Stochastic Quasi-Newton Methods

Transcription

1 Stochastic Quasi-Newton Methods Donald Goldfarb Department of IEOR Columbia University UCLA Distinguished Lecture Series May 17-19, / 35

2 Outline Stochastic Approximation Stochastic Gradient Descent Variance Reduction Techniques Newton-like and quasi-newton methods for convex stochastic optimization problems using limited memory block BFGS updates. Numerical results on problems from machine learning. Quasi-Newton methods for nonconvex stochastic optimization problems using damped and modified limited memory BFGS updates. 2 / 35

3 Stochastic optimization Stochastic optimization min f (x) = E[f (x, ξ)], ξ is random variable Or finite sum (with f i (x) f (x, ξ i ) for i = 1,..., n and very large n) min f (x) = 1 n f i (x) n f and f are very expensive to evaluate; stochastic gradient descent (SGD) methods choose a random subset S [n] and evaluate f S (x) = 1 S i S i=1 f i (x) and f S (x) = 1 S f i (x) Essentially, only noisy info about f, f and 2 f is available Challenge: how to smooth variability of stochastic methods Challenge: how to design methods that take advantage of noisy 2nd-order information? i S 3 / 35

4 Stochastic optimization Deterministic gradient method Stochastic gradient method 4 / 35

5 Stochastic Variance Reduced Gradients Stochastic methods converge slowly near the optimum due to the variance of the gradient estimates f S (x); hence requiring a decreasing step size. We use the control variates approach of Johnson and Zhang (2013) for a SGD method SVRG. It uses d = f S (x t ) f S (w k ) + f (w k ), where w k is a reference point, in place of f S (x t ). w k, and the full gradient, are computed after each full pass of the data, hence doubling the work of computing stochastic gradients. 5 / 35

6 Stochastic Average Gradient At iteration t - Sample i from {1,..., N} - update y t+1 i = f i (x t ) and y t+1 j - Compute g t+1 = 1 N N j=1 y t+1 i - Set x t+1 = x t α t+1 g t+1 = y t j for all j i Provable linear convergence in expectation. Other SGD variance reduction techniques have been recently proposes including the methods: SAGA, SDCA, S2GD. 6 / 35

7 Quasi-Newton Method for min f (x) : f C 1 Gradient method: Newton s method: Quasi-Newton method: x k+1 = x k α k f (x k ) x k+1 = x k α k [ 2 f (x k )] 1 f (x k ) x k+1 = x k α k B 1 k f (x k) where B k 0 approximates the Hessian matrix Update B k+1 s k = y k, (Secant equation) where s k = x k+1 x k = α k d k, and y k = f k+1 f k 7 / 35

8 BFGS BFGS quasi-newton method B k+1 = B k + y k y k s k y k B ks k s k B k s k B ks k where s k := x k+1 x k and y k := f (x k+1 ) f (x k ) B k+1 0 if B k 0 and s k y k > 0 (curvature condition) Secant equation has a solution if s k y k > 0 When f is strongly convex, s k y k > 0 holds automatically If f is nonconvex, use line search to guarantee s k y k > 0 H k+1 = (I s kyk sk y )H k (I y ksk k sk y ) + s ksk k sk y k 8 / 35

9 Prior work on Quasi-Newton Methods for Stochastic Optimization P1 N.N. Schraudolph, J. Yu and S.Günter. A stochastic quasi-newton method for online convex optim. Int l. Conf. AI & Stat., 2007 Modifies BFGS and L-BFGS updates by reducing the step s k and the last term in the update of H k, uses step size α k = β/k for small β > 0. P2 A. Bordes, L. Bottou and P. Gallinari. SGD-QN: Careful quasi-newton stochastic gradient descent. JMLR vol. 10, 2009 Uses a diagonal matrix approximation to [ 2 f ( )] 1 which is updated (hence, the name SGD-QN) on each iteration, α k = 1/(k + α). 9 / 35

10 Prior work on Quasi-Newton Methods for Stochastic Optimization P3 A. Mokhtari and A. Ribeiro. RES: Regularized stochastic BFGS algorithm. IEEE Trans. Signal Process., no. 10, Replaces y k by y k δs k for some δ > 0 in BFGS update and also adds δi to the update; uses α k = β/k; converges in expectation at sub-linear rate E(f (x k ) f ) C/k P4 A. Mokhtari and A. Ribeiro. Global convergence of online limited memory BFGS. to appear in J. Mach. Learn. Res., Uses L-BFGS without regularization and α k = β/k; converges in expectation at sub-linear rate E(f (x k ) f ) C/k 10 / 35

11 Prior work on Quasi-Newton Methods for Stochastic Optimization P5 R.H. Byrd, S.L. Hansen, J. Nocedal, and Y. Singer. A stochastic quasi-newton method for large-scale optim. arxiv v2, 2015 Averages iterates over L steps keeping H k fixed; uses average iterates to update H k using subsampled Hessian to compute y k ; α k = β/k; converges in expectation at a sub-linear rate E(f (x k ) f ) C/k P6 P. Moritz, R. Nishihara, M.I. Jordan. A linearly-convergent stochastic L-BFGS Algorithm, 2015 arxiv: v1 Combines [P5] with SVRG; uses fixed step size α; converges in expectation at a linear rate. 11 / 35

12 Using Stochastic 2nd-order information Assumption: f (x) = 1 n n i=1 f i(x) is strongly convex and twice continuously differentiable. Choose (compute) a sketching matrix S k (the columns of S k are a set of directions). We do not use differences in noisy gradients to estimate curvature, but rather compute the action of the sub-sampled Hessian on S k. i.e., compute Y k = 1 T i T 2 f i (x)s k, where T [n]. 12 / 35

13 Example of Hessian-Vector Computation In binary classification problem, sample function (logistic loss) where f (w; x i, z i ) = z i log(c(w; x i )) + (1 z i ) log(1 c(w; x i )) c(w; x i ) = exp( x i w), x i R n, w R n, z i {0, 1}, Gradient: f (w; x i, z i ) = (c(w; x i ) z i )x i Action of Hessian on s : 2 f (w; x i, z i )s = c(w; x i )(1 c(w; x i ))(x i s)x i 13 / 35

14 block BFGS The block BFGS method computes a least change update to the current approximation H k to the inverse Hessian matrix 2 f (x) at the current point x, by solving min H H k s.t., H = H, HY k = S k. where A = ( 2 f (x k )) 1/2 A( 2 f (x k )) 1/2 F (F = Frobenius) This gives the updating formula (analgous to the updates derived by Broyden, Fletcher, Goldfarb and Shanno, 1970). H k+1 = (I S k [S k Y k] 1 Y k )H k(i Y k [S k Y k] 1 S k )+S k[s k Y k] 1 S k or, by the Sherman-Morrison-Woodbury formula: B k+1 = B k B k S k [S k B ks k ] 1 S k B k + Y k [S k Y k] 1 Y k 14 / 35

15 Limited Memory Block BFGS After M block BFGS steps starting from H k+1 M, one can express H k+1 as H k+1 = V k H k V T k + S kλ k S T k. = V k V k 1 H k 1 V T k 1 V k + V k S k 1 Λ k 1 S T k 1 V T k + S kλ k S T k k+1 M = V k:k+1 M H k+1 M Vk:k+1 M T + i=k V k:i+1 S i Λ i S T i V T k:i+1, where V k = (I S k Λ k Y T k ) (1) and Λ k = (S T k Y k) 1 and V k:i = V k V i. 15 / 35

16 Limited Memory Block BFGS Intuition Hence, when the number of variables d is large, instead of storing the d d matrix H k, we store the previous M block curvature triples (S k+1 M, Y k+1 M, Λ k+1 M ),..., (S k, Y k, Λ k ). Then, analogously to the standard L-BFGS method, for any vector v R d, H k v can be computed efficiently using a two-loop block recursion (in O(Mp(d + p) + p 3 ) operations), if all S i R d p. Limited memory - least change aspect of BFGS is important Each block update acts like a sketching procedure. 16 / 35

17 Two Loop Recursion 17 / 35

18 Choices for the Sketching Matrix S k We employ one of the following strategies Gaussian: S k N (0, I ) has Gaussian entries sampled i.i.d at each iteration. Previous search directions s i delayed: Store the previous L search directions S k = [s k+1 L,..., s k ] then update H k only once every L iterations. Self-conditioning: Sample the columns of the Cholesky factors L k of H k (i.e., L k L T k = H k) uniformly at random. Fortunately we can maintain and update L k efficiently with limited memory. The matrix S is a sketching matrix, in the sense that we are sketching the, possibly very large equation 2 f (x)h = I to which the solution is the inverse Hessian. Right multiplying by S compresses/sketches the equation yielding 2 f (x)hs = S. 18 / 35

19 The Basic Algorithm 19 / 35

20 Convergence - Assumptions There exist constants λ, Λ R + such that f is λ strongly convex f (w) f (x) + f (x) T (w x) + λ 2 w x 2 2, (2) f is Λ smooth f (w) f (x) + f (x) T (w x) + Λ 2 w x 2 2, (3) These assumptions imply that λi 2 f S (w) ΛI, for all x R d, S [n], (4) from which we can prove that there exist constants γ, Γ R + such that for all k we have γi H k ΓI. (5) 20 / 35

21 Bounds on Spectrum of H k Lemma Assuming 0 < λ < Λ such that for all x R d and T [n], where λi 2 f T (x) ΛI γi H k ΓI 1 1+MΛ γ, Γ (1 + κ) 2M 1 (1 + λ(2 ) and κ = Λ/λ κ+κ) bounds in MNJ depend on problem dimension 1 (d+m)λ γ and Γ [(d+m)λ]d+m 1 λ d+m (dκ) d+m 21 / 35

22 Linear Convergence Theorem Suppose that the Assumptions hold. Let w be the unique minimizer of f (w). Then in our Algorithm, we have for all k 0 that Ef (w k ) f (w ) ρ k Ef (w 0 ) f (w ), where the convergence rate is given by ρ = 1/2mη + ηγ2 Λ(Λ λ) γλ ηγ 2 Λ 2 < 1, assuming we have chosen η < γλ/(2γ 2 Λ 2 ) and that we choose m large enough to satisfy m 1 2η (γλ ηγ 2 Λ(2Λ λ)), which is a positive lower bound given our restriction on η. 22 / 35

23 Numerical Experiments Empirical Risk Minimization Test Problems logistic loss with l 2 regularizer min w n log(1 + exp( y i a i, w )) + L w 2 2 i=1 given data: A = [a 1, a 2,, a n ] R d n y {0, 1} n. For each method, chose step size η {1,.5,.1,.05,..., , 10 8 } that gave best results Computed full gradient after each full data pass. Vertical axis in figures below: log(relative error) 23 / 35

24 error gisette-scale d = 5, 000, n = 6, 000 gauss_18_m_5 prev_18_m_5 fact_18_m_3 MNJ_bH_330 SVRG time (s) 24 / 35

25 error covtype-libsvm-binary d = 54, n = 581, 012 gauss_8_m_5 prev_8_m_5 fact_8_m_3 MNJ_bH_3815 SVRG time (s) 25 / 35

26 error Higgs d = 28, n = 11, 000, 000 gauss_4_m_5 prev_4_m_5 fact_4_m_3 MNJ_bH_16585 SVRG time (s) 26 / 35

27 error SUSY d = 18, n = 3, 548, 466 gauss_5_m_5 prev_5_m_5 fact_5_m_3 MNJ_bH_9420 SVRG time (s) 27 / 35

28 error epsilon-normalized d = 2, 000, n = 400, 000 gauss_45_m_5 prev_45_m_5 fact_45_m_3 MNJ_bH_3165 SVRG time (s) 28 / 35

29 error rcv1-training d = 47, 236, n = 20, 242 gauss_10_m_5 prev_10_m_5 fact_10_m_3 MNJ_bH_715 SVRG time (s) 29 / 35

30 error url-combined d = 3, 231, 961, n = 2, 396, 130 gauss_2_m_5 prev_2_m_5 fact_2_m_3 MNJ_bH_7740 SVRG time (s) 30 / 35

31 Contributions New metric learning framework. A block BFGS framework for gradually learning the metric of the underlying function using a sketched form of the subsampled Hessian matrix New limited memory block BFGS method. May also be of interest for non-stochastic optimization Several sketching matrix possibilities. More reasonable bounds on eigenvalues of H k more reasonable conditions for step size 31 / 35

32 Nonconvex stochastic optimization Most stochastic quasi-newton optimization methods are for strongly convex problems; this is needed to ensure a curvature condition required for the positive definiteness of B k (H k ) This is not possible for problems min f (x) E[F (x, ξ)], where f is nonconvex In deterministic setting, one can do line search to guarantee the curvature condition, and hence the positive definiteness of B k (H k ) Line search is not possible for stochastic optimization To address these issues we develop a stochastic damped and a stochastic modified L-BFGS method. 32 / 35

33 Stochastic Damped BFGS (Wang, Ma, G, Liu, 2015) Let y k = 1 m m i=1 ( f (x k+1, ξ k,i ) f (x k, ξ k,i )) and define ȳ k = θ k y k + (1 θ k )B k s k, where { 1, if sk θ k = y k 0.25sk B ks k, (0.75sk B ks k )/(sk B ks k sk y k), if sk y k < 0.25sk B ks k. Update H k : (replace y k by ȳ k ) H k+1 = (I ρ k s k ȳ k )H k(i ρ k ȳ k s k ) + ρ ks k s k where ρ k = 1/s k ȳk Implemented in a limited memory version Work in progress: combine with variance reduced stochastic gradients (SVRG) 33 / 35

34 Convergence of Stochastic Damped BFGS Method Assumptions [AS1] f C 1, bounded below, f is L Lipschitz continuous [AS2] For any iteration k, the stochastic gradient satisfies E ξk [ f (x k, ξ k )] = f (x k ) E ξk [ f (x k, ξ k ) f (x k ) 2 ] σ 2 Theorem (Global convergence): Assume AS1-AS2 hold, (and α k = β/k γ/(lγ 2 ) for all k), then there exist positive constants γ, Γ, such that γi H k ΓI, for all k, and lim inf k f (x k) = 0, with probability 1. Under additional assumption E ξk [ f (xk, ξ k ) 2] M lim f (x k) = 0, with probability 1. k We do not need to assume convexity of f 34 / 35

35 Block-L-BFGS Method for Non-Convex Stochastic Optimization Block-update H k+1 = (I S k Λ 1 k Y k )H k(i Y k Λ 1 k S k ) + S kλ 1 where Λ k = Sk Y k = Sk 2 f (x k )S k In non-convex case Λ k = Λ k may not be positive definite. Λ k 0 discovered while computing Cholesky factorization LDL of Λ k. If during Cholesky, d j δ or (LD 1/2 ) ij β are not satisfied, d j is increased by τ j. = (Λ k ) jj (Λ k ) jj + τ j has the effect of moving search direction H k+1 f (x k+1 ) toward one of negative curvature. k S k Modification based on Gershgorin disc also possible. 35 / 35