Proximal gradient methods for learning

Proximal gradient (forward backward splitting) methods for learning is an area of research in optimization and statistical learning theory which studies algorithms for a general class of convex regularization problems where the regularization penalty may not be differentiable. One such example is regularization (also known as Lasso) of the form

Proximal gradient methods offer a general framework for solving regularization problems from statistical learning theory with penalties that are tailored to a specific problem application.[1][2] Such customized penalties can help to induce certain structure in problem solutions, such as sparsity (in the case of lasso) or group structure (in the case of group lasso).

Relevant background

edit

Proximal gradient methods are applicable in a wide variety of scenarios for solving convex optimization problems of the form

 

where   is convex and differentiable with Lipschitz continuous gradient,   is a convex, lower semicontinuous function which is possibly nondifferentiable, and   is some set, typically a Hilbert space. The usual criterion of   minimizes   if and only if   in the convex, differentiable setting is now replaced by

 

where   denotes the subdifferential of a real-valued, convex function  .

Given a convex function   an important operator to consider is its proximal operator   defined by

 

which is well-defined because of the strict convexity of the   norm. The proximal operator can be seen as a generalization of a projection.[1][3][4] We see that the proximity operator is important because   is a minimizer to the problem   if and only if

  where   is any positive real number.[1]

Moreau decomposition

edit

One important technique related to proximal gradient methods is the Moreau decomposition, which decomposes the identity operator as the sum of two proximity operators.[1] Namely, let   be a lower semicontinuous, convex function on a vector space  . We define its Fenchel conjugate   to be the function

 

The general form of Moreau's decomposition states that for any   and any   that

 

which for   implies that  .[1][3] The Moreau decomposition can be seen to be a generalization of the usual orthogonal decomposition of a vector space, analogous with the fact that proximity operators are generalizations of projections.[1]

In certain situations it may be easier to compute the proximity operator for the conjugate   instead of the function  , and therefore the Moreau decomposition can be applied. This is the case for group lasso.

Lasso regularization

edit

Consider the regularized empirical risk minimization problem with square loss and with the   norm as the regularization penalty:

 

where   The   regularization problem is sometimes referred to as lasso (least absolute shrinkage and selection operator).[5] Such   regularization problems are interesting because they induce sparse solutions, that is, solutions   to the minimization problem have relatively few nonzero components. Lasso can be seen to be a convex relaxation of the non-convex problem

 

where   denotes the   "norm", which is the number of nonzero entries of the vector  . Sparse solutions are of particular interest in learning theory for interpretability of results: a sparse solution can identify a small number of important factors.[5]

Solving for L1 proximity operator

edit

For simplicity we restrict our attention to the problem where  . To solve the problem

 

we consider our objective function in two parts: a convex, differentiable term   and a convex function  . Note that   is not strictly convex.

Let us compute the proximity operator for  . First we find an alternative characterization of the proximity operator   as follows:

 

For   it is easy to compute  : the  th entry of   is precisely

 

Using the recharacterization of the proximity operator given above, for the choice of   and   we have that   is defined entrywise by

 

which is known as the soft thresholding operator  .[1][6]

Fixed point iterative schemes

edit

To finally solve the lasso problem we consider the fixed point equation shown earlier:

 

Given that we have computed the form of the proximity operator explicitly, then we can define a standard fixed point iteration procedure. Namely, fix some initial  , and for   define

 

Note here the effective trade-off between the empirical error term   and the regularization penalty  . This fixed point method has decoupled the effect of the two different convex functions which comprise the objective function into a gradient descent step ( ) and a soft thresholding step (via  ).

Convergence of this fixed point scheme is well-studied in the literature[1][6] and is guaranteed under appropriate choice of step size   and loss function (such as the square loss taken here). Accelerated methods were introduced by Nesterov in 1983 which improve the rate of convergence under certain regularity assumptions on  .[7] Such methods have been studied extensively in previous years.[8] For more general learning problems where the proximity operator cannot be computed explicitly for some regularization term  , such fixed point schemes can still be carried out using approximations to both the gradient and the proximity operator.[4][9]

Practical considerations

edit

There have been numerous developments within the past decade in convex optimization techniques which have influenced the application of proximal gradient methods in statistical learning theory. Here we survey a few important topics which can greatly improve practical algorithmic performance of these methods.[2][10]

Adaptive step size

edit

In the fixed point iteration scheme

 

one can allow variable step size   instead of a constant  . Numerous adaptive step size schemes have been proposed throughout the literature.[1][4][11][12] Applications of these schemes[2][13] suggest that these can offer substantial improvement in number of iterations required for fixed point convergence.

Elastic net (mixed norm regularization)

edit

Elastic net regularization offers an alternative to pure   regularization. The problem of lasso ( ) regularization involves the penalty term  , which is not strictly convex. Hence, solutions to   where   is some empirical loss function, need not be unique. This is often avoided by the inclusion of an additional strictly convex term, such as an   norm regularization penalty. For example, one can consider the problem

 

where   For   the penalty term   is now strictly convex, and hence the minimization problem now admits a unique solution. It has been observed that for sufficiently small  , the additional penalty term   acts as a preconditioner and can substantially improve convergence while not adversely affecting the sparsity of solutions.[2][14]

Exploiting group structure

edit

Proximal gradient methods provide a general framework which is applicable to a wide variety of problems in statistical learning theory. Certain problems in learning can often involve data which has additional structure that is known a priori. In the past several years there have been new developments which incorporate information about group structure to provide methods which are tailored to different applications. Here we survey a few such methods.

Group lasso

edit

Group lasso is a generalization of the lasso method when features are grouped into disjoint blocks.[15] Suppose the features are grouped into blocks  . Here we take as a regularization penalty

 

which is the sum of the   norm on corresponding feature vectors for the different groups. A similar proximity operator analysis as above can be used to compute the proximity operator for this penalty. Where the lasso penalty has a proximity operator which is soft thresholding on each individual component, the proximity operator for the group lasso is soft thresholding on each group. For the group   we have that proximity operator of   is given by

 

where   is the  th group.

In contrast to lasso, the derivation of the proximity operator for group lasso relies on the Moreau decomposition. Here the proximity operator of the conjugate of the group lasso penalty becomes a projection onto the ball of a dual norm.[2]

Other group structures

edit

In contrast to the group lasso problem, where features are grouped into disjoint blocks, it may be the case that grouped features are overlapping or have a nested structure. Such generalizations of group lasso have been considered in a variety of contexts.[16][17][18][19] For overlapping groups one common approach is known as latent group lasso which introduces latent variables to account for overlap.[20][21] Nested group structures are studied in hierarchical structure prediction and with directed acyclic graphs.[18]

See also

edit

References

edit
  1. ^ a b c d e f g h i Combettes, Patrick L.; Wajs, Valérie R. (2005). "Signal Recovering by Proximal Forward-Backward Splitting". Multiscale Model. Simul. 4 (4): 1168–1200. doi:10.1137/050626090. S2CID 15064954.
  2. ^ a b c d e Mosci, S.; Rosasco, L.; Matteo, S.; Verri, A.; Villa, S. (2010). "Solving Structured Sparsity Regularization with Proximal Methods". Machine Learning and Knowledge Discovery in Databases. Lecture Notes in Computer Science. Vol. 6322. pp. 418–433. doi:10.1007/978-3-642-15883-4_27. ISBN 978-3-642-15882-7.
  3. ^ a b Moreau, J.-J. (1962). "Fonctions convexes duales et points proximaux dans un espace hilbertien". Comptes Rendus de l'Académie des Sciences, Série A. 255: 2897–2899. MR 0144188. Zbl 0118.10502.
  4. ^ a b c Bauschke, H.H., and Combettes, P.L. (2011). Convex analysis and monotone operator theory in Hilbert spaces. Springer.{{cite book}}: CS1 maint: multiple names: authors list (link)
  5. ^ a b Tibshirani, R. (1996). "Regression shrinkage and selection via the lasso". J. R. Stat. Soc. Ser. B. 1. 58 (1): 267–288. doi:10.1111/j.2517-6161.1996.tb02080.x.
  6. ^ a b Daubechies, I.; Defrise, M.; De Mol, C. (2004). "An iterative thresholding algorithm for linear inverse problem with a sparsity constraint". Comm. Pure Appl. Math. 57 (11): 1413–1457. arXiv:math/0307152. doi:10.1002/cpa.20042. S2CID 1438417.
  7. ^ Nesterov, Yurii (1983). "A method of solving a convex programming problem with convergence rate  ". Soviet Mathematics - Doklady. 27 (2): 372–376.
  8. ^ Nesterov, Yurii (2004). Introductory Lectures on Convex Optimization. Kluwer Academic Publisher.
  9. ^ Villa, S.; Salzo, S.; Baldassarre, L.; Verri, A. (2013). "Accelerated and inexact forward-backward algorithms". SIAM J. Optim. 23 (3): 1607–1633. CiteSeerX 10.1.1.416.3633. doi:10.1137/110844805. S2CID 11379846.
  10. ^ Bach, F.; Jenatton, R.; Mairal, J.; Obozinski, Gl. (2011). "Optimization with sparsity-inducing penalties". Foundations and Trends in Machine Learning. 4 (1): 1–106. arXiv:1108.0775. Bibcode:2011arXiv1108.0775B. doi:10.1561/2200000015. S2CID 56356708.
  11. ^ Loris, I.; Bertero, M.; De Mol, C.; Zanella, R.; Zanni, L. (2009). "Accelerating gradient projection methods for  -constrained signal recovery by steplength selection rules". Applied & Comp. Harmonic Analysis. 27 (2): 247–254. arXiv:0902.4424. doi:10.1016/j.acha.2009.02.003. S2CID 18093882.
  12. ^ Wright, S.J.; Nowak, R.D.; Figueiredo, M.A.T. (2009). "Sparse reconstruction by separable approximation". IEEE Trans. Image Process. 57 (7): 2479–2493. Bibcode:2009ITSP...57.2479W. CiteSeerX 10.1.1.115.9334. doi:10.1109/TSP.2009.2016892. S2CID 7399917.
  13. ^ Loris, Ignace (2009). "On the performance of algorithms for the minimization of  -penalized functionals". Inverse Problems. 25 (3): 035008. arXiv:0710.4082. Bibcode:2009InvPr..25c5008L. doi:10.1088/0266-5611/25/3/035008. S2CID 14213443.
  14. ^ De Mol, C.; De Vito, E.; Rosasco, L. (2009). "Elastic-net regularization in learning theory". J. Complexity. 25 (2): 201–230. arXiv:0807.3423. doi:10.1016/j.jco.2009.01.002. S2CID 7167292.
  15. ^ Yuan, M.; Lin, Y. (2006). "Model selection and estimation in regression with grouped variables". J. R. Stat. Soc. B. 68 (1): 49–67. doi:10.1111/j.1467-9868.2005.00532.x. S2CID 6162124.
  16. ^ Chen, X.; Lin, Q.; Kim, S.; Carbonell, J.G.; Xing, E.P. (2012). "Smoothing proximal gradient method for general structured sparse regression". Ann. Appl. Stat. 6 (2): 719–752. arXiv:1005.4717. doi:10.1214/11-AOAS514. S2CID 870800.
  17. ^ Mosci, S.; Villa, S.; Verri, A.; Rosasco, L. (2010). "A primal-dual algorithm for group sparse regularization with overlapping groups". NIPS. 23: 2604–2612.
  18. ^ a b Jenatton, R.; Audibert, J.-Y.; Bach, F. (2011). "Structured variable selection with sparsity-inducing norms". J. Mach. Learn. Res. 12: 2777–2824. arXiv:0904.3523. Bibcode:2009arXiv0904.3523J.
  19. ^ Zhao, P.; Rocha, G.; Yu, B. (2009). "The composite absolute penalties family for grouped and hierarchical variable selection". Ann. Stat. 37 (6A): 3468–3497. arXiv:0909.0411. Bibcode:2009arXiv0909.0411Z. doi:10.1214/07-AOS584. S2CID 9319285.
  20. ^ Obozinski, Guillaume; Jacob, Laurent; Vert, Jean-Philippe (2011). "Group Lasso with Overlaps: The Latent Group Lasso approach". arXiv:1110.0413 [stat.ML].
  21. ^ Villa, Silvia; Rosasco, Lorenzo; Mosci, Sofia; Verri, Alessandro (2012). "Proximal methods for the latent group lasso penalty". arXiv:1209.0368 [math.OC].