Large-scale machine learning and convex optimization

Transcription

1 Large-scale machine learning and convex optimization Francis Bach INRIA - Ecole Normale Supérieure, Paris, France Eurandom - March 2014 Slides available at

2 Context Machine learning for big data Large-scale machine learning: large d, large n, large k d : dimension of each observation (input) n : number of observations k : number of tasks (dimension of outputs) Examples: computer vision, bioinformatics, text processing Ideal running-time complexity: O(dn + kn) Going back to simple methods Stochastic gradient methods (Robbins and Monro, 1951) Mixing statistics and optimization Using smoothness to go beyond stochastic gradient descent

3 Object recognition

4 Learning for bioinformatics - Proteins Crucial components of cell life Predicting multiple functions and interactions Massive data: up to 1 millions for humans! Complex data Amino-acid sequence Link with DNA Tri-dimensional molecule

5 Search engines - advertising

6 Advertising - recommendation

7 Context Machine learning for big data Large-scale machine learning: large d, large n, large k d : dimension of each observation (input) n : number of observations k : number of tasks (dimension of outputs) Examples: computer vision, bioinformatics, text processing Ideal running-time complexity: O(dn + kn) Going back to simple methods Stochastic gradient methods (Robbins and Monro, 1951) Mixing statistics and optimization Using smoothness to go beyond stochastic gradient descent

8 Context Machine learning for big data Large-scale machine learning: large d, large n, large k d : dimension of each observation (input) n : number of observations k : number of tasks (dimension of outputs) Examples: computer vision, bioinformatics, text processing Ideal running-time complexity: O(dn + kn) Going back to simple methods Stochastic gradient methods (Robbins and Monro, 1951) Mixing statistics and optimization Using smoothness to go beyond stochastic gradient descent

9 Outline 1. Large-scale machine learning and optimization Traditional statistical analysis Classical methods for convex optimization 2. Non-smooth stochastic approximation Stochastic (sub)gradient and averaging Non-asymptotic results and lower bounds Strongly convex vs. non-strongly convex 3. Smooth stochastic approximation algorithms Asymptotic and non-asymptotic results 4. Beyond decaying step-sizes 5. Finite data sets

10 Supervised machine learning Data: n observations (x i,y i ) X Y, i = 1,...,n, i.i.d. Prediction as a linear function θ Φ(x) of features Φ(x) R d (regularized) empirical risk minimization: find ˆθ solution of 1 n min l ( y i,θ Φ(x i ) ) + µω(θ) θ R d n i=1 convex data fitting term + regularizer

11 Usual losses Regression: y R, prediction ŷ = θ Φ(x) quadratic loss 1 2 (y ŷ)2 = 1 2 (y θ Φ(x)) 2

12 Usual losses Regression: y R, prediction ŷ = θ Φ(x) quadratic loss 1 2 (y ŷ)2 = 1 2 (y θ Φ(x)) 2 Classification : y { 1,1}, prediction ŷ = sign(θ Φ(x)) loss of the form l(yθ Φ(x)) True 0-1 loss: l(yθ Φ(x)) = 1 yθ Φ(x)<0 Usual convex losses: hinge square logistic

13 Usual regularizers Main goal: avoid overfitting (squared) Euclidean norm: θ 2 2 = d j=1 θ j 2 Numerically well-behaved Representer theorem and kernel methods : θ = n i=1 α iφ(x i ) See, e.g., Schölkopf and Smola (2001); Shawe-Taylor and Cristianini (2004) Sparsity-inducing norms Main example: l 1 -norm θ 1 = d j=1 θ j Perform model selection as well as regularization Non-smooth optimization and structured sparsity See, e.g., Bach, Jenatton, Mairal, and Obozinski (2011, 2012)

14 Supervised machine learning Data: n observations (x i,y i ) X Y, i = 1,...,n, i.i.d. Prediction as a linear function θ Φ(x) of features Φ(x) R d (regularized) empirical risk minimization: find ˆθ solution of 1 n min l ( y i,θ Φ(x i ) ) + µω(θ) θ R d n i=1 convex data fitting term + regularizer

15 Supervised machine learning Data: n observations (x i,y i ) X Y, i = 1,...,n, i.i.d. Prediction as a linear function θ Φ(x) of features Φ(x) R d (regularized) empirical risk minimization: find ˆθ solution of 1 n min l ( y i,θ Φ(x i ) ) + µω(θ) θ R d n i=1 convex data fitting term + regularizer Empirical risk: ˆf(θ) = 1 n n i=1 l(y i,θ Φ(x i )) training cost Expected risk: f(θ) = E (x,y) l(y,θ Φ(x)) testing cost Two fundamental questions: (1) computing ˆθ and (2) analyzing ˆθ May be tackled simultaneously

16 Supervised machine learning Data: n observations (x i,y i ) X Y, i = 1,...,n, i.i.d. Prediction as a linear function θ Φ(x) of features Φ(x) R d (regularized) empirical risk minimization: find ˆθ solution of 1 n min l ( y i,θ Φ(x i ) ) + µω(θ) θ R d n i=1 convex data fitting term + regularizer Empirical risk: ˆf(θ) = 1 n n i=1 l(y i,θ Φ(x i )) training cost Expected risk: f(θ) = E (x,y) l(y,θ Φ(x)) testing cost Two fundamental questions: (1) computing ˆθ and (2) analyzing ˆθ May be tackled simultaneously

17 Supervised machine learning Data: n observations (x i,y i ) X Y, i = 1,...,n, i.i.d. Prediction as a linear function θ Φ(x) of features Φ(x) R d (regularized) empirical risk minimization: find ˆθ solution of 1 n min l ( y i,θ Φ(x i ) ) such that Ω(θ) D θ R d n i=1 convex data fitting term + constraint Empirical risk: ˆf(θ) = 1 n n i=1 l(y i,θ Φ(x i )) training cost Expected risk: f(θ) = E (x,y) l(y,θ Φ(x)) testing cost Two fundamental questions: (1) computing ˆθ and (2) analyzing ˆθ May be tackled simultaneously

18 Analysis of empirical risk minimization Approximation and estimation errors: C = {θ R d,ω(θ) D} [ ] [ ] f(ˆθ) min θ R df(θ) = f(ˆθ) min f(θ) + minf(θ) min θ C θ C θ R df(θ) NB: may replace min by best (non-linear) predictions θ Rdf(θ) 1. Uniform deviation bounds, with ˆθ argmin θ C ˆf(θ) f(ˆθ) min θ C Typically slow rate O ( 1 n ) f(θ) 2sup θ C ˆf(θ) f(θ) 2. More refined concentration results with faster rates

19 Slow rate for supervised learning Assumptions (f is the expected risk, ˆf the empirical risk) Ω(θ) = θ 2 (Euclidean norm) Linear predictors: θ(x) = θ Φ(x), with Φ(x) 2 R a.s. G-Lipschitz loss: f and ˆf are GR-Lipschitz on C = { θ 2 D} No assumptions regarding convexity

20 Slow rate for supervised learning Assumptions (f is the expected risk, ˆf the empirical risk) Ω(θ) = θ 2 (Euclidean norm) Linear predictors: θ(x) = θ Φ(x), with Φ(x) 2 R a.s. G-Lipschitz loss: f and ˆf are GR-Lipschitz on C = { θ 2 D} No assumptions regarding convexity With probability greater than 1 δ sup θ C ˆf(θ) f(θ) GRD n [2+ Expectated estimation error: E [ sup θ C 2log 2 δ ] ˆf(θ) f(θ) ] 4GRD n Using Rademacher averages (see, e.g., Boucheron et al., 2005) Lipschitz functions slow rate

21 Fast rate for supervised learning Assumptions (f is the expected risk, ˆf the empirical risk) Same as before (bounded features, Lipschitz loss) Regularized risks: f µ (θ) = f(θ)+ µ 2 θ 2 2 and ˆf µ (θ) = ˆf(θ)+ µ 2 θ 2 2 Convexity For any a > 0, with probability greater than 1 δ, for all θ R d, f µ (θ) min (η) (1+a)(ˆf µ (θ) min ˆfµ (η))+ 8(1+ 1 a )G2 R 2 (32+log 1 δ ) η R dfµ η R d µn Results from Sridharan, Srebro, and Shalev-Shwartz (2008) see also Boucheron and Massart (2011) and references therein Strongly convex functions fast rate Warning: µ should decrease with n to reduce approximation error

22 Complexity results in convex optimization Assumption: f convex on R d Classical generic algorithms (sub)gradient method/descent Accelerated gradient descent Newton method Key additional properties of f Lipschitz continuity, smoothness or strong convexity Key insight from Bottou and Bousquet (2008) In machine learning, no need to optimize below estimation error Key reference: Nesterov (2004)

23 Lipschitz continuity Bounded gradients of f (Lipschitz-continuity): the function f if convex, differentiable and has (sub)gradients uniformly bounded by B on the ball of center 0 and radius D: Machine learning θ R d, θ 2 D f (θ) 2 B with f(θ) = 1 n n i=1 l(y i,θ Φ(x i )) G-Lipschitz loss and R-bounded data: B = GR

24 Smoothness and strong convexity A function f : R d R is L-smooth if and only if it is differentiable and its gradient is L-Lipschitz-continuous θ 1,θ 2 R d, f (θ 1 ) f (θ 2 ) 2 L θ 1 θ 2 2 If f is twice differentiable: θ R d, f (θ) L Id smooth non smooth

25 Smoothness and strong convexity A function f : R d R is L-smooth if and only if it is differentiable and its gradient is L-Lipschitz-continuous θ 1,θ 2 R d, f (θ 1 ) f (θ 2 ) 2 L θ 1 θ 2 2 If f is twice differentiable: θ R d, f (θ) L Id Machine learning with f(θ) = 1 n n i=1 l(y i,θ Φ(x i )) Hessian covariance matrix 1 n n i=1 Φ(x i)φ(x i ) l-smooth loss and R-bounded data: L = lr 2

26 Smoothness and strong convexity A function f : R d R is µ-strongly convex if and only if θ 1,θ 2 R d, f(θ 1 ) f(θ 2 )+f (θ 2 ) (θ 1 θ 2 )+ µ 2 θ 1 θ If f is twice differentiable: θ R d, f (θ) µ Id convex strongly convex

27 Smoothness and strong convexity A function f : R d R is µ-strongly convex if and only if θ 1,θ 2 R d, f(θ 1 ) f(θ 2 )+f (θ 2 ) (θ 1 θ 2 )+ µ 2 θ 1 θ If f is twice differentiable: θ R d, f (θ) µ Id Machine learning with f(θ) = 1 n n i=1 l(y i,θ Φ(x i )) Hessian covariance matrix 1 n n i=1 Φ(x i)φ(x i ) Data with invertible covariance matrix (low correlation/dimension)

28 Smoothness and strong convexity A function f : R d R is µ-strongly convex if and only if θ 1,θ 2 R d, f(θ 1 ) f(θ 2 )+f (θ 2 ) (θ 1 θ 2 )+ µ 2 θ 1 θ If f is twice differentiable: θ R d, f (θ) µ Id Machine learning with f(θ) = 1 n n i=1 l(y i, θ Φ(x i )) Hessian covariance matrix 1 n n i=1 Φ(x i)φ(x i ) Data with invertible covariance matrix (low correlation/dimension) Adding regularization by µ 2 θ 2 creates additional bias unless µ is small

29 Summary of smoothness/convexity assumptions Bounded gradients of f (Lipschitz-continuity): the function f if convex, differentiable and has (sub)gradients uniformly bounded by B on the ball of center 0 and radius D: θ R d, θ 2 D f (θ) 2 B Smoothness of f: the function f is convex, differentiable with L-Lipschitz-continuous gradient f : θ 1,θ 2 R d, f (θ 1 ) f (θ 2 ) 2 L θ 1 θ 2 2 Strong convexity of f: The function f is strongly convex with respect to the norm, with convexity constant µ > 0: θ 1,θ 2 R d, f(θ 1 ) f(θ 2 )+f (θ 2 ) (θ 1 θ 2 )+ µ 2 θ 1 θ 2 2 2

30 Assumptions Subgradient method/descent f convex and B-Lipschitz-continuous on { θ 2 D} ( Algorithm: θ t = Π D θ t 1 2D ) B t f (θ t 1 ) Π D : orthogonal projection onto { θ 2 D} Bound: Three-line proof f ( 1 t t 1 k=0 θ k ) f(θ ) 2DB t Best possible convergence rate after O(d) iterations

31 Subgradient method/descent - proof - I Iteration: θ t = Π D (θ t 1 γ t f (θ t 1 )) with γ t = 2D B t Assumption: f (θ) 2 B and θ 2 D θ t θ 2 2 θ t 1 θ γ t f (θ t 1 ) 2 2 by contractivity of projections θ t 1 θ 2 2+B 2 γ 2 t 2γ t (θ t 1 θ ) f (θ t 1 ) because f (θ t 1 ) 2 B θ t 1 θ 2 2+B 2 γ 2 t 2γ t [ f(θt 1 ) f(θ ) ] (property of subgradients) leading to f(θ t 1 ) f(θ ) B2 γ t [ θt 1 θ 2 2γ 2 θ t θ 2 ] 2 t

32 Starting from Subgradient method/descent - proof - II t [ f(θu 1 ) f(θ ) ] u=1 f(θ t 1 ) f(θ ) B2 γ t 2 = = Using convexity: f t u=1 t u=1 ( 1 t t u=1 t u=1 B 2 γ u 2 B 2 γ u 2 B 2 γ u 2 B 2 γ u 2 t 1 k=0 + + t u=1 t 1 u=1 + θ k ) t 1 u=1 + 1 [ θt 1 θ 2 2γ 2 θ t θ 2 ] 2 t 1 [ θu 1 θ 2 2γ 2 θ u θ 2 ] 2 u ( θ u θ ) θ 0 θ θ t θ 2 2 2γ u+1 2γ u 2γ 1 2γ t 4D 2( 1 2γ u+1 1 2γ u ) + 4D 2 2γ 1 + 4D2 2γ t 2DB t with γ t = 2D B t f(θ ) 2DB t

33 Subgradient descent for machine learning Assumptions (f is the expected risk, ˆf the empirical risk) Linear predictors: θ(x) = θ Φ(x), with Φ(x) 2 R a.s. ˆf(θ) = 1 n n i=1 l(y i,φ(x i ) θ) G-Lipschitz loss: f and ˆf are GR-Lipschitz on C = { θ 2 D} Statistics: with probability greater than 1 δ ˆf(θ) f(θ) GRD [2+ n sup θ C 2log 2 δ ] Optimization: after t iterations of subgradient method ˆf(ˆθ) min η C ˆf(η) GRD t t = n iterations, with total running-time complexity of O(n 2 d)

34 Subgradient descent - strong convexity Assumptions f convex and B-Lipschitz-continuous on { θ 2 D} f µ-strongly convex ) 2 Algorithm: θ t = Π D (θ t 1 µ(t+1) f (θ t 1 ) Bound: f ( 2 t(t+1) t k=1 kθ k 1 ) f(θ ) 2B2 µ(t+1) Three-line proof Best possible convergence rate after O(d) iterations

35 Subgradient method - strong convexity - proof - I Iteration: θ t = Π D (θ t 1 γ t f (θ t 1 )) with γ t = 2 µ(t+1) Assumption: f (θ) 2 B and θ 2 D and µ-strong convexity of f θ t θ 2 2 θ t 1 θ γ t f (θ t 1 ) 2 2 by contractivity of projections θ t 1 θ 2 2+B 2 γt 2 2γ t (θ t 1 θ ) f (θ t 1 ) because f (θ t 1 ) 2 B θ t 1 θ 2 2+B 2 γt 2 [ 2γ t f(θt 1 ) f(θ )+ µ 2 θ t 1 θ 2 ] 2 (property of subgradients and strong convexity) leading to f(θ t 1 ) f(θ ) B2 γ t B 2 µ(t+1) + µ 2 [1 γ t µ ] θ t 1 θ γ t θ t θ 2 2 [t 1 2 ] θt 1 θ 2 2 µ(t+1) θ t θ 2 2 4

36 Subgradient method - strong convexity - proof - II From f(θ t 1 ) f(θ ) B2 µ(t+1) + µ 2 [t 1 2 ] θt 1 θ 2 2 µ(t+1) θ t θ t u [ f(θ u 1 ) f(θ ) ] u=1 u t=1 B 2 u µ(u+1) t [ u(u 1) θu 1 θ 2 2 u(u+1) θ u θ 2 2] u=1 B2 t µ + 1 [ 0 t(t+1) θt θ 2 B 4 2] 2 t µ Using convexity: f ( 2 t(t+1) t u=1 uθ u 1 ) f(θ ) 2B2 t+1

37 (smooth) gradient descent Assumptions f convex with L-Lipschitz-continuous gradient Minimum attained at θ Algorithm: Bound: Three-line proof θ t = θ t 1 1 L f (θ t 1 ) f(θ t ) f(θ ) 2L θ 0 θ 2 t+4 Not best possible convergence rate after O(d) iterations

38 (smooth) gradient descent - strong convexity Assumptions f convex with L-Lipschitz-continuous gradient f µ-strongly convex Algorithm: Bound: θ t = θ t 1 1 L f (θ t 1 ) f(θ t ) f(θ ) (1 µ/l) t[ f(θ 0 ) f(θ ) ] Three-line proof Adaptivity of gradient descent to problem difficulty Line search

39 Accelerated gradient methods (Nesterov, 1983) Assumptions f convex with L-Lipschitz-cont. gradient, min. attained at θ Algorithm: θ t = η t 1 1 L f (η t 1 ) Bound: η t = θ t + t 1 t+2 (θ t θ t 1 ) f(θ t ) f(θ ) 2L θ 0 θ 2 (t+1) 2 Ten-line proof (see, e.g., Schmidt, Le Roux, and Bach, 2011) Not improvable Extension to strongly convex functions

40 Optimization for sparsity-inducing norms (see Bach, Jenatton, Mairal, and Obozinski, 2011) Gradient descent as a proximal method (differentiable functions) θ t+1 = arg min θ R df(θ t)+(θ θ t ) f(θ t )+ L 2 θ θ t 2 2 θ t+1 = θ t 1 L f(θ t)

41 Optimization for sparsity-inducing norms (see Bach, Jenatton, Mairal, and Obozinski, 2011) Gradient descent as a proximal method (differentiable functions) θ t+1 = arg min θ R df(θ t)+(θ θ t ) f(θ t )+ L 2 θ θ t 2 2 θ t+1 = θ t 1 L f(θ t) Problems of the form: min θ R df(θ)+µω(θ) θ t+1 = arg min θ R df(θ t)+(θ θ t ) f(θ t )+µω(θ)+ L 2 θ θ t 2 2 Ω(θ) = θ 1 Thresholded gradient descent Similar convergence rates than smooth optimization Acceleration methods (Nesterov, 2007; Beck and Teboulle, 2009)

42 Summary: minimizing convex functions Assumption: f convex Gradient descent: θ t = θ t 1 γ t f (θ t 1 ) O(1/ t) convergence rate for non-smooth convex functions O(1/t) convergence rate for smooth convex functions O(e ρt ) convergence rate for strongly smooth convex functions Newton method: θ t = θ t 1 f (θ t 1 ) 1 f (θ t 1 ) O ( e ρ2t ) convergence rate

43 Summary: minimizing convex functions Assumption: f convex Gradient descent: θ t = θ t 1 γ t f (θ t 1 ) O(1/ t) convergence rate for non-smooth convex functions O(1/t) convergence rate for smooth convex functions O(e ρt ) convergence rate for strongly smooth convex functions Newton method: θ t = θ t 1 f (θ t 1 ) 1 f (θ t 1 ) O ( e ρ2t ) convergence rate Key insights from Bottou and Bousquet (2008) 1. In machine learning, no need to optimize below statistical error 2. In machine learning, cost functions are averages Stochastic approximation

44 Outline 1. Large-scale machine learning and optimization Traditional statistical analysis Classical methods for convex optimization 2. Non-smooth stochastic approximation Stochastic (sub)gradient and averaging Non-asymptotic results and lower bounds Strongly convex vs. non-strongly convex 3. Smooth stochastic approximation algorithms Asymptotic and non-asymptotic results 4. Beyond decaying step-sizes 5. Finite data sets

45 Stochastic approximation Goal: Minimizing a function f defined on R d given only unbiased estimates f n(θ n ) of its gradients f (θ n ) at certain points θ n R d Stochastic approximation (much) broader applicability beyond convex optimization θ n = θ n 1 γ n h n (θ n 1 ) with E [ h n (θ n 1 ) θ n 1 ] = h(θn 1 ) Beyond convex problems, i.i.d assumption, finite dimension, etc. Typically asymptotic results See, e.g., Kushner and Yin (2003); Benveniste et al. (2012)

46 Stochastic approximation Goal: Minimizing a function f defined on R d given only unbiased estimates f n(θ n ) of its gradients f (θ n ) at certain points θ n R d Machine learning - statistics loss for a single pair of observations: f n (θ) = l(y n,θ Φ(x n )) f(θ) = Ef n (θ) = El(y n,θ Φ(x n )) = generalization error Expected gradient: f (θ) = Ef n(θ) = E { l (y n,θ Φ(x n ))Φ(x n ) } Non-asymptotic results Number of iterations = number of observations

47 Stochastic approximation Relationship to online learning Minimize f(θ) = E z l(θ,z) = generalization error of θ Using the gradients of single i.i.d. observations

48 Stochastic approximation Relationship to online learning Minimize f(θ) = E z l(θ,z) = generalization error of θ Using the gradients of single i.i.d. observations Batch learning Finite set of observations: z 1,...,z n Empirical risk: ˆf(θ) = 1 n n k=1 l(θ,z i) Estimator ˆθ = Minimizer of ˆf(θ) over a certain class Θ Generalization bound using uniform concentration results

49 Stochastic approximation Relationship to online learning Minimize f(θ) = E z l(θ,z) = generalization error of θ Using the gradients of single i.i.d. observations Batch learning Finite set of observations: z 1,...,z n Empirical risk: ˆf(θ) = 1 n n k=1 l(θ,z i) Estimator ˆθ = Minimizer of ˆf(θ) over a certain class Θ Generalization bound using uniform concentration results Online learning Update ˆθ n after each new (potentially adversarial) observation z n Cumulative loss: 1 n n k=1 l(ˆθ k 1,z k ) Online to batch through averaging (Cesa-Bianchi et al., 2004)

50 Convex stochastic approximation Key properties of f and/or f n Smoothness: f B-Lipschitz continuous, f L-Lipschitz continuous Strong convexity: f µ-strongly convex

51 Convex stochastic approximation Key properties of f and/or f n Smoothness: f B-Lipschitz continuous, f L-Lipschitz continuous Strong convexity: f µ-strongly convex Key algorithm: Stochastic gradient descent (a.k.a. Robbins-Monro) θ n = θ n 1 γ n f n(θ n 1 ) Polyak-Ruppert averaging: θ n = 1 n 1 n k=0 θ k Which learning rate sequence γ n? Classical setting: γ n = Cn α

52 Convex stochastic approximation Key properties of f and/or f n Smoothness: f B-Lipschitz continuous, f L-Lipschitz continuous Strong convexity: f µ-strongly convex Key algorithm: Stochastic gradient descent (a.k.a. Robbins-Monro) θ n = θ n 1 γ n f n(θ n 1 ) Polyak-Ruppert averaging: θ n = 1 n 1 n k=0 θ k Which learning rate sequence γ n? Classical setting: γ n = Cn α Desirable practical behavior Applicable (at least) to classical supervised learning problems Robustness to (potentially unknown) constants (L,B,µ) Adaptivity to difficulty of the problem (e.g., strong convexity)

53 Assumptions Stochastic subgradient descent/method f n convex and B-Lipschitz-continuous on { θ 2 D} (f n ) i.i.d. functions such that Ef n = f θ global optimum of f on { θ 2 D} ( Algorithm: θ n = Π D θ n 1 2D ) B n f n(θ n 1 ) Bound: Ef ( 1 n n 1 k=0 θ k ) f(θ ) 2DB n Same three-line proof as in the deterministic case Minimax convergence rate Running-time complexity: O(dn) after n iterations

54 Stochastic subgradient method - proof - I Iteration: θ n = Π D (θ n 1 γ n f n(θ n 1 )) with γ n = 2D B n F n : information up to time n f n(θ) 2 B and θ 2 D, unbiased gradients/functions E(f n F n 1 ) = f θ n θ 2 2 θ n 1 θ γ n f (θ n 1 ) 2 2 by contractivity of projections θ n 1 θ 2 2 +B 2 γ 2 n 2γ n (θ n 1 θ ) f n(θ n 1 ) because f n(θ n 1 ) 2 B E [ θ n θ 2 2 F n 1 ] θn 1 θ 2 2+B 2 γ 2 n 2γ n (θ n 1 θ ) f (θ n 1 ) θ n 1 θ 2 2+B 2 γ 2 n 2γ n [ f(θn 1 ) f(θ ) ] (subgradient property) E θ n θ 2 2 E θ n 1 θ 2 2+B 2 γ 2 n 2γ n [ Ef(θn 1 ) f(θ ) ] leading to Ef(θ n 1 ) f(θ ) B2 γ n [ ] E θn 1 θ 2 2γ 2 E θ n θ 2 2 n

55 Stochastic subgradient method - proof - II Starting from Ef(θ n 1 ) f(θ ) B2 γ n [ ] E θn 1 θ 2 2 E θ n θ 2 2 2γ n n [ Ef(θu 1 ) f(θ ) ] u=1 n u=1 n u=1 B 2 γ u 2 B 2 γ u 2 + n u=1 1 [ E θu 1 θ 2 2γ 2 E θ u θ 2 ] 2 u + 4D2 2γ n 2DB n with γ n = 2D B n Using convexity: Ef ( 1 n n 1 k=0 θ k ) f(θ ) 2DB n

56 Stochastic subgradient descent - strong convexity - I Assumptions f n convex and B-Lipschitz-continuous (f n ) i.i.d. functions such that Ef n = f f µ-strongly convex on { θ 2 D} θ global optimum of f over { θ 2 D} ) 2 Algorithm: θ n = Π D (θ n 1 µ(n+1) f n(θ n 1 ) Bound: ( 2 Ef n(n+1) n ) kθ k 1 k=1 f(θ ) 2B2 µ(n+1) Same three-line proof than in the deterministic case Minimax convergence rate

57 Stochastic subgradient descent - strong convexity - II Assumptions f n convex and B-Lipschitz-continuous (f n ) i.i.d. functions such that Ef n = f θ global optimum of g = f + µ No compactness assumption - no projections Algorithm: 2 2 [ ] θ n = θ n 1 µ(n+1) g n(θ n 1 ) = θ n 1 f µ(n+1) n (θ n 1 )+µθ n 1 ( 2 Bound: Eg n(n+1) n ) kθ k 1 k=1 Minimax convergence rate g(θ ) 2B2 µ(n+1)

58 Outline 1. Large-scale machine learning and optimization Traditional statistical analysis Classical methods for convex optimization 2. Non-smooth stochastic approximation Stochastic (sub)gradient and averaging Non-asymptotic results and lower bounds Strongly convex vs. non-strongly convex 3. Smooth stochastic approximation algorithms Asymptotic and non-asymptotic results 4. Beyond decaying step-sizes 5. Finite data sets

59 Convex stochastic approximation Existing work Known global minimax rates of convergence for non-smooth problems (Nemirovsky and Yudin, 1983; Agarwal et al., 2012) Strongly convex: O((µn) 1 ) Attainedbyaveragedstochasticgradientdescentwithγ n (µn) 1 Non-strongly convex: O(n 1/2 ) Attained by averaged stochastic gradient descent with γ n n 1/2 Bottou and Le Cun (2005); Bottou and Bousquet (2008); Hazan et al. (2007); Shalev-Shwartz and Srebro (2008); Shalev-Shwartz et al. (2007, 2009); Xiao (2010); Duchi and Singer (2009); Nesterov and Vial (2008); Nemirovski et al. (2009)

60 Convex stochastic approximation Existing work Known global minimax rates of convergence for non-smooth problems (Nemirovsky and Yudin, 1983; Agarwal et al., 2012) Strongly convex: O((µn) 1 ) Attainedbyaveragedstochasticgradientdescentwithγ n (µn) 1 Non-strongly convex: O(n 1/2 ) Attained by averaged stochastic gradient descent with γ n n 1/2 Many contributions in optimization and online learning: Bottou and Le Cun (2005); Bottou and Bousquet (2008); Hazan et al. (2007); Shalev-Shwartz and Srebro (2008); Shalev-Shwartz et al. (2007, 2009); Xiao (2010); Duchi and Singer (2009); Nesterov and Vial (2008); Nemirovski et al. (2009)

61 Convex stochastic approximation Existing work Known global minimax rates of convergence for non-smooth problems (Nemirovsky and Yudin, 1983; Agarwal et al., 2012) Strongly convex: O((µn) 1 ) Attainedbyaveragedstochasticgradientdescentwithγ n (µn) 1 Non-strongly convex: O(n 1/2 ) Attained by averaged stochastic gradient descent with γ n n 1/2 Asymptotic analysis of averaging (Polyak and Juditsky, 1992; Ruppert, 1988) All step sizes γ n = Cn α with α (1/2,1) lead to O(n 1 ) for smooth strongly convex problems A single algorithm with global adaptive convergence rate for smooth problems?

62 Convex stochastic approximation Existing work Known global minimax rates of convergence for non-smooth problems (Nemirovsky and Yudin, 1983; Agarwal et al., 2012) Strongly convex: O((µn) 1 ) Attainedbyaveragedstochasticgradientdescentwithγ n (µn) 1 Non-strongly convex: O(n 1/2 ) Attained by averaged stochastic gradient descent with γ n n 1/2 Asymptotic analysis of averaging (Polyak and Juditsky, 1992; Ruppert, 1988) All step sizes γ n = Cn α with α (1/2,1) lead to O(n 1 ) for smooth strongly convex problems Non-asymptotic analysis for smooth problems? problems with convergence rate O(min{1/µn,1/ n})

63 Smoothness/convexity assumptions Iteration: θ n = θ n 1 γ n f n(θ n 1 ) Polyak-Ruppert averaging: θ n = 1 n n 1 k=0 θ k Smoothness of f n : For each n 1, the function f n is a.s. convex, differentiable with L-Lipschitz-continuous gradient f n: Smooth loss and bounded data Strong convexity of f: The function f is strongly convex with respect to the norm, with convexity constant µ > 0: Invertible population covariance matrix or regularization by µ 2 θ 2

64 Summary of new results (Bach and Moulines, 2011) Stochastic gradient descent with learning rate γ n = Cn α Strongly convex smooth objective functions Old: O(n 1 ) rate achieved without averaging for α = 1 New: O(n 1 ) rate achieved with averaging for α [1/2,1] Non-asymptotic analysis with explicit constants Forgetting of initial conditions Robustness to the choice of C

65 Summary of new results (Bach and Moulines, 2011) Stochastic gradient descent with learning rate γ n = Cn α Strongly convex smooth objective functions Old: O(n 1 ) rate achieved without averaging for α = 1 New: O(n 1 ) rate achieved with averaging for α [1/2,1] Non-asymptotic analysis with explicit constants Forgetting of initial conditions Robustness to the choice of C Convergence rates for E θ n θ 2 and E θ n θ 2 ( σ 2 γ ) n no averaging: O +O(e µnγ n ) θ 0 θ 2 µ averaging: trh(θ ) 1 ( +µ 1 O(n 2α +n 2+α θ0 θ 2 ) )+O n µ 2 n 2

66 Robustness to wrong constants for γ n = Cn α f(θ) = 1 2 θ 2 with i.i.d. Gaussian noise (d = 1) Left: α = 1/2 Right: α = 1 log[f(θ n ) f ] 5 0 α = 1/2 sgd C=1/5 ave C=1/5 sgd C=1 ave C=1 sgd C=5 ave C=5 log[f(θ n ) f ] 5 0 α = 1 sgd C=1/5 ave C=1/5 sgd C=1 ave C=1 sgd C=5 ave C= log(n) log(n) See also

67 Summary of new results (Bach and Moulines, 2011) Stochastic gradient descent with learning rate γ n = Cn α Strongly convex smooth objective functions Old: O(n 1 ) rate achieved without averaging for α = 1 New: O(n 1 ) rate achieved with averaging for α [1/2,1] Non-asymptotic analysis with explicit constants

68 Summary of new results (Bach and Moulines, 2011) Stochastic gradient descent with learning rate γ n = Cn α Strongly convex smooth objective functions Old: O(n 1 ) rate achieved without averaging for α = 1 New: O(n 1 ) rate achieved with averaging for α [1/2,1] Non-asymptotic analysis with explicit constants Non-strongly convex smooth objective functions Old: O(n 1/2 ) rate achieved with averaging for α = 1/2 New: O(max{n 1/2 3α/2,n α/2,n α 1 }) rate achieved without averaging for α [1/3,1] Take-home message Use α = 1/2 with averaging to be adaptive to strong convexity

69 Beyond stochastic gradient method Adding a proximal step Goal: min θ R df(θ)+ω(θ) = Ef n(θ)+ω(θ) Replace recursion θ n = θ n 1 γ n f n(θ n ) by θ n = min θ R d θ θ n 1 +γ n f n(θ n ) 2 2 +CΩ(θ) Xiao (2010); Hu et al. (2009) May be accelerated Ghadimi and Lan (2013) Related frameworks Regularized dual averaging (Nesterov, 2009; Xiao, 2010) Mirror descent (Nemirovski et al., 2009; Lan et al., 2012)

70 Outline 1. Large-scale machine learning and optimization Traditional statistical analysis Classical methods for convex optimization 2. Non-smooth stochastic approximation Stochastic (sub)gradient and averaging Non-asymptotic results and lower bounds Strongly convex vs. non-strongly convex 3. Smooth stochastic approximation algorithms Asymptotic and non-asymptotic results 4. Beyond decaying step-sizes 5. Finite data sets

71 Convex stochastic approximation Existing work Known global minimax rates of convergence for non-smooth problems (Nemirovsky and Yudin, 1983; Agarwal et al., 2012) Strongly convex: O((µn) 1 ) Attainedbyaveragedstochasticgradientdescentwithγ n (µn) 1 Non-strongly convex: O(n 1/2 ) Attained by averaged stochastic gradient descent with γ n n 1/2 Asymptotic analysis of averaging (Polyak and Juditsky, 1992; Ruppert, 1988) All step sizes γ n = Cn α with α (1/2,1) lead to O(n 1 ) for smooth strongly convex problems A single adaptive algorithm for smooth problems with convergence rate O(min{1/µn,1/ n}) in all situations?

72 Adaptive algorithm for logistic regression Logistic regression: (Φ(x n ),y n ) R d { 1,1} Single data point: f n (θ) = log(1+exp( y n θ Φ(x n ))) Generalization error: f(θ) = Ef n (θ)

73 Adaptive algorithm for logistic regression Logistic regression: (Φ(x n ),y n ) R d { 1,1} Single data point: f n (θ) = log(1+exp( y n θ Φ(x n ))) Generalization error: f(θ) = Ef n (θ) Cannot be strongly convex local strong convexity unless restricted to θ Φ(x n ) M (and with constants e M ) µ = lowest eigenvalue of the Hessian at the optimum f (θ ) logistic loss

74 Adaptive algorithm for logistic regression Logistic regression: (Φ(x n ),y n ) R d { 1,1} Single data point: f n (θ) = log(1+exp( y n θ Φ(x n ))) Generalization error: f(θ) = Ef n (θ) Cannot be strongly convex local strong convexity unless restricted to θ Φ(x n ) M (and with constants e M ) µ = lowest eigenvalue of the Hessian at the optimum f (θ ) n steps of averaged SGD with constant step-size 1/ ( 2R 2 n ) with R = radius of data (Bach, 2013): Ef( θ n ) f(θ ) min { } 1 n, R2 (15+5R θ0 θ ) 4 nµ Proof based on self-concordance (Nesterov and Nemirovski, 1994)

75 Adaptive algorithm for logistic regression Logistic regression: (Φ(x n ),y n ) R d { 1,1} Single data point: f n (θ) = log(1+exp( y n θ Φ(x n ))) Generalization error: f(θ) = Ef n (θ) Cannot be strongly convex local strong convexity unless restricted to θ Φ(x n ) M (and with constants e M ) µ = lowest eigenvalue of the Hessian at the optimum f (θ ) n steps of averaged SGD with constant step-size 1/ ( 2R 2 n ) with R = radius of data (Bach, 2013): Ef( θ n ) f(θ ) min { } 1 n, R2 (15+5R θ0 θ ) 4 nµ A single adaptive algorithm for smooth problems with convergence rate O(1/n) in all situations?

76 Least-mean-square (LMS) algorithm Least-squares: f(θ) = 1 2 E[ (y n Φ(x n ) θ) 2] with θ R d SGD = least-mean-square algorithm (see, e.g., Macchi, 1995) usually studied without averaging and decreasing step-sizes with strong convexity assumption E [ Φ(x n )Φ(x n ) ] = H µ Id

77 Least-mean-square (LMS) algorithm Least-squares: f(θ) = 1 2 E[ (y n Φ(x n ) θ) 2] with θ R d SGD = least-mean-square algorithm (see, e.g., Macchi, 1995) usually studied without averaging and decreasing step-sizes with strong convexity assumption E [ Φ(x n )Φ(x n ) ] = H µ Id New analysis for averaging and constant step-size γ = 1/(4R 2 ) Assume Φ(x n ) R and y n Φ(x n ) θ σ almost surely No assumption regarding lowest eigenvalues of H Main result: Ef( θ n ) f(θ ) 4σ2 d n + 2R2 θ 0 θ 2 n Matches statistical lower bound (Tsybakov, 2003) Non-asymptotic robust version of Györfi and Walk (1996)

78 Least-mean-square (LMS) algorithm Least-squares: f(θ) = 1 2 E[ (y n Φ(x n ) θ) 2] with θ R d SGD = least-mean-square algorithm (see, e.g., Macchi, 1995) usually studied without averaging and decreasing step-sizes with strong convexity assumption E [ Φ(x n )Φ(x n ) ] = H µ Id New analysis for averaging and constant step-size γ = 1/(4R 2 ) Assume Φ(x n ) R and y n Φ(x n ) θ σ almost surely No assumption regarding lowest eigenvalues of H Main result: Ef( θ n ) f(θ ) 4σ2 d n + 2R2 θ 0 θ 2 n Improvement of bias term (Flammarion and Bach, 2014): { R 2 θ 0 θ 2 min, R4 (θ 0 θ ) H 1 (θ 0 θ ) } n n 2

79 Least-mean-square (LMS) algorithm Least-squares: f(θ) = 1 2 E[ (y n Φ(x n ) θ) 2] with θ R d SGD = least-mean-square algorithm (see, e.g., Macchi, 1995) usually studied without averaging and decreasing step-sizes with strong convexity assumption E [ Φ(x n )Φ(x n ) ] = H µ Id New analysis for averaging and constant step-size γ = 1/(4R 2 ) Assume Φ(x n ) R and y n Φ(x n ) θ σ almost surely No assumption regarding lowest eigenvalues of H Main result: Ef( θ n ) f(θ ) 4σ2 d n + 2R2 θ 0 θ 2 n Extension to Hilbert spaces (Dieuleveult and Bach, 2014): Achieves minimax statistical rates given decay of spectrum of H

80 Least-squares - Proof technique LMS recursion with ε n = y n Φ(x n ) θ : θ n θ = [ I γφ(x n )Φ(x n ) ] (θ n 1 θ )+γε n Φ(x n ) Simplified LMS recursion: with H = E [ Φ(x n )Φ(x n ) ] θ n θ = [ I γh ] (θ n 1 θ )+γε n Φ(x n ) Direct proof technique of Polyak and Juditsky (1992), e.g., θ n θ = [ I γh ] n (θ0 θ )+γ Exact computations n k=1 [ I γh ] n kεk Φ(x k ) Infinite expansion of Aguech, Moulines, and Priouret(2000) in powers of γ

81 Markov chain interpretation of constant step sizes LMS recursion for f n (θ) = 1 2 ( yn Φ(x n ) θ ) 2 θ n = θ n 1 γ ( Φ(x n ) θ n 1 y n ) Φ(xn ) The sequence (θ n ) n is a homogeneous Markov chain convergence to a stationary distribution π γ with expectation θ def γ = θπ γ (dθ)

82 Markov chain interpretation of constant step sizes LMS recursion for f n (θ) = 1 2 ( yn Φ(x n ) θ ) 2 θ n = θ n 1 γ ( Φ(x n ) θ n 1 y n ) Φ(xn ) The sequence (θ n ) n is a homogeneous Markov chain convergence to a stationary distribution π γ with expectation θ def γ = θπ γ (dθ) For least-squares, θ γ = θ θ n does not converge to θ but oscillates around it oscillations of order γ cf. Kaczmarz method (Strohmer and Vershynin, 2009) Ergodic theorem: Averaged iterates converge to θ γ = θ at rate O(1/n)

83 Simulations - synthetic examples Gaussian distributions - d = 20 0 synthetic square log 10 [f(θ) f(θ * )] /2R 2 1/8R 2 4 1/32R 2 1/2R 2 n 1/ log 10 (n)

84 Simulations - benchmarks alpha (d = 500, n = ), news (d = , n = ) log 10 [f(θ) f(θ * )] alpha square C=1 test 1/R 2 1/R 2 n 1/2 SAG log (n) alpha square C=opt test C/R 2 C/R 2 n 1/2 SAG log (n) news square C=1 test 0.2 news square C=opt test log 10 [f(θ) f(θ * )] /R 2 1/R 2 n 1/2 SAG log (n) C/R 2 C/R 2 n 1/2 SAG log 10 (n)

85 Beyond least-squares - Markov chain interpretation Recursion θ n = θ n 1 γf n(θ n 1 ) also defines a Markov chain Stationary distribution π γ such that f (θ)π γ (dθ) = 0 When f is not linear, f ( θπ γ (dθ)) f (θ)π γ (dθ) = 0

86 Beyond least-squares - Markov chain interpretation Recursion θ n = θ n 1 γf n(θ n 1 ) also defines a Markov chain Stationary distribution π γ such that f (θ)π γ (dθ) = 0 When f is not linear, f ( θπ γ (dθ)) f (θ)π γ (dθ) = 0 θ n oscillates around the wrong value θ γ θ moreover, θ θ n = O p ( γ) Ergodic theorem averaged iterates converge to θ γ θ at rate O(1/n) moreover, θ θ γ = O(γ) (Bach, 2013) NB: coherent with earlier results by Nedic and Bertsekas (2000)

87 Simulations - synthetic examples Gaussian distributions - d = 20 0 synthetic logistic 1 log 10 [f(θ) f(θ * )] 1 2 1/2R 2 3 1/8R 2 4 1/32R 2 5 1/2R 2 n 1/ log 10 (n)

88 Restoring convergence through online Newton steps Known facts 1. Averaged SGD with γ n n 1/2 leads to robust rate O(n 1/2 ) for all convex functions 2. Averaged SGD with γ n constant leads to robust rate O(n 1 ) for all convex quadratic functions 3. Newton s method squares the error at each iteration for smooth functions 4. A single step of Newton s method is equivalent to minimizing the quadratic Taylor expansion Online Newton step Rate: O((n 1/2 ) 2 +n 1 ) = O(n 1 ) Complexity: O(d) per iteration for linear predictions

89 Restoring convergence through online Newton steps Known facts 1. Averaged SGD with γ n n 1/2 leads to robust rate O(n 1/2 ) for all convex functions 2. Averaged SGD with γ n constant leads to robust rate O(n 1 ) for all convex quadratic functions 3. Newton s method squares the error at each iteration for smooth functions 4. A single step of Newton s method is equivalent to minimizing the quadratic Taylor expansion Online Newton step Rate: O((n 1/2 ) 2 +n 1 ) = O(n 1 ) Complexity: O(d) per iteration for linear predictions

90 Restoring convergence through online Newton steps The Newton step for f = Ef n (θ) def = E [ l(y n,θ Φ(x n )) ] at θ is equivalent to minimizing the quadratic approximation g(θ) = f( θ)+f ( θ) (θ θ)+ 1 2 (θ θ) f ( θ)(θ θ) = f( θ)+ef n( θ) (θ θ)+ 1 2 (θ θ) Ef n( θ)(θ θ) [ = E f( θ)+f n( θ) (θ θ)+ 1 2 (θ θ) f n( θ)(θ θ) ]

91 Restoring convergence through online Newton steps The Newton step for f = Ef n (θ) def = E [ l(y n,θ Φ(x n )) ] at θ is equivalent to minimizing the quadratic approximation g(θ) = f( θ)+f ( θ) (θ θ)+ 1 2 (θ θ) f ( θ)(θ θ) = f( θ)+ef n( θ) (θ θ)+ 1 2 (θ θ) Ef n( θ)(θ θ) [ = E f( θ)+f n( θ) (θ θ)+ 1 2 (θ θ) f n( θ)(θ θ) ] Complexity of least-mean-square recursion for g is O(d) θ n = θ n 1 γ [ f n( θ)+f n( θ)(θ n 1 θ) ] f n( θ) = l (y n,φ(x n ) θ)φ(xn )Φ(x n ) has rank one New online Newton step without computing/inverting Hessians

92 Choice of support point for online Newton step Two-stage procedure (1) Run n/2 iterations of averaged SGD to obtain θ (2) Run n/2 iterations of averaged constant step-size LMS Reminiscent of one-step estimators (see, e.g., Van der Vaart, 2000) Provable convergence rate of O(d/n) for logistic regression Additional assumptions but no strong convexity

93 Logistic regression - Proof technique Using generalized self-concordance of ϕ : u log(1+e u ): ϕ (u) ϕ (u) NB: difference with regular self-concordance: ϕ (u) 2ϕ (u) 3/2 Using novel high-probability convergence results for regular averaged stochastic gradient descent Requires assumption on the kurtosis in every direction, i.e., E(Φ(x n ) η) 4 κ [ E(Φ(x n ) η) 2] 2

94 Choice of support point for online Newton step Two-stage procedure (1) Run n/2 iterations of averaged SGD to obtain θ (2) Run n/2 iterations of averaged constant step-size LMS Reminiscent of one-step estimators (see, e.g., Van der Vaart, 2000) Provable convergence rate of O(d/n) for logistic regression Additional assumptions but no strong convexity

95 Choice of support point for online Newton step Two-stage procedure (1) Run n/2 iterations of averaged SGD to obtain θ (2) Run n/2 iterations of averaged constant step-size LMS Reminiscent of one-step estimators (see, e.g., Van der Vaart, 2000) Provable convergence rate of O(d/n) for logistic regression Additional assumptions but no strong convexity Update at each iteration using the current averaged iterate Recursion: θ n = θ n 1 γ [ f n( θ n 1 )+f n( θ n 1 )(θ n 1 θ n 1 ) ] No provable convergence rate (yet) but best practical behavior Note (dis)similarity with regular SGD: θ n = θ n 1 γf n(θ n 1 )

96 Simulations - synthetic examples Gaussian distributions - d = 20 0 synthetic logistic 1 0 synthetic logistic 2 log 10 [f(θ) f(θ * )] /2R 2 3 1/8R 2 3 every 2 p 4 1/32R 2 every iter. 4 2 step 5 1/2R 2 n 1/2 5 2 step dbl log 10 (n) log 10 (n) log 10 [f(θ) f(θ * )] 1

97 Simulations - benchmarks alpha (d = 500, n = ), news (d = , n = ) alpha logistic C=1 test alpha logistic C=opt test log 10 [f(θ) f(θ * )] /R /R 2 n 1/2 SAG 2 Adagrad Newton log (n) C/R C/R 2 n 1/2 SAG 2 Adagrad Newton log (n) news logistic C=1 test 0.2 news logistic C=opt test log 10 [f(θ) f(θ * )] /R /R 2 n 1/2 SAG 0.8 Adagrad 1 Newton log (n) C/R C/R 2 n 1/2 SAG 0.8 Adagrad 1 Newton log 10 (n)

98 Outline 1. Large-scale machine learning and optimization Traditional statistical analysis Classical methods for convex optimization 2. Non-smooth stochastic approximation Stochastic (sub)gradient and averaging Non-asymptotic results and lower bounds Strongly convex vs. non-strongly convex 3. Smooth stochastic approximation algorithms Asymptotic and non-asymptotic results 4. Beyond decaying step-sizes 5. Finite data sets

99 Going beyond a single pass over the data Stochastic approximation Assumes infinite data stream Observations are used only once Directly minimizes testing cost E (x,y) l(y,θ Φ(x))

100 Going beyond a single pass over the data Stochastic approximation Assumes infinite data stream Observations are used only once Directly minimizes testing cost E (x,y) l(y,θ Φ(x)) Machine learning practice Finite data set (x 1,y 1,...,x n,y n ) Multiple passes Minimizes training cost 1 n n i=1 l(y i,θ Φ(x i )) Need to regularize (e.g., by the l 2 -norm) to avoid overfitting Goal: minimize g(θ) = 1 n n f i (θ) i=1

101 Stochastic vs. deterministic methods Minimizing g(θ) = 1 n n f i (θ) with f i (θ) = l ( y i,θ Φ(x i ) ) +µω(θ) i=1 Batchgradientdescent:θ t = θ t 1 γ t g (θ t 1 ) = θ t 1 γ t n Linear (e.g., exponential) convergence rate in O(e αt ) Iteration complexity is linear in n (with line search) n f i(θ t 1 ) i=1

102 Stochastic vs. deterministic methods Minimizing g(θ) = 1 n n f i (θ) with f i (θ) = l ( y i,θ Φ(x i ) ) +µω(θ) i=1 Batchgradientdescent:θ t = θ t 1 γ t g (θ t 1 ) = θ t 1 γ t n Linear (e.g., exponential) convergence rate in O(e αt ) Iteration complexity is linear in n (with line search) n f i(θ t 1 ) i=1 Stochastic gradient descent: θ t = θ t 1 γ t f i(t) (θ t 1) Sampling with replacement: i(t) random element of {1,...,n} Convergence rate in O(1/t) Iteration complexity is independent of n (step size selection?)

103 Stochastic vs. deterministic methods Goal = best of both worlds: Linear rate with O(1) iteration cost Robustness to step size log(excess cost) stochastic deterministic time

104 Stochastic vs. deterministic methods Goal = best of both worlds: Linear rate with O(1) iteration cost Robustness to step size log(excess cost) hybrid time stochastic deterministic

105 Accelerating gradient methods - Related work Nesterov acceleration Nesterov (1983, 2004) Better linear rate but still O(n) iteration cost Hybrid methods, incremental average gradient, increasing batch size Bertsekas (1997); Blatt et al. (2008); Friedlander and Schmidt (2011) Linear rate, but iterations make full passes through the data.

106 Accelerating gradient methods - Related work Momentum, gradient/iterate averaging, stochastic version of accelerated batch gradient methods Polyak and Juditsky (1992); Tseng (1998); Sunehag et al. (2009); Ghadimi and Lan (2010); Xiao (2010) Can improve constants, but still have sublinear O(1/t) rate Constant step-size stochastic gradient (SG), accelerated SG Kesten (1958); Delyon and Juditsky (1993); Solodov (1998); Nedic and Bertsekas (2000) Linear convergence, but only up to a fixed tolerance. Stochastic methods in the dual Shalev-Shwartz and Zhang (2012) Similar linear rate but limited choice for the f i s

107 Stochastic average gradient (Le Roux, Schmidt, and Bach, 2012) Stochastic average gradient (SAG) iteration Keep in memory the gradients of all functions f i, i = 1,...,n Random selection i(t) {1,...,n} with replacement { Iteration: θ t = θ t 1 γ n t f yi t with yi t n = i (θ t 1 ) if i = i(t) otherwise i=1 y t 1 i

108 Stochastic average gradient (Le Roux, Schmidt, and Bach, 2012) Stochastic average gradient (SAG) iteration Keep in memory the gradients of all functions f i, i = 1,...,n Random selection i(t) {1,...,n} with replacement { Iteration: θ t = θ t 1 γ n t f yi t n with yt i = i (θ t 1 ) if i = i(t) otherwise i=1 y t 1 i Stochastic version of incremental average gradient(blatt et al., 2008) Extra memory requirement Supervised machine learning If f i (θ) = l i (y i,φ(x i ) θ), then f i (θ) = l i (y i,φ(x i ) θ)φ(x i ) Only need to store n real numbers

109 Stochastic average gradient - Convergence analysis Assumptions Each f i is L-smooth, i = 1,...,n g= 1 n n i=1 f i is µ-strongly convex (with potentially µ = 0) constant step size γ t = 1/(16L) initialization with one pass of averaged SGD

110 Stochastic average gradient - Convergence analysis Assumptions Each f i is L-smooth, i = 1,...,n g= 1 n n i=1 f i is µ-strongly convex (with potentially µ = 0) constant step size γ t = 1/(16L) initialization with one pass of averaged SGD Strongly convex case (Le Roux et al., 2012, 2013) E [ g(θ t ) g(θ ) ] ( 8σ 2 nµ + 4L θ 0 θ 2 n ) exp ( { 1 tmin 8n, µ }) 16L Linear (exponential) convergence rate with ( O(1) iteration { cost 1 After one pass, reduction of cost by exp min 8, nµ }) 16L

111 Stochastic average gradient - Convergence analysis Assumptions Each f i is L-smooth, i = 1,...,n g= 1 n n i=1 f i is µ-strongly convex (with potentially µ = 0) constant step size γ t = 1/(16L) initialization with one pass of averaged SGD Non-strongly convex case (Le Roux et al., 2013) E [ g(θ t ) g(θ ) ] 48 σ2 +L θ 0 θ 2 n n t Improvement over regular batch and stochastic gradient Adaptivity to potentially hidden strong convexity

112 Convergence analysis - Proof sketch Main step: find good Lyapunov function J(θ t,y t 1,...,y t n) such that E [ J(θ t,y t 1,...,y t n) F t 1 ] < J(θt 1,y t 1 1,...,y t 1 n ) no natural candidates Computer-aided proof Parameterize function J(θ t,y t 1,...,y t n) = g(θ t ) g(θ )+quadratic Solve semidefinite program to obtain candidates (that depend on n,µ,l) Check validity with symbolic computations

113 Rate of convergence comparison Assume that L = 100, µ =.01, and n = Full gradient method has rate ( 1 µ L) = Accelerated gradient method has rate ( 1 µ ) L = Running n iterations of SAG for the same cost has rate ( 1 1 n 8n) = Fastest possible first-order method has rate ( ) 2 L µ L+ µ = Beating two lower bounds (with additional assumptions) (1) stochastic gradient and (2) full gradient

114 Stochastic average gradient Implementation details and extensions The algorithm can use sparsity in the features to reduce the storage and iteration cost Grouping functions together can further reduce the memory requirement We have obtained good performance when L is not known with a heuristic line-search Algorithm allows non-uniform sampling Possibility of making proximal, coordinate-wise, and Newton-like variants

115 spam dataset (n = , d = )

116 Summary and future work Constant-step-size averaged stochastic gradient descent Reaches convergence rate O(1/n) in all regimes Improves on the O(1/ n) lower-bound of non-smooth problems Efficient online Newton step for non-quadratic problems Robustness to step-size selection Going beyond a single pass through the data

117 Summary and future work Constant-step-size averaged stochastic gradient descent Reaches convergence rate O(1/n) in all regimes Improves on the O(1/ n) lower-bound of non-smooth problems Efficient online Newton step for non-quadratic problems Robustness to step-size selection Going beyond a single pass through the data Extensions and future work Pre-conditioning Proximal extensions fo non-differentiable terms kernels and non-parametric estimation line-search parallelization

118 Outline 1. Large-scale machine learning and optimization Traditional statistical analysis Classical methods for convex optimization 2. Non-smooth stochastic approximation Stochastic (sub)gradient and averaging Non-asymptotic results and lower bounds Strongly convex vs. non-strongly convex 3. Smooth stochastic approximation algorithms Asymptotic and non-asymptotic results 4. Beyond decaying step-sizes 5. Finite data sets

119 Conclusions Machine learning and convex optimization Statistics with or without optimization? Significance of mixing algorithms with analysis Benefits of mixing algorithms with analysis Open problems Non-parametric stochastic approximation Going beyond a single pass over the data (testing performance) Characterization of implicit regularization of online methods Further links between convex optimization and online learning/bandits

120 References A. Agarwal, P. L. Bartlett, P. Ravikumar, and M. J. Wainwright. Information-theoretic lower bounds on the oracle complexity of stochastic convex optimization. Information Theory, IEEE Transactions on, 58(5): , R. Aguech, E. Moulines, and P. Priouret. On a perturbation approach for the analysis of stochastic tracking algorithms. SIAM J. Control and Optimization, 39(3): , F. Bach. Adaptivity of averaged stochastic gradient descent to local strong convexity for logistic regression. Technical Report , HAL, F. Bach and E. Moulines. Non-asymptotic analysis of stochastic approximation algorithms for machine learning. In Adv. NIPS, F. Bach, R. Jenatton, J. Mairal, and G. Obozinski. Optimization with sparsity-inducing penalties. Technical Report , HAL, F. Bach, R. Jenatton, J. Mairal, and G. Obozinski. Structured sparsity through convex optimization, A. Beck and M. Teboulle. A fast iterative shrinkage-thresholding algorithm for linear inverse problems. SIAM Journal on Imaging Sciences, 2(1): , Albert Benveniste, Michel Métivier, and Pierre Priouret. Adaptive algorithms and stochastic approximations. Springer Publishing Company, Incorporated, D. P. Bertsekas. A new class of incremental gradient methods for least squares problems. SIAM Journal on Optimization, 7(4): , 1997.

121 D. Blatt, A. O. Hero, and H. Gauchman. A convergent incremental gradient method with a constant step size. SIAM Journal on Optimization, 18(1):29 51, L. Bottou and O. Bousquet. The tradeoffs of large scale learning. In Adv. NIPS, L. Bottou and Y. Le Cun. On-line learning for very large data sets. Applied Stochastic Models in Business and Industry, 21(2): , S. Boucheron and P. Massart. A high-dimensional wilks phenomenon. Probability theory and related fields, 150(3-4): , S. Boucheron, O. Bousquet, G. Lugosi, et al. Theory of classification: A survey of some recent advances. ESAIM Probability and statistics, 9: , N. Cesa-Bianchi, A. Conconi, and C. Gentile. On the generalization ability of on-line learning algorithms. Information Theory, IEEE Transactions on, 50(9): , B. Delyon and A. Juditsky. Accelerated stochastic approximation. SIAM Journal on Optimization, 3: , J. Duchi and Y. Singer. Efficient online and batch learning using forward backward splitting. Journal of Machine Learning Research, 10: , ISSN M. P. Friedlander and M. Schmidt. Hybrid deterministic-stochastic methods for data fitting. arxiv: , S. Ghadimi and G. Lan. Optimal stochastic approximation algorithms for strongly convex stochastic composite optimization. Optimization Online, July, Saeed Ghadimi and Guanghui Lan. Optimal stochastic approximation algorithms for strongly convex stochastic composite optimization, ii: shrinking procedures and optimal algorithms. SIAM Journal