An Evolving Gradient Resampling Method for Machine Learning. Jorge Nocedal

Transcription

1 An Evolving Gradient Resampling Method for Machine Learning Jorge Nocedal Northwestern University NIPS, Montreal

2 Collaborators Figen Oztoprak Stefan Solntsev Richard Byrd 2

3 Outline 1. How to improve upon the stochastic gradient method for risk minimization 2. Noise reduction methods Dynamic Sampling (batching) Aggregated Gradient methods (SAG, SVRG, etc) 3. Second order methods 4. Propose a noise reduction method that re-uses old gradients and also employs dynamic sampling 3

4 Organization of optimization methods Stochastic Gradient Method noise reduction Batch Gradient Method condition number Stochastic Newton Method 4

5 Second-order methods Stochastic Gradient Method Batch Gradient Method Averaging (Polyak-Ruppert) Momentum Natural gradient, Fischer quasi-newton inexact Newton (Hessian-free) Stochastic Newton 5

6 Noise reducing methods Stochastic Gradient Method Batch Gradient Method Dynamic sampling methods Stochastic Newton Method aggregated gradient methods This Talk: combine both ideas Why? 6

7 Objective Function min w F(w) = E[ f (w;ξ)] ξ = (x, y) random variable with distribution P f ( ;ξ) composition of loss l and prediction h w k+1 = w k α k f (w k ;ξ k ) stochastic gradient method (SG) Sample gradient approximation - batch (or mini-batch) F s (w) = 1 S f (w;ξ i ) i S w k+1 = w k α k F S (w k ) batch (mini) method 7

8 Transient behavior of SG Expected function decrease E[F(w k+1 ) F(w k )] α k F(w k ) 2 2 +α k 2 E f (w k,ξ k ) 2 Initially, gradient decrease dominates; then variance in gradient hinters progress (area of confusion) To ensure convergence α k 0 in SG method to control variance. Steplength selected to achieve fast initial progress, but this will slow progress in later stages Dynamic sampling methods reduce gradient variance by increasing batch. What is the right rate? 8

9 Proposal: Gradient accuracy conditions Geometric noise reduction Consider stochastic gradient method with fixed steplength w k+1 = w k αg(w k,ξ k ) If the variance of stochastic gradient decreases geometrically, the method yields linear convergence Lemma If M > 0, ζ (0,1) s.t. Then E[ g(w k,ξ k ) 2 2 ] F(w k ) 2 2 M ζ k 1 Schmidt et al Pasupathy et al E[F(w k ) F * ] νρ k 1 Extension of classical convergence result for gradient method where error in gradient estimates decreases sufficiently rapidly to preserve linear convergence

10 Proposal: Gradient accuracy conditions Optimal work complexity We can ensure variance condition E[ g(w k,ξ k ) 2 2 ] F(w k ) 2 2 M ζ k 1 by letting S k = a k 1 a > 1 F s (w) = 1 f (w;ξ i ) S i S Moreover, we obtain optimal complexity bounds The total number of stochastic gradient evaluations to achieve E[F(w k ) F * ] ε is O(1/ ε) with favorable constants a [1,1 βcµ 2 ] 1 Pasupathy, Glynn et al 2014 Friedlander, Schmidt 2012 Homem-de-Mello, Shapiro 2012 Byrd, Chin, N., Wu 2013

11 Theorem: Suppose F is strongly convex. Consider w k +1 = w k (1/ L)g k where S k is chosen so that variance condition holds and Then S k γ k for γ >1.! E[F(w k ) F(w * )] Cρ k ρ < 1! the number of gradient samples to achieve ε accuracy is O κ ωd ε λ d = no. of variables κ = condition number, λ = smallest eigenv of Hessian Var l(w k ;i) 1 ω (population)

12 Dynamic sampling (batching) At every iteration, choose a subset S of {1,,n} and apply one step of an optimization algorithm to the function F S (w k ) = 1 f (w;ξ i ), S i S At the start, a small sample size S is chosen If optimization step is likely to reduce F(w), sample size is kept unchanged; new sample S is chosen; next optimization step taken Else, a larger sample size is chosen, a new random sample S is selected, a new iterate computed Many optimization methods can be used. This approach creates the opportunity of employing second order methods 12

13 How to implement this in practice? 1. Predetermine a geometric increase, tuning parameter S k = a k 1 a > 1 2. Use angle (i.e. variance test) Ensure bound is satisfied in expectation g(w k ) F(w k ) θ g k θ < 1 Popular: combination of these two strategies. 13

14 Numerical test Numerical Tests: Newton-CG method with dynamic sampling, Armijo line search w k+1 = w k α k 2 F H k (w k ) 1 g k α k 1 Test Problem From Google VoiceSearch 191,607 training points 129 classes; 235 features 30,315 parameters (variables) Small version of production problem Multi-class logistic regression Initial batch size: 1%; Hessian sample 10% 14

15 Function Dynamic New method Newton-CG L-BFGS (m=2,20) Batch L-BFGS Stochastic gradient descent Classical Newton-CG Dynamic change of sample sizes based on variance estimate Time 15

16 However, not completely satisfactory More investigation is needed. Particularly: Transition between stochastic and batch regimes Coordination between step size and batch size Use of second order information (one stochastic gradient is not too noisy) Can the idea of re-using gradients in a gradient aggregation approach help? 16

17 Transition from stochastic to batch regimes Stochastic process gradient methods 1 [ ] S k m SGD twilight zone α k = 1 α k = 1 / k Gradient aggregation could smooth transition 17

18 Randomized Aggregated Gradient Methods (for empirical risk min) Expected Risk: F (w) = E [ f (w;ξ)] Empirical Risk: F m (w) = 1 f (w;ξ i ) = 1 f i (w) m i=1 m i=1 n n SAG, SAGA, SVRG, etc focus on minimizing empirical risk Iteration: w k+1 = w k α y k y k combination of gradients f i evaluated at previous iterates φ j y k = 1 m [ f (w ) f (φ j j k j k 1 m ) + 1 m i f i (φ k 1 )] i=1 Choose j at random SAG 18

19 Example of Gradient Aggregation Methods y k = 1 m [ f j j (w k ) f j (φ k 1 m ) + 1 m i f i (φ k 1 )] i=1 j SAG m SAG, SAGA, SVRG F m (w) = 1 f (w;ξ i ) = 1 f i (w) m i=1 m i=1 Achieve linear rate of convergence in expectation (after a full initialization pass) 19

20 EGR Method The Evolving Gradient Resampling Method for Expected Risk Minimization 20

21 Proposed algorithm 1. Minimizes expected risk (not training error) 2. Stores previous gradients and updates several (s k ) at each 3. iteration 4. Additional (u k ) gradients are computed at current iterate 5. Total amount of stored gradients increases monotonically 6. Shares properties with dynamic sampling and gradient aggregation methods Goal: analyze an algorithm of this generality (interesting in its own right) Finding right balance between re-using old information and batching can result in efficient method. 21

22 The EGR method y k = 1 t k + u k ( j Sk [ f j (w k ) f j (φ k 1 j )]+ t k i f i (φ k 1 i=1 f j (w k ) ) ) + j U k t k number of gradients in storage at start of iteration k U k indices of new gradients sampled at w k u k = U k S k indices of previously computed gradients that are updated s k = S k s k u k j j Evaluated at w k t k How should s k and u k be controled? Related work: Frostig et al 2014, Babanezhad et al

23 Algorithms included in framework stochastic gradient method: s k = 0, u k = 1 dynamic sampling method: s k = 0, u k = function(k) aggregated gradient: s k = constant, u k = 0 s k u k EGR lin: s k = 0, u k = r EGR quad: s k = rk, u k = r EGR exp: s k = u k a k j j t k 23

24 Assumptions: 1) s k = u k = a k a! + geometric growth 2) F strongly convex, f i Lipschitz continuous gradients 3) tr (var [ f i (w)]) v 2 w Lemma: E[E k [e k ]] E[ w k w * ] M σ k E[E k [e k 1 ]] E[ w k 1 w * ] σ k 1 Lyapunov function e k = 1 t k+1 f i ( φ i k ) f i (w k ) σ k = v 2 / t k+1 t k+1 i=1

25 1 η 1 η (1+ α L) 1+ η 1+ η α L 1 η 1+ η α M = α L 1 αµ α η η : probability that an old gradient is recomputed α : steplength L: Lipschitz constant µ : strong convexity parameter Lemma. For sufficiently small α the spectral radius of M satisfies ρ M < 1

26 Theorem: If α k is chosen small enough E w k w * c k β R-linear convergence SAG special case: t k = m, u k = 0, s k = constant Simple proof of R-linear convergence of SAG but with a larger constant

27 Related Methods Streaming SVRG (Frostig et al 2014) Stop wasting gradients (Babanezhad et al 2015) j SVRG m y k = 1 m [ f (w ) f (φ j j k j k 1 m ) + 1 m i f i (φ k 1 )] i=1 27

28 SAG(A) initialization phase 1. Sample j {1,...,m} at random 2. Compute f j (w k ) 3. If j has been sampled earlier, replace old gradient 4. Else add new f j (w k ) to aggregated gradient (memory grows) j SAG m y k = 1 m [ f (w ) f (φ j j k j k 1 m ) + 1 m i f i (φ k 1 )] i=1 28

29 Numerical Result 1. Comparison of EGR with various growth rates 2. Comparison with SGD 3. Comparison with SAG-init and SAGA-init Goal: analyze an algorithm of this generality (interesting in its own right) Finding right balance between re-using old information and batching can result in efficient method. 29

30 Susy EGR vs SG vs Dynamic Sampling 30

31 Random Comparing with SAG initialization 31

32 Random Larger initial batch for EGR 32

33 Alpha 33

34 EGR with various growth rates (Alpha) 34

35 Random 35

36 The End 36