Proximal Minimization by Incremental Surrogate Optimization (MISO)

Transcription

1 Proximal Minimization by Incremental Surrogate Optimization (MISO) (and a few variants) Julien Mairal Inria, Grenoble ICCOPT, Tokyo, 2016 Julien Mairal, Inria MISO 1/26

2 Motivation: large-scale machine learning Minimizing large finite sums of functions Given data points x i, i =1,...,n, learnsomemodel parameters θ in R p by minimizing 1 min θ R p n n l(x i,θ)+ψ(θ), i=1 where l measures the data fit, andψ is a regularization function. Minimizing expectations If the amount of data is infinite, we may also need to minimize the expected cost min θ R p E x[l(x,θ)] + ψ(θ), leading to a stochastic optimization problem. Julien Mairal, Inria MISO 2/26

3 Motivation: large-scale machine learning Afewexamplesfromtheconvexworld 1 n 1 Ridge regression: min x R p n 2 (y i θ, x i ) 2 + λ 2 θ 2 2. Linear SVM: Logistic regression: 1 min x R p n 1 min x R p n i=1 n max(0, 1 y i θ, x i ) + λ 2 θ 2 2. i=1 n i=1 ( ) log 1+e y i θ,x i + λ 2 θ 2 2. Julien Mairal, Inria MISO 3/26

4 Motivation: large-scale machine learning Afewexamplesfromtheconvexworld 1 n 1 Ridge regression: min x R p n 2 (y i θ, x i ) 2 + λ 2 θ 2 2. Linear SVM: Logistic regression: 1 min x R p n 1 min x R p n i=1 n max(0, 1 y i θ, x i ) 2 + λ 2 θ 2 2. i=1 n i=1 ( ) log 1+e y i θ,x i + λ 2 θ 2 2. Julien Mairal, Inria MISO 3/26

5 Motivation: large-scale machine learning Afewexamplesfromtheconvexworld 1 n 1 Ridge regression: min x R p n 2 (y i θ, x i ) 2 + λ θ 1. Linear SVM: Logistic regression: 1 min x R p n 1 min x R p n i=1 n max(0, 1 y i θ, x i ) 2 + λ θ 1. i=1 n i=1 ( ) log 1+e y i θ,x i + λ θ 1. Julien Mairal, Inria MISO 3/26

6 Methodology We will consider optimization methods that iteratively build a model of the objective before updating the variable: θ t arg min g t (θ), θ R p where g t is easy to minimize and exploits the objective structure: large finite sum, expectation, (strong) convexity, composite? There is a large body of related work Kelley s and bundle methods; incremental and online EM algorithms; incremental and stochastic proximal gradient methods; variance-reduction techniques for minimizing finite sums. [Neal and Hinton, 1998, Duchi and Singer, 2009, Bertsekas, 2011, Schmidt et al., 2013, Defazio et al., 2014a, Shalev-Shwartz and Zhang, 2012, Lan,2012,2015]... Julien Mairal, Inria MISO 4/26

7 Outline of the talk 1) stochastic majorization-minimization min E x[l(x,θ)] + ψ(θ), θ R p where l is not necessarily smooth or convex. 2) incremental majorization-minimization 1 n min l(x i,θ)+ψ(θ). θ R p n i=1 The MISO algorithm for non-convex functions. 3) faster schemes for composite strongly-convex functions Another MISO algorithm for strongly-convex functions. 4)?? Julien Mairal, Inria MISO 5/26

8 Outline of the talk 1) stochastic majorization-minimization min E x[l(x,θ)] + ψ(θ), θ R p where l is not necessarily smooth or convex. 2) incremental majorization-minimization 1 n min l(x i,θ)+ψ(θ). θ R p n i=1 The MISO algorithm for non-convex functions. 3) faster schemes for composite strongly-convex functions Another MISO algorithm for strongly-convex functions. 4)?? Julien Mairal, Inria MISO 6/26

9 Majorization-minimization principle g t (θ) f (θ) θ t 1 θ t Iteratively minimize locally tight upper bounds of the objective. The objective monotonically decreases. Under some assumptions, we get similar convergence rates as classical first-order approaches in the convex case. Julien Mairal, Inria MISO 7/26

10 Setting: first-order surrogate functions h t (θ) g t (θ) θ t 1 f (θ) θ t g t (θ t ) f (θ t )forθ t in arg min θ Θ g t (θ); the approximation error h t = gt f is differentiable, and h t is L-Lipschitz. Moreover, h t (θ t 1 )=0and h t (θ t 1 ) = 0; we may also need g t to be strongly convex. Julien Mairal, Inria MISO 8/26

11 Examples of first-order surrogate functions Lipschitz gradient surrogates: f is L-smooth (differentiable + L-Lipschitz gradient). g : θ f (κ)+ f (κ) (θ κ)+ L 2 θ κ 2 2. Minimizing g yields a gradient descent step θ κ 1 L f (κ). Proximal gradient surrogates: f = f + ψ with f smooth. g : θ f (κ)+ f (κ) (θ κ)+ L 2 θ κ ψ(θ). Minimizing g amounts to one step of the forward-backward, ISTA, or proximal gradient descent algorithm. [Nesterov, 2004, 2013, Beck and Teboulle, 2009, Wright et al., 2009]... Julien Mairal, Inria MISO 9/26

12 Examples of first-order surrogate functions Linearizing concave functions and dc-programming: f = f 1 + f 2 with f 2 smooth and concave. g : θ f 1 (θ)+f 2 (κ)+ f 2 (κ) (θ κ). when f 1 is convex, the algorithm is called dc-programming. Quadratic surrogates: f is twice differentiable, and H is a uniform upper bound of 2 f : g : θ f (κ)+ f (κ) (θ κ)+ 1 2 (θ κ) H(θ κ). Upper-bounds based on Jensen s inequality... Julien Mairal, Inria MISO 10/26

13 Theoretical guarantees of the basic MM algorithm When using first-order surrogates, for convex problems: f (θ t ) f = O(L/t). for µ-strongly convex ones: O((1 µ/l) t ). for non-convex problems: f (θ t ) monotonically decreases and lim inf inf f (θ t,θ θ t ) 0, (1) t + θ Θ θ θ t 2 which we call asymptotic stationary point condition. Directional derivative f (θ + εκ) f (θ) f (θ, κ) = lim. ε 0 + ε when Θ = R p and f is smooth, (1) is equivalent to f (θ t ) 0. Julien Mairal, Inria MISO 11/26

14 Stochastic majorization minimization [Mairal, 2013] Assume that f is an expectation: Recipe f (θ) =E x [l(θ, x)]. Draw a single function f t : θ l(θ, x t ) at iteration t; Choose a first-order surrogate function g t for f t at θ t 1 ; Update the model g t =(1 w t )g t 1 + w t g t with appropriate w t ; Update θ t by minimizing g t. Related Work online-em; online matrix factorization. [Neal and Hinton, 1998, Mairal et al., 2010, Razaviyayn et al., 2013]... Julien Mairal, Inria MISO 12/26

15 Stochastic majorization minimization [Mairal, 2013] Theoretical Guarantees - Non-Convex Problems under a set of reasonable assumptions, f (θ t ) almost surely converges; the function g t asymptotically behaves as a first-order surrogate; asymptotic stationary point conditions hold almost surely. Theoretical Guarantees - Convex Problems under a few assumptions, for proximal gradient surrogates, we obtain similar expected rates as SGD with averaging: O(1/t) for strongly convex problems, O(log(t)/ t)forconvex ones. The most interesting feature of this principle is probably the ability to deal with some non-smooth non-convex problems. Julien Mairal, Inria MISO 13/26

16 Stochastic majorization minimization [Mairal, 2013] Update Rule for Proximal Gradient Surrogate θ t arg min θ Θ t i=1 w i t [ fi (θ i 1 ) θ + L 2 θ θ i ψ(θ)]. (SMM) Other schemes in the literature [Duchi and Singer, 2009]: θ t arg min f t (θ t 1 ) θ + 1 2η t θ θ t ψ(θ), θ Θ (FOBOS) or regularized dual averaging (RDA) of Xiao [2010]: θ t arg min θ Θ 1 t t i=1 f i (θ i 1 ) θ + 1 2η t θ ψ(θ). (RDA) or others... Julien Mairal, Inria MISO 14/26

17 Outline of the talk 1) stochastic majorization-minimization min E x[l(x,θ)] + ψ(θ), θ R p where l is not necessarily smooth or convex. 2) incremental majorization-minimization 1 n min l(x i,θ)+ψ(θ). θ R p n i=1 The MISO algorithm for non-convex functions. 3) faster schemes for composite strongly-convex functions Another MISO algorithm for strongly-convex functions. 4)?? Julien Mairal, Inria MISO 15/26

18 MISO (MM) for non-convex optimization [Mairal, 2015] Assume that f splits into many components: f (θ) = 1 n n f i (θ). i=1 Recipe Draw at random a single index i t at iteration t; Compute a first-order surrogate g it t of f it at θ t 1 ; Incrementally update the approximate surrogate g t = 1 n n i=1 Update θ t by minimizing g t. gt i = g t it (gt g it t 1 n ). Julien Mairal, Inria MISO 16/26

19 MISO (MM) for non-convex optimization [Mairal, 2015] Theoretical Guarantees - Non-Convex Problems same as the basic MM algorithm with probability one. Theoretical Guarantees - Convex Problems when using proximal gradient surrogates, for convex problems, f ( ˆθ t ) f = O(nL/t). for µ-strongly convex problems, f (θ t ) f = O((1 µ/(nl)) t ). The computational complexity is the same as ISTA. Related work for non-convex problems incremental EM; more specific incremental MM algorithms. [Neal and Hinton, 1998, Ahn et al., 2006]. Julien Mairal, Inria MISO 17/26

20 Outline of the talk 1) stochastic majorization-minimization min E x[l(x,θ)] + ψ(θ), θ R p where l is not necessarily smooth or convex. 2) incremental majorization-minimization 1 n min l(x i,θ)+ψ(θ). θ R p n i=1 The MISO algorithm for non-convex functions. 3) faster schemes for composite strongly-convex functions Another MISO algorithm for strongly-convex functions. 4)?? Julien Mairal, Inria MISO 18/26

21 MISO for µ-strongly convex smooth functions Strong convexity provides simple quadratic surrogate functions: g i t : θ f i (θ t 1 )+ f i (θ t 1 ) (θ θ t 1 )+ µ 2 θ θ t (2) This time, the model of the objective is a lower bound. Proposition: MISO with lower bounds [Mairal, 2015] When the functions f i are µ-strongly convex, L-smooth, and non-negative, MISO with the surrogates (2) guarantees that E[f (θ t ) f ] under the big data condition n 2L/µ. Remark ( 1 1 ) t nf, 3n When n 2L/µ, the algorithm may diverge. Julien Mairal, Inria MISO 19/26

22 MISO for µ-strongly convex composite functions [Lin, Mairal, and Harchaoui, 2015] First goal: allow a composite term ψ f (θ) = 1 n f i (θ)+ψ(θ), n i=1 by simply using the composite lower-bounds g i t : θ f i (θ t 1 )+ f i (θ t 1 ) (θ θ t 1 )+ µ 2 θ θ t ψ(θ). ( ) Second goal: remove the condition n 2L/µ ( with δ =min 1, gt i : θ (1 δ)gt 1(θ)+δ( ), i (3) ) instead of δ = 1 previously. µn 2(L µ) Julien Mairal, Inria MISO 20/26

23 MISO for µ-strongly convex composite functions [Lin, Mairal, and Harchaoui, 2015] Convergence of MISO-prox When the functions f i are µ-strongly convex, L-smooth, MISO-prox with the surrogates (3) guarantees that E[f (θ t )] f 1 { µ τ (1 τ)t+1 (f (θ 0 ) g 0 (θ 0 )) with τ min 4L, 1 }. 2n Furthermore, we also have fast convergence of the certificate E[f (θ t ) g t (θ t )] 1 τ (1 τ)t (f g 0 (θ 0 )). Julien Mairal, Inria MISO 21/26

24 MISO for µ-strongly convex composite functions [Lin, Mairal, and Harchaoui, 2015] Relation with SDCA [Shalev-Shwartz and Zhang, 2012]. Variant 5 of SDCA is identical to MISO-Prox with δ = µn L+µn ; The construction is primal. The proof of convergence and the algorithm do not use duality, whereas SDCA is a dual ascent technique; g t (θ t ) is a lower-bound of f ; it plays the same role as the dual lower bound in SDCA, but is easier to evaluate. Another viewpoint about SDCA without duality [Shalev-Shwartz, 2015]. Julien Mairal, Inria MISO 22/26

25 MISO for µ-strongly convex composite functions We may now compare the expected complexity, using the fact that incremental algorithms require to compute a single f i per iteration. grad. desc., ISTA, MISO-MM FISTA, acc. grad. desc. SVRG, SAG, SAGA, SDCA, MISOµ, Finito ( µ>0 O n L µ log ( ) ) 1 ε ( O n L µ log ( ) ) 1 ε ( ( ) O max n, L µ log ( ) ) 1 ε SVRG, SAG, SAGA, SDCA, MISO, Finito improve upon FISTA when ( max n, L ) L L n µ µ µ n, [Schmidt et al., 2013, Xiao and Zhang, 2014, Defazio et al., 2014a,b, Shalev-Shwartz and Zhang, 2012, Zhang and Xiao, 2015] Julien Mairal, Inria MISO 23/26

26 Outline of the talk 1) stochastic majorization-minimization min E x[l(x,θ)] + ψ(θ), θ R p where l is not necessarily smooth or convex. 2) incremental majorization-minimization 1 n min l(x i,θ)+ψ(θ). θ R p n i=1 The MISO algorithm for non-convex functions. 3) faster schemes for composite strongly-convex functions Another MISO algorithm for strongly-convex functions. 4)?? Julien Mairal, Inria MISO 24/26

27 Can we do better? [Lin, Mairal, and Harchaoui, 2015] SVRG, SAG, SAGA, SDCA, MISO, Finito improve upon FISTA but they are not accelerated in the sense of Nesterov. [Beck and Teboulle, 2009, Nesterov, 2013, Shalev-Shwartz and Zhang, 2014, Lan, 2015, Agarwal and Bottou, 2015, Allen-Zhu, 2016] Julien Mairal, Inria MISO 25/26

28 Can we do better? [Lin, Mairal, and Harchaoui, 2015] SVRG, SAG, SAGA, SDCA, MISO, Finito improve upon FISTA but they are not accelerated in the sense of Nesterov. How to improve the previous complexities? 1 read classical paper about accelerated gradient methods; [Beck and Teboulle, 2009, Nesterov, 2013, Shalev-Shwartz and Zhang, 2014, Lan, 2015, Agarwal and Bottou, 2015, Allen-Zhu, 2016] Julien Mairal, Inria MISO 25/26

29 Can we do better? [Lin, Mairal, and Harchaoui, 2015] SVRG, SAG, SAGA, SDCA, MISO, Finito improve upon FISTA but they are not accelerated in the sense of Nesterov. How to improve the previous complexities? 1 read classical paper about accelerated gradient methods; 2 stay in the room and listen to G. Lan s talk; [Beck and Teboulle, 2009, Nesterov, 2013, Shalev-Shwartz and Zhang, 2014, Lan, 2015, Agarwal and Bottou, 2015, Allen-Zhu, 2016] Julien Mairal, Inria MISO 25/26

30 Can we do better? [Lin, Mairal, and Harchaoui, 2015] SVRG, SAG, SAGA, SDCA, MISO, Finito improve upon FISTA but they are not accelerated in the sense of Nesterov. How to improve the previous complexities? 1 read classical paper about accelerated gradient methods; 2 stay in the room and listen to G. Lan s talk; 3 Listen to Hongzhou s Lin talk tomorrow. Tuesday, 2:45pm, room 5A [Beck and Teboulle, 2009, Nesterov, 2013, Shalev-Shwartz and Zhang, 2014, Lan, 2015, Agarwal and Bottou, 2015, Allen-Zhu, 2016] Julien Mairal, Inria MISO 25/26

31 Conclusion a large class of majorization-minimization algorithms for non-convex, possibly non-smooth, optimization; fast algorithms for minimizing large sums of convex functions (using lower bounds). see Hongzhou Lin s talk on acceleration tomorrow. Related publications J. Mairal. Optimization with First-Order Surrogate Functions. ICML, J. Mairal. Stochastic Majorization-Minimization Algorithms for Large-Scale Optimization. NIPS, J. Mairal. Incremental Majorization-Minimization Optimization with Application to Large-Scale Machine Learning. SIAM Journal on Optimization, 2015; H. Lin, J. Mairal, and Z. Harchaoui. A Universal Catalyst for First-Order Optimization. NIPS, 2015; Julien Mairal, Inria MISO 26/26

32 Online Sparse Matrix Factorization Consider some signals x in R m. We want to find a dictionary D in R m p. The quality of D is measured through the loss l(x, D) = 1 min α R K 2 x Dα λ 1 α 1 + λ 2 2 α 2 2. Then, learning the dictionary amounts to solving min E x [l(x, D)] + ϕ(d), D D Why is it a matrix factorization problem? [ 1 1 min D D,A R p n n 2 X DA 2 F + n i=1 λ 1 α i 1 + λ 2 2 α i 2 2 ] + ϕ(d). Julien Mairal, Inria MISO 27/26

33 Online Sparse Matrix Factorization when D = {D R m p s.t. d j 2 1} and ϕ = 0, the problem is called sparse coding or dictionary learning [Olshausen and Field, 1996, Elad and Aharon, 2006, Mairal et al., 2010]. non-negativity constraints can be easily added. It yields an online nonnegative matrix factorization algorithm. ϕ can be a function encouraging a particular structure in D [Jenatton et al., 2011]. Julien Mairal, Inria MISO 28/26

34 Online Sparse Matrix Factorization Dictionary Learning on Natural Image Patches Consider n = whitened natural image patches of size m = We learn a dictionary with K = 256 elements. 0s on an old laptop 1.2GHz dual-core CPU. (initialization) Julien Mairal, Inria MISO 29/26

35 Online Sparse Matrix Factorization Dictionary Learning on Natural Image Patches Consider n = whitened natural image patches of size m = We learn a dictionary with K = 256 elements. 1.15s on an old laptop 1.2GHz dual-core CPU (0.1 pass) Julien Mairal, Inria MISO 30/26

36 Online Sparse Matrix Factorization Dictionary Learning on Natural Image Patches Consider n = whitened natural image patches of size m = We learn a dictionary with K = 256 elements. 5.97s on an old laptop 1.2GHz dual-core CPU (0.5 pass) Julien Mairal, Inria MISO 31/26

37 Online Sparse Matrix Factorization Dictionary Learning on Natural Image Patches Consider n = whitened natural image patches of size m = We learn a dictionary with K = 256 elements s on an old laptop 1.2GHz dual-core CPU (1 pass) Julien Mairal, Inria MISO 32/26

38 Online Sparse Matrix Factorization Dictionary Learning on Natural Image Patches Consider n = whitened natural image patches of size m = We learn a dictionary with K = 256 elements s on an old laptop 1.2GHz dual-core CPU (2 passes) Julien Mairal, Inria MISO 33/26

39 Online Sparse Matrix Factorization Dictionary Learning on Natural Image Patches Consider n = whitened natural image patches of size m = We learn a dictionary with K = 256 elements s on an old laptop 1.2GHz dual-core CPU (5 passes) Julien Mairal, Inria MISO 34/26

40 References I A. Agarwal and L. Bottou. A lower bound for the optimization of finite sums. In Proceedings of the International Conference on Machine Learning (ICML), Sangtae Ahn, Jeffrey A Fessler, Doron Blatt, and Alfred O Hero. Convergent incremental optimization transfer algorithms: Application to tomography. IEEE Transactions on Medical Imaging, 25(3): ,2006. Z. Allen-Zhu. Katyusha: The first direct acceleration of stochastic gradient methods. ArXiv , A. Beck and M. Teboulle. A fast iterative shrinkage-thresholding algorithm for linear inverse problems. SIAM Journal on Imaging Sciences, 2(1): , D. P. Bertsekas. Incremental gradient, subgradient, and proximal methods for convex optimization: A survey. Optimization for Machine Learning, 2010: 1 38, Julien Mairal, Inria MISO 35/26

41 References II A. Defazio, F. Bach, and S. Lacoste-Julien. SAGA: A fast incremental gradient method with support for non-strongly convex composite objectives. In Advances in Neural Information Processing Systems (NIPS), 2014a. A. J. Defazio, T. S. Caetano, and J. Domke. Finito: A faster, permutable incremental gradient method for big data problems. In Proceedings of the International Conference on Machine Learning (ICML), 2014b. J. Duchi and Y. Singer. Efficient online and batch learning using forward backward splitting. Journal of Machine Learning Research, 10: , M. Elad and M. Aharon. Image denoising via sparse and redundant representations over learned dictionaries. IEEE Transactions on Image Processing, 15(12): ,2006. R. Jenatton, J-Y. Audibert, and F. Bach. Structured variable selection with sparsity-inducing norms. Journal of Machine Learning Research, 12: , Guanghui Lan. An optimal method for stochastic composite optimization. Mathematical Programming, 133(1-2): ,2012. Julien Mairal, Inria MISO 36/26

42 References III Guanghui Lan. An optimal randomized incremental gradient method. arxiv preprint arxiv: , H. Lin, J. Mairal, and Z. Harchaoui. A universal catalyst for first-order optimization. In Advances in Neural Information Processing Systems, J. Mairal. Stochastic majorization-minimization algorithms for large-scale optimization. In Advances in Neural Information Processing Systems (NIPS), J. Mairal. Incremental majorization-minimization optimization with application to large-scale machine learning. SIAM Journal on Optimization, 25(2): , J. Mairal, F. Bach, J. Ponce, and G. Sapiro. Online learning for matrix factorization and sparse coding. Journal of Machine Learning Research, 11: 19 60, R.M. Neal and G.E. Hinton. A view of the EM algorithm that justifies incremental, sparse, and other variants. Learning in graphical models, 89, Julien Mairal, Inria MISO 37/26

43 References IV Y. Nesterov. Introductory lectures on convex optimization: a basic course. Kluwer Academic Publishers, Y. Nesterov. Gradient methods for minimizing composite objective function. Mathematical Programming, 140(1): ,2013. B. A. Olshausen and D. J. Field. Emergence of simple-cell receptive field properties by learning a sparse code for natural images. Nature, 381: , Meisam Razaviyayn, Maziar Sanjabi, and Zhi-Quan Luo. A stochastic successive minimization method for nonsmooth nonconvex optimization with applications to transceiver design in wireless communication networks. arxiv preprint arxiv: , M. Schmidt, N. Le Roux, and F. Bach. Minimizing finite sums with the stochastic average gradient. arxiv: , S. Shalev-Shwartz. Sdca without duality. arxiv 1502:06177, S. Shalev-Shwartz and T. Zhang. Proximal stochastic dual coordinate ascent. arxiv: , Julien Mairal, Inria MISO 38/26

44 References V S. Shalev-Shwartz and T. Zhang. Accelerated proximal stochastic dual coordinate ascent for regularized loss minimization. Mathematical Programming, pages1 41,2014. S.J. Wright, R.D. Nowak, and M.A.T. Figueiredo. Sparse reconstruction by separable approximation. IEEE Transactions on Signal Processing, 57(7): , L. Xiao. Dual averaging methods for regularized stochastic learning and online optimization. Journal of Machine Learning Research, 11: , L. Xiao and T. Zhang. A proximal stochastic gradient method with progressive variance reduction. SIAM Journal on Optimization, 24(4): , Y. Zhang and L. Xiao. Stochastic primal-dual coordinate method for regularized empirical risk minimization. In Proceedings of the International Conference on Machine Learning (ICML), Julien Mairal, Inria MISO 39/26