Optimization. Benjamin Recht University of California, Berkeley Stephen Wright University of Wisconsin-Madison

Transcription

1 Optimization Benjamin Recht University of California, Berkeley Stephen Wright University of Wisconsin-Madison

2 optimization () cost constraints might be too much to cover in 3 hours

3 optimization (for big data?) cost E ξ [(, ξ)] conv(a) constraints distribution over ξ is well-behaved A is simple (low-cardinality, low-dimension, low-complexity) E ξ [(, ξ)] + () closely related cousin where P is a simple convex function

4 Support Vector Machines cancer vs other illness fraud vs normal purchase up-going vs down-going muons = max(, )+λ sample average over observed data and labels regularizer to select low-complexity models

5 LASSO Compressed Sensing Sparse Modeling reduce number of measurements required for signal acquisition search for a sparse set of markers for classification = ( )

6 Matrix Completion M = M ij known for black cells M ij unknown for white cells Rows index features Columns index examples Entries specified on set E How do you fill in the missing data? M = L R * k x n kn entries k x r r x n r(k+n) parameters minimize (,) ( ) + X

7 Graph Cuts Image Segmentation Entity Resolution Topic Modeling Bhusnurmath and Taylor, 2008 (,) [, ] = =

8 optimization () cost constraints might be too much to cover in 3 hours optimization is ubiquitous optimization is modular optimization is declarative

9 [ + ] []+α [] Today: gradient descent conjugate/ accelerated stochastic/ sub projected/ proximal mix and match

10 tomorrow: duality min () = max () min max L(, ) = max min L(, ) find problems that always lower bound the optimal value. puts problem in NP conp information from one problem informs the other some times easier to solve one than the other basis of many proof techniques in data science (and tons of other areas too!)

11 what we ll be skipping... 2nd order/newton/bfgs interior point methods/ellipsoid methods active set methods, manifold identification branch and bound integrating combinatorial thinking derivative-free optimization soup of heuristics (simulated annealing, genetic algorithms,...) modeling

12 optimality conditions () R Search for () = Turns a geometric problem into an algebraic problem: solve for the point where the gradient vanishes. Is necessary for optimality (sufficient for convex, smooth f )

13 [ + ] []+α [] gradient descent Assume there exits an where ( )= D Suppose the map ψ() = α () is contractive on D () () < for some run gradient descent starting at [] D [ + ] = [] ([]) = ([]) ( ) [] ( )= contractivity + [] linear rate

14 If f is 2x differentiable, contractivity means f is convex on D ( + ) () for all t>0 lim + ( + ) () = lim + = () ( ( + ) ()) () () + convex Lipschitz gradients

15 convexity ( +( )) ()+( )() () ()+ () ( ) + l strong convexity = Lipschitz gradients () () () ()+ () ( )+ condition number of Hessian follows from Taylor s theorem With step size α = l+, [] + []. ([]) + []

16 Note on convergence rate With step size α = [] + l+,. [] If you don t know the exact stepsize, can we achieve the rate? Exact line search: at each iteration, find the α that minimizes f(x+αd). Backtracking line search: Reduce α by constant multiple until the function value sufficiently decreases. Both achieve linear rate of convergence. More sophisticated line searches often used in practice, but none improve over this rate in the worst case.

17

18

19 acceleration/multistep gradient method akin to an ODE [ + ] =[] = () ([]) to prevent oscillation, add a second order term = () [ + ] =[] ([]) + ([] [ ]) [ + ] = []+ [] [] = ([]) + [ ] heavy ball method (constant α,β) when f is quadratic, this is Chebyshev s iterative method

20

21

22 analysis [ + ] = []+ [] [] = ([]) + [ ] heavy ball method (constant α,β) Analyze by defining a composite error vector: := [] [ ] Then [ + ] =[]+( [] ) where := ( )+( + )

23 [ + ] =[]+( [] ) B has the same eigenvalues as analysis (cont.) +( + ) = diag(,,..., ) where λi are the eigenvalues of ( ) Choose α,β to explicitly minimize the max eigenvalue of B to obtain = ( + / ) = +. Leads to linear convergence for [] + with rate approximately

24 about those rates... Best steepest descent: Linear rate approx Heavy-ball: Linear rate approx Big difference! To yield [] < [] + + log(/ ) gradient descent log(/ ) heavy ball A factor of κ 1/2 difference. e.g. if κ=100, need 10 times fewer steps.

25 conjugate gradients [ + ] = []+ [] [] = ([]) + [ ] Choose αk by line search (to reduce f) Choose βk such that p[k] is approximately conjugate to p[1],..., p[k-1] (really only makes sense for quadratics, but whatever...) Does not achieve a better rate than heavy ball Gets around having to know parameters Convergence proofs very sketchy (except when f is quadratic) and need elaborate line search to guarantee local convergence.

26 optimal method Nesterov s optimal method (1983,2004) = [ + ] = []+ [] [] = ([]+ ([] [ ])) + [ ] heavy ball with extragradient step + =( +) + + = ( ) + + = + + = = + FISTA (Beck and Teboulle 2007) + Recent fixes use line search to find parameters and still achieve optimal rate (modulo log factors) Analysis based on estimate sequences, using simple quadratic approximations to f

27 why optimal? you can t beat the heavy ball convergence rate using only gradients and function evaluations. () = + = ( + ) + + () ( + ) + start at x[0] = e1. after k steps, x[j] = 0 for j>k+1 norm of the optimal solution on the unseen coordinates tends to ( + )

28 not strongly convex (l=0) gradient descent: ([]) [] + optimal method: ([]) [] ( + ) Big difference! To yield ([]) < gradient descent [] optimal method [] A factor of ϵ 1/2 difference. e.g. if ϵ=0.0001, need 100 times fewer steps.

29 nonconvexity can still efficiently find a point where () in time (/ ) not convex n.b. nonconvexity really `lets you model anything () =,= () = for all Q checking if 0 is a local minimum in NP-hard

30 stochastic gradient E [(, )] Stochastic Gradient Descent: For each k, sample ξk and compute [ + ] =[] ([], ) Robbins and Monro (1950) Adaptive Filtering (1960s-1990s) Back Propagation in Neural Networks (1980s) Online Learning, Stochastic Approximation (2000s)

31 Support Vector Machines cancer vs other illness fraud vs normal purchase up-going vs down-going muons = max(, )+λ = max(, )+ Step 1: Pick i and compute the sign of the assignment: ˆ = sign( ) Step 2: If ˆ =, ( ) +

32 matrix completion M = L R * k x n k x r r x n Entries Specified on set E (,) ( ) + X Idea: approximate X LR T (L,R) (,) (L R ) + L + R

33 SGD code for matrix completion M = L R * k x n k x r r x n (L,R) (,) (L R ) + L + R Step 1: Pick (u,v) and compute residual: =(L R ) Step 2: Take a mixture of current model and corrected model: L ( )L R R ( )R L

34 Mixture of hundreds of models, including nuclear norm nuclear norm (a.k.a. SVD)

35 SGD and BIG Data E [(, )] For each k, sample ξk and compute [ + ] =[] ([], ) Ideal for big data analysis: small, predictable memory footprint robustness against noise in data rapid learning rates one algorithm! Why should this work?

36 Example: Computing the mean ( = ) = Stepsize = 1/2k = ( )= = ( )/ =. = ( )/ = = ( )/ =. In general, if we minimize SGD returns: = ( = ) =

37 Stepsize = 1/2 = cos +sin () = = Choose directions in order 7 8 Choose a direction uniformly with replacement

38 convergence of sgd () Assume f is strongly convex with parameter l and has Lipschitz gradients with parameter L Assume at each iteration we sample G(x), an unbiased estimate of f(x), independent of x and the past iterates Assume G(x) M almost surely. [ + ] =[] ([])

39 [ + ] = [] ([]) = [] ([] ) ([]) + ([]). Define = E [] + E[([] ) ([])] +. By iterating expectation: E[([] ) ([])] = E [ ] E [([] ) ([]) [ ] ] = E[([] ) ([])] By strong convexity: ([]) ([] ) ([]) ( )+. + ( ) +

40 + ( ) + Large steps: >, = E[ [] ] max, [] < Small steps:, constant stepsize E[ [] ] ( ) [] + Achieves 1/k rate if run in epochs of diminishing stepsize

41 R () = = () Algorithm Time per iteration Error after T iterations Error after items N Newton O(d 2 N+d 3 ) C I 2 Gradient O(dN) C G SGD O(d) (or constant) C S N

42 extensions non-smooth, non-strongly convex (1/ k) non-convex (converges asymptotically) stochastic coordinate descent (special decomposition of f) parallelization

43 projected gradient Suppose it is easy to solve Π () unique solution () convex smooth projected gradient method: [ + ] Π ([]+α [])

44 [ + ] Π ([]+α []) Key Lemma: Π () Π () Assume minimizer of f Ω Assume f is strongly convex [ + ] = ([] (([])) ( ) [] ([]) = ([]) ( ) [] non-expansive ( )= contractivity + [] linear rate

45 ()+() ()+() ([]) + ([]) ( []) + α [] + () Define prox () = arg min + () Solving the approximation yields [ + ] =prox α ([] α ([]))

46 proximal mapping prox () = arg min + () () = () = prox () = () prox () = + < >

47 ()+() ()+() ([]) + ([]) ( []) + α [] + () Define prox () = arg min + () Solving the approximation yields [ + ] =prox α ([] α ([])) Key Lemma: prox () prox () immediately implies earlier analysis works for proximal gradient. projected gradient is a special case inherits rates of convergence from f (i.e., P=0)

48 More variants mirror descent: use a general distance () ( )+ ( ), + D(, ) ADMM: combine multiple prox operators for complicated constraints.

49 [ + ] []+α [] gradient descent conjugate/ accelerated stochastic/ sub projected/ proximal mix and match Everything here combines, and you get the expected rates out.