Stochastic gradient descent and robustness to ill-conditioning

Transcription

1 Stochastic gradient descent and robustness to ill-conditioning Francis Bach INRIA - Ecole Normale Supérieure, Paris, France ÉCOLE NORMALE SUPÉRIEURE Joint work with Aymeric Dieuleveut, Nicolas Flammarion, Eric Moulines - Venezia, May 2016

2 Big data revolution? A new scientific context Data everywhere: size does not (always) matter Science and industry Size and variety Learning from examples n observations in dimension p

3 Search engines - Advertising

4 Visual object recognition

5 Bioinformatics Protein: Crucial elements of cell life Massive data: 2 millions for humans Complex data

6 Context Machine learning for big data Large-scale machine learning: large p, large n p : dimension of each observation (input) n : number of observations Examples: computer vision, bioinformatics, advertising Ideal running-time complexity: O(pn) Going back to simple methods Stochastic gradient methods (Robbins and Monro, 1951) Mixing statistics and optimization Using smoothness to go beyond stochastic gradient descent

7 Context Machine learning for big data Large-scale machine learning: large p, large n p : dimension of each observation (input) n : number of observations Examples: computer vision, bioinformatics, advertising Ideal running-time complexity: O(pn) 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 p, large n p : dimension of each observation (input) n : number of observations Examples: computer vision, bioinformatics, advertising Ideal running-time complexity: O(pn) 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 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 p Explicit features adapted to inputs (can be learned as well) Using Hilbert spaces for non-linear / non-parametric estimation

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 p (regularized) empirical risk minimization: find ˆθ solution of 1 n min l ( y i, θ,φ(x i ) ) + µω(θ) θ R p 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(y θ,φ(x) ) True 0-1 loss: l(y θ,φ(x) ) = 1 y θ,φ(x) <0 Usual convex losses: hinge square logistic

13 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 p (regularized) empirical risk minimization: find ˆθ solution of 1 n min l ( y i, θ,φ(x i ) ) + µω(θ) θ R p n i=1 convex data fitting term + regularizer

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 p (regularized) empirical risk minimization: find ˆθ solution of 1 n min l ( y i, θ,φ(x i ) ) + µω(θ) θ R p 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

15 Smoothness and strong convexity A function g : R p R is L-smooth if and only if it is twice differentiable and θ R p, eigenvalues [ g (θ) ] L smooth non smooth

16 Smoothness and strong convexity A function g : R p R is L-smooth if and only if it is twice differentiable and θ R p, eigenvalues [ g (θ) ] L Machine learning with g(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 ) Bounded data: Φ(x) R L = O(R 2 )

17 Smoothness and strong convexity A twice differentiable function g : R p R is µ-strongly convex if and only if θ R p, eigenvalues [ g (θ) ] µ convex strongly convex

18 Smoothness and strong convexity A twice differentiable function g : R p R is µ-strongly convex if and only if θ R p, eigenvalues [ g (θ) ] µ (large µ/l) (small µ/l)

19 Smoothness and strong convexity A twice differentiable function g : R p R is µ-strongly convex if and only if θ R p, eigenvalues [ g (θ) ] µ Machine learning with g(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)

20 Smoothness and strong convexity A twice differentiable function g : R p R is µ-strongly convex if and only if θ R p, eigenvalues [ g (θ) ] µ Machine learning with g(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

21 Iterative methods for minimizing smooth functions Assumption: g convex and smooth on R p Gradient descent: θ t = θ t 1 γ t g (θ t 1 ) O(1/t) convergence rate for convex functions O(e (µ/l)t ) convergence rate for strongly convex functions

22 Iterative methods for minimizing smooth functions Assumption: g convex and smooth on R p Gradient descent: θ t = θ t 1 γ t g (θ t 1 ) O(1/t) convergence rate for convex functions O(e (µ/l)t ) convergence rate for strongly convex functions Newton method: θ t = θ t 1 g (θ t 1 ) 1 g (θ t 1 ) O ( e ρ2t ) convergence rate

23 Iterative methods for minimizing smooth functions Assumption: g convex and smooth on R p Gradient descent: θ t = θ t 1 γ t g (θ t 1 ) O(1/t) convergence rate for convex functions O(e (µ/l)t ) convergence rate for strongly convex functions Newton method: θ t = θ t 1 g (θ t 1 ) 1 g (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

24 Stochastic approximation Goal: Minimizing a function f defined on R p given only unbiased estimates f n(θ n ) of its gradients f (θ n ) at certain points θ n R p

25 Stochastic approximation Goal: Minimizing a function f defined on R p given only unbiased estimates f n(θ n ) of its gradients f (θ n ) at certain points θ n R p Machine learning - statistics f(θ) = Ef n (θ) = El(y n, θ,φ(x n ) ) = generalization error Loss for a single pair of observations: f n (θ) = l(y n, θ,φ(x n ) ) Expected gradient: f (θ) = Ef n(θ) = E { l (y n, θ,φ(x n ) )Φ(x n ) } Beyond convex optimization: see, e.g., Benveniste et al. (2012)

26 Convex stochastic approximation Key assumption: smoothness and/or strong convexity 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

27 Convex stochastic approximation Key assumption: smoothness and/or strong convexity 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 Running-time = O(np) Single pass through the data One line of code among many

28 Convex stochastic approximation Existing analysis 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

29 Convex stochastic approximation Existing analysis 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?

30 Convex stochastic approximation Existing analysis 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 for smooth problems with global convergence rate O(1/n) in all situations?

31 Least-mean-square algorithm Least-squares: f(θ) = 1 2 E[ (y n Φ(x n ),θ ) 2] with θ R p 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

32 Least-mean-square algorithm Least-squares: f(θ) = 1 2 E[ (y n Φ(x n ),θ ) 2] with θ R p 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 p n + 4R2 θ 0 θ 2 n Matches statistical lower bound (Tsybakov, 2003) Non-asymptotic robust version of Györfi and Walk (1996)

33 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 θ γ θ 0

34 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 θ θ 0

35 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 θ n θ θ 0

36 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 Ergodic theorem: Averaged iterates converge to θ γ = θ at rate O(1/n)

37 Simulations - synthetic examples Gaussian distributions - p = 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)

38 Simulations - benchmarks alpha (p = 500, n = ), news (p = , 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)

39 Isn t least-squares regression a regression?

40 Isn t least-squares regression a regression? Least-squares regression Simpler to analyze and understand Explicit relationship to bias/variance trade-offs (next slides) Many important loss functions are not quadratic Beyond least-squares with online Newton steps Complexity of O(p) per iteration with rate O(p/n) See Bach and Moulines (2013) for details

41 Optimal bounds for least-squares? Least-squares: cannot beat σ 2 p/n (Tsybakov, 2003). Really?

42 Optimal bounds for least-squares? Least-squares: cannot beat σ 2 p/n (Tsybakov, 2003). Really? Refined analysis (Défossez and Bach, 2015) Ef( θ n ) f(θ ) σ2 p n + H 1/2 (θ 0 θ ) 2 2 γ 2 n 2 In practice: bias may be larger than variance, σ 2 p/n pessimistic

43 Optimal bounds for least-squares? Least-squares: cannot beat σ 2 p/n (Tsybakov, 2003). Really? Refined analysis (Défossez and Bach, 2015) Ef( θ n ) f(θ ) σ2 p n + H 1/2 (θ 0 θ ) 2 2 γ 2 n 2 In practice: bias may be larger than variance, σ 2 p/n pessimistic Refined assumptions with adaptivity (Dieuleveut and Bach, 2014) Ef( θ n ) f(θ ) σ2 γ 1/α trh 1/α + H1/2 r (θ 0 θ ) 2 2 n 1 1/α γ 2r n 2min{r,1} SGD is adaptive to the covariance matrix eigenvalue decay Leads to optimal rates for non-parametric regression

44 Achieving optimal bias and variance terms Current results with averaged SGD (ill-conditioned problems) Variance (starting from optimal θ ) = σ2 p n { R 2 θ 0 θ 2 Bias(nonoise)= min, R4 θ 0 θ,h 1 (θ 0 θ ) } n n 2

45 Achieving optimal bias and variance terms Current results with averaged SGD (ill-conditioned problems) Variance (starting from optimal θ ) = σ2 p n { R 2 θ 0 θ 2 Bias(nonoise)= min, R4 θ 0 θ,h 1 (θ 0 θ ) } n n 2

46 Achieving optimal bias and variance terms Current results with averaged SGD (ill-conditioned problems) Variance (starting from optimal θ ) = σ2 p n { R 2 θ 0 θ 2 Bias(nonoise)= min, R4 θ 0 θ,h 1 (θ 0 θ ) } n n 2 Averaged gradient descent (Bach and Moulines, 2013) Bias R 2 θ 0 θ 2 n Variance σ 2 p n

47 Achieving optimal bias and variance terms Averaged gradient descent (Bach and Moulines, 2013) Bias R 2 θ 0 θ 2 n Variance σ 2 p n

48 Achieving optimal bias and variance terms Averaged gradient descent (Bach and Moulines, 2013) Accelerated gradient descent (Nesterov, 1983) Bias R 2 θ 0 θ 2 n R 2 θ 0 θ 2 n 2 Variance σ 2 p n σ 2 p Acceleration is notoriously non-robust to noise (d Aspremont, 2008; Schmidt et al., 2011) For non-structured noise, see Lan (2012)

49 Achieving optimal bias and variance terms Averaged gradient descent (Bach and Moulines, 2013) Accelerated gradient descent (Nesterov, 1983) Between averaging and acceleration (Flammarion and Bach, 2015) Bias R 2 θ 0 θ 2 n R 2 θ 0 θ 2 n 2 R 2 θ 0 θ 2 n 1+α Variance σ 2 p n σ 2 p σ 2 p n 1 α

50 Achieving optimal bias and variance terms Averaged gradient descent (Bach and Moulines, 2013) Accelerated gradient descent (Nesterov, 1983) Between averaging and acceleration (Flammarion and Bach, 2015) Averaging and acceleration (Dieuleveut, Flammarion, and Bach, 2016) Bias R 2 θ 0 θ 2 n R 2 θ 0 θ 2 n 2 R 2 θ 0 θ 2 n 1+α R 2 θ 0 θ 2 n 2 Variance σ 2 p n σ 2 p σ 2 p n 1 α σ 2 p n

51 Conclusions 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 and adaptivity

52 Conclusions 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 and adaptivity Extensions and future work Going beyond a single pass (Le Roux, Schmidt, and Bach, 2012; Defazio, Bach, and Lacoste-Julien, 2014) Proximal extensions fo non-differentiable terms Kernels and nonparametric estimation (Dieuleveut and Bach, 2014) Parallelization Non-convex problems

53 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): , F. Bach and E. Moulines. Non-strongly-convex smooth stochastic approximation with convergence rate o(1/n). Technical Report , HAL, Albert Benveniste, Michel Métivier, and Pierre Priouret. Adaptive algorithms and stochastic approximations. Springer Publishing Company, Incorporated, L. Bottou and O. Bousquet. The tradeoffs of large scale learning. In Adv. NIPS, A. d Aspremont. Smooth optimization with approximate gradient. SIAM J. Optim., 19(3): , Aaron Defazio, Francis Bach, and Simon Lacoste-Julien. Saga: A fast incremental gradient method with support for non-strongly convex composite objectives. In Advances in Neural Information Processing Systems, pages , A. Défossez and F. Bach. Constant step size least-mean-square: Bias-variance trade-offs and optimal sampling distributions A. Dieuleveut and F. Bach. Non-parametric Stochastic Approximation with Large Step sizes. Technical report, ArXiv, A. Dieuleveut, N. Flammarion, and F. Bach. Harder, better, faster, stronger convergence rates for least-squares regression. Technical Report , arxiv, 2016.

54 N. Flammarion and F. Bach. From averaging to acceleration, there is only a step-size. arxiv preprint arxiv: , L. Györfi and H. Walk. On the averaged stochastic approximation for linear regression. SIAM Journal on Control and Optimization, 34(1):31 61, G. Lan. An optimal method for stochastic composite optimization. Math. Program., 133(1-2, Ser. A): , N. Le Roux, M. Schmidt, and F. Bach. A stochastic gradient method with an exponential convergence rate for strongly-convex optimization with finite training sets. In Adv. NIPS, O. Macchi. Adaptive processing: The least mean squares approach with applications in transmission. Wiley West Sussex, A. S. Nemirovsky and D. B. Yudin. Problem complexity and method efficiency in optimization. Wiley & Sons, Y. Nesterov. A method for solving a convex programming problem with rate of convergence O(1/k 2 ). Soviet Math. Doklady, 269(3): , B. T. Polyak and A. B. Juditsky. Acceleration of stochastic approximation by averaging. SIAM Journal on Control and Optimization, 30(4): , H. Robbins and S. Monro. A stochastic approximation method. Ann. Math. Statistics, 22: , ISSN D. Ruppert. Efficient estimations from a slowly convergent Robbins-Monro process. Technical Report 781, Cornell University Operations Research and Industrial Engineering, M. Schmidt, N. Le Roux, and F. Bach. Convergence rates for inexact proximal-gradient method. In

55 Adv. NIPS, A. B. Tsybakov. Optimal rates of aggregation