

associated with our regularization term λΩ, is the function

that maps a vector u ∈ Rpto the (unique) solution of

min

v∈Rp

1

2?u − v?2

2+ λΩ(v).

(2)

In the simpler setting where G is the set of singletons, Ω is

the ℓ1norm, and the proximal operator is the (elementwise)

soft-thresholding operator uj ?→ sign(uj)max(|uj| −

λ,0), j ∈ {1,...,p}. Similarly, when the groups in G

form a partition of the set of variables, we have a group

Lasso like penalty, and the proximal operator can be com-

puted in closed-form (see Bengio et al., 2009 and refer-

ences therein). This is a priori not possible anymore as

soon as some groups in G overlap, which is always the case

in our hierarchical setting with tree-structured groups.

3.2. Primal-Dual Interpretation

We now show that Eq. (2) can be solved with a primal-

dual approach. The procedure solves a dual formulation of

Eq. (2) involving the dual norm4of ?.?, denoted by ?.?∗,

and defined by ?κ?∗= max?z?≤1z⊤κ for any vector κ

in Rp. The formulation is described in the following lemma

that relies on conic duality (Boyd & Vandenberghe, 2004):

Lemma 1 (Dual of the proximal problem)

Let u ∈ Rpand let us consider the problem

max

ξ∈Rp×|G|−1

2

??u −

?

g∈G

ξg?2

2− ?u?2

2

?

s.t. ∀g ∈ G, ?ξg?∗≤ λwgand ξg

j= 0 if j / ∈ g,

(3)

where ξ = (ξg)g∈Gand ξg

the vector ξgin Rp. Then, problems (2) and (3) are dual to

each other and strong duality holds. In addition, the pair

of primal-dual variables {v,ξ} is optimal if and only if ξ

is a feasible point of the optimization problem (3), and

jdenotes the j-th coordinate of

v = u −?

∀g ∈ G,

g∈Gξg,

v⊤

or v|g= 0.

?

|gξg= ?v|g??ξg?∗and ?ξg?∗= λwg,

4It is easy to show that the dual norm of the ℓ2norm is the ℓ2

norm itself. The dual norm of the ℓ∞is the ℓ1norm.

Page 4

Proximal Methods for Sparse Hierarchical Dictionary Learning

For space limitation reasons, we have omitted all the de-

tailed proofs from this section. They will be available in

a longer version of this paper. Note that we focus here on

specific tree-structured groups, but the previous lemma is

valid regardless of the nature of G.

ThestructureofthedualproblemofEq.(3), i.e., thesepara-

bility of the (convex) constraints for each vector ξg, g ∈ G,

makes itpossibletouseblockcoordinateascent(Bertsekas,

1999). Such a procedure is presented in Algorithm 1. It op-

timizes sequentially Eq. (3) with respect to the variable ξg,

while keeping fixed the other variables ξh, for h ?= g. It is

easy to see from Eq. (3) that such an update for a group g

in G amounts to the orthogonal projection of the vector

u|g−?

norm ?.?∗. We denote this projection Π∗

h?=gξh

|gonto the ball of radius λwg of the dual

λwg.

Algorithm 1 Block coordinate ascent in the dual

Inputs: u ∈ Rpand set of groups G.

Outputs: (v,ξ) (primal-dual solutions).

Initialization: v = u, ξ = 0.

while ( maximum number of iterations not reached ) do

for g ∈ G do

v ← u −?

ξg← Π∗

end for

end while

v ← u −?

h?=gξh.

λwg(v|g).

g∈Gξg.

3.3. Convergence in One Pass

In general, Algorithm 1 is not guaranteed to solve exactly

Eq. (2) in a finite number of iterations. However, when ?.?

is the ℓ2or ℓ∞norm, and provided that the groups in G

are appropriately ordered, we now prove that only one pass

of Algorithm 1, i.e., only one iteration over all groups, is

sufficient to obtain the exact solution of Eq. (2). This result

constitutes the main technical contribution of the paper.

Beforestatingthisresult, weneedtointroduceakeylemma

that shows that, given two nested groups g,h such that g ⊆

h ⊆ {1,...,p}, if ξgis updated before ξhin Algorithm 1,

then the optimality condition for ξgis not perturbed by the

update of ξh.

Lemma 2 (Projections with nested groups)

Let ?.? denote either the ℓ2or ℓ∞norm, and g and h be

two nested groups—that is, g ⊆ h ⊆ {1,...,p}. Let v be a

vector in Rp, and let us consider the successive projections

κg △= Π∗

tg(v|g) and κh△= Π∗

th(v|h− κg)

with tg,th> 0. Then, we have as well κg=Π∗

tg(v|g−κh

|g).

The previous lemma establishes the convergence in one

pass of Algorithm 1 in the case where G contains only two

nested groups g ⊆ h, provided that ξgis computed be-

fore ξh. In the following proposition, this lemma is ex-

tended to general tree-structured sets of groups G:

Proposition 1 (Convergence in one pass) Suppose

the groups in G are ordered according to ? and that

the norm ?.? is either the ℓ2or ℓ∞norm5. Then, after

initializing ξ to 0, one pass of Algorithm 1 with the order ?

gives the solution of Eq. (2).

that

Proof sketch.

ceed by induction, by showing that we keep the optimality

conditions of Eq. (3) satisfied after each update in Algo-

rithm 1. The induction is initialized by the leaves. Once

the induction reaches the last group, i.e., after one com-

pletepassover G, thedualvariableξ satisfiestheoptimality

conditions for Eq. (3), which implies that {v,ξ} is optimal.

Since strong duality holds, v is the solution of Eq. (2).

The proof relies on Lemma 2. We pro-

3.4. Efficient Computation of the Proximal Operator

Since one pass of Algorithm 1 involves |G| = p projections

onto the ball of the dual norm (respectively the ℓ2and the

ℓ1norms) of vectors in Rp, a naive implementation leads to

a complexity in O(p2), since each of these projections can

be obtained in O(p) operations (see Mairal et al., 2010, and

references therein). However, the primal solution v, which

is the quantity of interest, can be obtained with a better

complexity, as exposed below:

Proposition 2 (Complexity of the procedure)

i) For the ℓ2norm, the primal solution v of Algorithm 1 can

be obtained in O(p) operations.

ii) For the ℓ∞norm, v can be obtained in O(pd) opera-

tions, where d is the depth of the tree.

The linear complexity in the ℓ2norm case results from a

recursive implementation. It exploits the fact that each pro-

jection amounts to a scaling, whose factor can be found

without explicitly performing the full projection at each it-

eration. As for the ℓ∞norm, since all the groups at a depth

k ∈ {1,...,d} do not overlap, the cost for performing all

the projections at this depth k is O(p), which leads to a

total complexity of O(dp). Note that d could depend on

p as well. For instance, in an unbalanced case, the worse

case could be d = O(p), in a balanced tree, one could have

d = O(log(p)). In practice, the structures we have consid-

ered are relatively flat, with a depth not exceeding d = 5.

Details will be provided in a longer version of this paper.

3.5. Learning the Dictionary

We alternate between the updates of D and A, minimizing

over one while keeping the other variable fixed.

5Interestingly, we have observed that this was not true in gen-

eral when ?.? is an ℓqnorm, for q ?= 2 and q ?= ∞.

Page 5

Proximal Methods for Sparse Hierarchical Dictionary Learning

Updating D.

inversion free procedure of Mairal et al. (2010) for up-

dating the dictionary. This method consists in a block-

coordinate scheme over the columns of D. Specifically,

we assume that the domain set D has the form

We have chosen to follow the matrix-

Dµ

△= {D ∈ Rm×p, µ?dj?1+ (1 − µ)?dj?2

2≤ 1}, (4)

or D+

particular domain sets is motivated by the experiments of

Section 4. For natural image patches, the dictionary ele-

ments are usually constrained to be in the unit ℓ2norm ball

(i.e., D = D0), while for topic modeling, the dictionary

elements are distributions of words and therefore belong

to the simplex (i.e., D = D+

1). The update of each dic-

tionary element amounts to performing a Euclidean pro-

jection, which can be computed efficiently (Mairal et al.,

2010). Concerning the stopping criterion, we follow the

strategy from the same authors and go over the columns

of D only a few times, typically 5 in our experiments.

µ

△= Dµ∩Rm×p

+

, with µ ∈ [0,1]. The choice for these

Updating the vectors αi.

the columns of A is built upon the results derived in Sec-

tion 3.2. We have shown that the proximal operator from

Eq. (2) can be computed exactly and efficiently. It makes

it possible to use fast proximal techniques, suited to non-

smooth convex optimization.

The procedure for updating

Specifically, we have tried the accelerated scheme from

both Nesterov (2007) and Beck & Teboulle (2009), and fi-

nally opted for the latter since, for a comparable level of

precision, fewer calls of the proximal operator are required.

The procedure from Beck & Teboulle (2009) basically fol-

lows Section 3.1, except that the proximal operator is not

directly applied on the current estimate, but on an auxiliary

sequence of points that linearly combines past estimates.

This algorithm has an optimal convergence rate in the class

of first-order techniques, and also allows warm restarts,

which is crucial in our alternating scheme. Furthermore,

positivity constraints can be added on the domain of A,

by noticing that for our norm Ω and any u ∈ Rp, adding

these constraints when computing the proximal operator is

equivalent to solving

min

v∈Rp

1

2?[u]+− v?2

2+ λΩ(v),

with ([u]+)j

decompositions to model text corpora in Section 4.

△= max{uj,0}. We will indeed use positive

Finally, we monitor the convergence of the algorithm by

checking the relative decrease in the cost function. We also

investigated the derivation of a duality gap, but this implies

the computation of the dual norm Ω∗for which no closed-

form is available; computing approximations of Ω∗based

on bounds from Jenatton et al. (2009) turned out to be too

slow for our experiments.

4. Experiments

4.1. Natural Image Patches

This experiment studies whether a hierarchical structure

can help dictionaries for denoising natural image patches,

and in which noise regime the potential gain is significant.

We aim at reconstructing corrupted patches from a test set,

after having learned dictionaries on a training set of non-

corrupted patches. Thoughnottypicalinmachinelearning,

this setting is reasonable in the context of images, where

lots of non-corrupted patches are easily available.6

We have extracted 100,000 patches of size m = 8×8 pix-

els from the Berkeley segmentation database of natural im-

ages7, which contains a high variability of scenes. We have

then split this dataset into a training set Xtr, a validation

set Xval, and a test set Xte, respectively of size 50,000,

25,000, and 25,000 patches. All the patches are centered

and normalized to have unit ℓ2norm.

For the first experiment, the dictionary D is learned on Xtr

using the formulation of Eq. (1), with µ = 0 for Dµde-

fined in Eq. (4). The validation and test sets are corrupted

by removing a certain percentage of pixels, the task being

to reconstruct the missing pixels from the known pixels.

We thus introduce for each element x of the validation/test

set, a vector ˜ x, equal to x for the known pixel values and

0 otherwise. In the same way, we define˜D as the matrix

equal to D, except for the rows corresponding to missing

pixel values, which are set to 0. By decomposing ˜ x on˜D,

we obtain a sparse code α, and the estimate of the recon-

structed patch is defined as Dα. Note that this procedure

assumes that we know which pixel is missing and which is

not for every element x.

The parameters of the experiment are the regularization pa-

rameter λtrused during the train step, the regularization

parameter λteused during the validation/test step, and the

structure of the tree. For every reported result, these pa-

rameters have been selected by taking the ones offering

the best performance on the validation set, before report-

ing any result from the test set. The values for the reg-

ularization parameters λtr,λtewere tested on a logarith-

mic scale {2−10,2−9,...,22}, and then further refined on

a finer logarithmic scale of factor 2−1/4. For simplicity

reasons, we have chosen arbitrarily to use the ℓ∞-norm

in the structured norm Ω, with all the weights equal to

one. We have tested 21 balanced tree structures of depth

3 and 4, with different branching factors p1,p2,...,pd−1,

where d is the depth of the tree and pk, k ∈ {1,...,d − 1}

6Note that we study the ability of the model to reconstruct

independent patches, and additional work is required to apply our

framework to a full image processing task, where patches usually

overlap (Elad & Aharon, 2006).

7www.eecs.berkeley.edu/Research/Projects/CS/vision/bsds/

Page 6

Proximal Methods for Sparse Hierarchical Dictionary Learning

Table 1. Quantitative results of the reconstruction task on natural

image patches. First row: percentage of missing pixels. Second

and third rows: mean square error multiplied by 100, respectively

for classical sparse coding, and tree-structured sparse coding.

noise

flat 19.3 ± 0.1 26.8 ± 0.1 36.7 ± 0.1 50.6 ± 0.0 72.1 ± 0.0

tree 18.6 ± 0.1 25.7 ± 0.1 35.0 ± 0.1 48.0 ± 0.0 65.9 ± 0.3

50 %60 %70 %80 %90 %

is the number of children for the nodes at depth k. The

branching factors tested for the trees of depth 3 where

p1 ∈ {5,10,20,40,60,80,100}, p2 ∈ {2,3}, and for

trees of depth 4, p1 ∈ {5,10,20,40}, p2 ∈ {2,3} and

p3= 2, giving 21 possible structures associated with dic-

tionaries with at most 401 elements. For each tree struc-

ture, we evaluated the performance obtained with the tree-

structured dictionary along with the non-structured dictio-

nary containing the same number of elements. These ex-

periments were carried out four times, each time with a

different initialization, and with a different noise realiza-

tion. Quantitative results are reported on Table 1. For ev-

ery number of missing pixels, the tree-structured dictionary

outperforms the “unstructured one”, and the most signifi-

cant improvement is obtained in the noisiest setting. Note

that having more dictionary elements is worthwhile when

using the tree structure. To study the influence of the cho-

sen structure, we have reported on Figure 2 the results ob-

tained by the 14 tested structures of depth 3, along with

those obtained with the unstructured dictionaries contain-

ing the same number of elements, when 90% of the pixels

are missing. For every number of dictionary elements, the

tree-structured dictionary significantly outperforms the un-

structured ones. An example of a learned tree-structured

dictionary is presented on Figure 3. Dictionary elements

naturally organize in groups of patches, with often low fre-

quencies near the root of the tree, and high frequencies near

the leaves. Dictionary elements tend to be highly correlated

with their parents.

12345678 9 10 11 12 13 14

50

60

70

80

Figure 2. Mean square error multiplied by 100 obtained with 14

structures with error bars, sorted by number of dictionary ele-

ments. Red plain bars represents the tree-structured dictionaries.

White bars correspond to the flat dictionary model containing the

same number of dictionary as the tree-structured one. For read-

ability purpose, the y-axis of the graph starts at 50.

Figure 3. Learned dictionary with tree structure of depth 4. The

rootofthetreeisinthemiddleofthefigure. Thebranchingfactors

are p1 = 10, p2 = 2, p3 = 2. The dictionary is learned on

50,000 patches of size 16 × 16 pixels.

4.2. Text Documents

This second experimental section shows that our approach

can also be applied to model text corpora. The goal of

probabilistic topic models is to find a low-dimensional rep-

resentation of a collection of documents, where the rep-

resentation should provide a semantic description of the

collection. Within a parametric Bayesian framework, la-

tent Dirichlet allocation (LDA) (Blei et al., 2003) models

documents as a mixture of a predefined number of latent

topics that are distributions over a fixed vocabulary. When

one marginalizes over the topic random variable, one gets

multinomial PCA (Buntine, 2002). The number of topics is

usually small compared to the size of the vocabulary (e.g.,

100 against 10,000), so that the topic proportions of each

document give a compact representation of the corpus. For

instance, these new features can be used to feed a classifier

in a subsequent classification task. We will similarly use

our dictionary learning approach to find low-dimensional

representations of text corpora.

Suppose that the signals X = [x1,...,xn] ∈ Rm×nare n

documents over a vocabulary of m words, the k-th compo-

nent of xistanding for the frequency of the k-th word in

the document i. If we further assume that the entries of D

and A are nonnegative, and that the dictionary elements dj

have unit ℓ1norm, the decomposition DA can be seen as

a mixture of p topics. The regularization Ω organizes these

topics in a tree, so that, if a document involves a certain

topic, then all its ancestors in the tree are also present in

Page 7

Proximal Methods for Sparse Hierarchical Dictionary Learning

Figure 4. Example of a topic hierarchy estimated from 1714 NIPS

proceedings papers (from 1988 through 1999). Each node corre-

sponds to a topic whose 5 most important words are displayed.

Single characters such as n,t,r are part of the vocabulary and

often appear in NIPS papers, and their place in the hierarchy is

semantically relevant to children topics.

the topic decomposition. Since the hierarchy is shared by

all documents, the topics located at the top of the tree will

be part of every decomposition, and should therefore corre-

spond to topics common to all documents. Conversely, the

deeper the topics in the tree, the more specific they should

be. It is worth mentioning the extension of LDA that con-

sidershierarchiesoftopicsfromanon-parametricBayesian

viewpoint (Blei et al., 2010). We plan to compare to this

model in future work.

Visualization of NIPS proceedings. We first qualitatively

illustrateour dictionary learning approach on the NIPS pro-

ceedings papers from 1988 through 19998. After removing

words appearing fewer than 10 times, the dataset is com-

posed of 1714 articles, with a vocabulary of 8274 words.

As explained above, we consider D+

Rp×n

+

. Figure 4 displays an example of a learned dictio-

nary with 13 topics, obtained by using the ℓ∞norm in Ω

and selecting manually λ = 2−15. Similarly to Blei et al.

(2010), we interestingly capture the stopwords at the root

of the tree, and the different subdomains of the conference

such as neuroscience, optimization or learning theory.

1and take A to be

Posting classification. We now consider a binary clas-

sification task of postings from the 20 Newsgroups data

set9. We classify the postings from the two newsgroups

alt.atheism and talk.religion.misc, following the setting of

8http://psiexp.ss.uci.edu/research/programs data/toolbox.htm

9See http://people.csail.mit.edu/jrennie/20Newsgroups/

37

Number of Topics

15 3163

60

70

80

90

100

Classification Accuracy (%)

PCA + SVM

NMF + SVM

LDA + SVM

SpDL + SVM

SpHDL + SVM

Figure 5. Binary classification of two newsgroups: classification

accuracy for different dimensionality reduction techniques cou-

pled with a linear SVM classifier. The bars and the errors are

respectively the mean and the standard deviation, based on 10 ran-

dom split of the data set. Best seen in color.

Zhu et al. (2009). After removing words appearing fewer

than 10 times and standard stopwords, these postings form

a data set of 1425 documents over a vocabulary of 13312

words.We compare different dimensionality reduction

techniques that we use to feed a linear SVM classifier,

i.e., we consider (i) LDA (with the code from Blei et al.,

2003), (ii) principal component analysis (PCA), (iii) non-

negative matrix factorization (NMF), (iv) standard sparse

dictionary learning (denoted by SpDL) and (v) our sparse

hierarchical approach (denoted by SpHDL). Both SpDL

and SpHDL are optimized over D+

the weights wgequal to 1. We proceed as follows: given

a random split into a training/test set of 1000/425 post-

ings, and given a number of topics p (also the number of

components for PCA, NMF, SpDL and SpHDL), we train

a SVM classifier based on the low-dimensional representa-

tion of the postings. This is performed on the training set of

1000 postings, where the parameters, λ∈{2−26,...,2−5}

and/or Csvm∈ {4−3,...,41} are selected by 5-fold cross-

validation. We report in Figure 5 the average classifica-

tion scores on the test set of 425 postings, based on 10

random splits, for different number of topics. Unlike the

experiment on the image patches, we consider only one

tree structure, namely complete binary trees with depths in

{1,...,5}. The results from Figure 5 show that SpDL and

SpHDLperformbetterthantheotherdimensionalityreduc-

tion techniques on this task. As a baseline, the SVM classi-

fier applied directly to the raw data (the 13312 words) ob-

tains a score of 90.9±1.1, which is better than all the tested

methods, but without dimensionality reduction (as already

reported by Blei et al., 2003). Moreover, the error bars in-

dicate that, though nonconvex, SpDL and SpHDL do not

seem to suffer much from instability issues. Even if SpDL

and SpHDL perform similarly, SpHDL has the advantage

to give a more interpretable topic mixture in terms of hier-

archy, which standard unstructured sparse coding cannot.

1and A = Rp×n

+

, with

Page 8

Proximal Methods for Sparse Hierarchical Dictionary Learning

5. Discussion

We have shown in this paper that tree-structured sparse de-

composition problems can be solved at the same computa-

tional cost as addressing classical decomposition based on

the ℓ1norm. We have used this approach to learn dictionar-

ies embedded in trees, with application to representation of

natural image patches and text documents.

We believe that the connection established between sparse

methods and probabilistic topic models should prove

fruitful as the two lines of work have focused on dif-

ferent aspects of the same unsupervised learning prob-

lem: our approach is based on convex optimization tools,

and provides experimentally more stable data representa-

tions. Moreover, it can be easily extended with the same

tools to other types of structures corresponding to other

norms (Jenatton et al., 2009; Jacob et al., 2009). However,

it is not able to learn elegantly and automatically model

parameters such as dictionary size of tree topology, which

Bayesian methods can. Finally, another interesting com-

mon line of research to pursue is the supervised design

of dictionaries, which has been proved useful in the two

frameworks (Mairal et al., 2009; Blei & McAuliffe., 2008).

Acknowledgments

This paper was partially supported by grants from the

Agence Nationale de la Recherche (MGA Project) and

from the European Research Council (SIERRA Project).

References

Bach, F. High-dimensional non-linear variable selection

through hierarchical kernel learning. Technical report,

arXiv:0909.0844, 2009.

Beck, A. and Teboulle, M.

thresholding algorithm for linear inverse problems.

SIAM J. Imag. Sci., 2(1):183–202, 2009.

A fast iterative shrinkage-

Bengio, S., Pereira, F., Singer, Y., and Strelow, D. Group

sparse coding. In Adv. NIPS, 2009.

Bengio, Y. LearningdeeparchitecturesforAI. Foundations

and Trends in Machine Learning, 2(1), 2009.

Bertsekas, D.P. Nonlinearprogramming. AthenaScientific

Belmont, 1999.

Blei, D., Ng, A., and Jordan, M. Latent dirichlet allocation.

J. Mach. Learn. Res., 3:993–1022, 2003.

Blei, D. and McAuliffe, J. Supervised topic models. In

Adv. NIPS, 2008.

Blei, D., Griffiths, T., and Jordan, M.

nese restaurant process and Bayesian nonparametric in-

ference of topic hierarchies. Journal of the ACM, 2010.

The nested chi-

Boyd, S. P. and Vandenberghe, L. Convex optimization.

Cambridge University Press, 2004.

Buntine, W.L. Variational Extensions to EM and Multino-

mial PCA. In Proc. ECML, 2002.

Elad, M. and Aharon, M.

and redundant representations over learned dictionaries.

IEEE Trans. Image Process., 54(12):3736–3745, 2006.

Image denoising via sparse

Jacob, L., Obozinski, G., and Vert, J.-P. Group Lasso with

overlap and graph Lasso. In Proc. ICML, 2009.

Jenatton, R., Audibert, J.-Y., and Bach, F. Structured vari-

able selection with sparsity-inducing norms. Technical

report, arXiv:0904.3523, 2009.

Jenatton, R., Obozinski, G., and Bach, F. Structured sparse

principal component analysis. In Proc. AISTATS, 2010.

Ji, S., and Ye, J. An accelerated gradient method for trace

norm minimization. In Proc. ICML, 2009.

Kavukcuoglu, K., Ranzato, M., Fergus, R., and LeCun,

Y. Learning invariant features through topographic fil-

ter maps. In Proc. CVPR, 2009.

Kim, S. and Xing, E. P. Tree-guided group lasso for multi-

taskregressionwithstructuredsparsity. Technicalreport,

arXiv:0909.1373, 2009.

Lee, D.D.andSeung, H.S. Learningthepartsofobjectsby

non-negative matrix factorization. Nature, 401(6755):

788–791, 1999.

Lee, H., Battle, A., Raina, R., andNg, A.Y. Efficientsparse

coding algorithms. In Adv. NIPS, 2007.

Mairal, J., Bach, F., Ponce, J., Sapiro, G., and Zisserman,

A. Supervised Dictionary Learning. In Adv. NIPS, 2009.

Mairal, J., Bach, F., Ponce, J., and Sapiro, G. Online learn-

ing for matrix factorization and sparse coding. J. Mach.

Learn. Res., 11, 19–60, 2010.

Nesterov, Y. Gradient methods for minimizing composite

objective function. Technical report, CORE, 2007.

Olshausen, B. A. and Field, D. J. Sparse coding with an

overcomplete basis set: A strategy employed by V1? Vi-

sion Research, 37:3311–3325, 1997.

Zhao, P., Rocha, G., and Yu, B. The composite absolute

penaltiesfamilyforgroupedandhierarchicalvariablese-

lection. Ann. Stat., 37(6A):3468–3497, 2009.

Zhu, J., Ahmed, A., and Xing, E. P. MedLDA: maximum

margin supervised topic models for regression and clas-

sification. In Proc. ICML, 2009.