A Kernel on Persistence Diagrams for Machine Learning

Transcription

1 A Kernel on Persistence Diagrams for Machine Learning Jan Reininghaus 1 Stefan Huber 1 Roland Kwitt 2 Ulrich Bauer 1 1 Institute of Science and Technology Austria 2 FB Computer Science Universität Salzburg, Austria Toposys Annual Meeting IST Austria September 9, 2014 Jan Reininghaus, Stefan Huber, Roland Kwitt, Ulrich Bauer: A kernel on persistence diagrams 1 of 25

2 Pipeline in topological machine learning Task: shape retrieval Task: object recognition Task: texture recognition Topological data analysis Jan Reininghaus, Stefan Huber, Roland Kwitt, Ulrich Bauer: A kernel on persistence diagrams Introduction 2 of 25

3 Pipeline in topological machine learning Task: shape retrieval SVM PCA Task: object recognition k-means Task: texture recognition Topological data analysis Machine learning Jan Reininghaus, Stefan Huber, Roland Kwitt, Ulrich Bauer: A kernel on persistence diagrams Introduction 2 of 25

4 Pipeline in topological machine learning Task: shape retrieval SVM Kernel PCA Task: object recognition k-means Task: texture recognition Topological data analysis Machine learning We want a robust, multi-scale kernel on the set of persistence diagrams! Jan Reininghaus, Stefan Huber, Roland Kwitt, Ulrich Bauer: A kernel on persistence diagrams Introduction 2 of 25

5 Background on kernels Definition Given a non-empty set X, a (positive definite) kernel is a symmetric, positive-definite function k : X X R, i.e. k(x, x ) = k(x, x) x, x X (1) [k(x i, x j )] n,n i,j=1,1 is pos. semi-def. x 1,..., x n X, n > 0. (2) Example: An inner product on any vector space V is a kernel. Jan Reininghaus, Stefan Huber, Roland Kwitt, Ulrich Bauer: A kernel on persistence diagrams Kernels 3 of 25

6 Background on kernels Theorem For any kernel k on a set X there exists a Hilbert space H and a mapping ϕ: X H with k(x, x ) = (ϕ(x), ϕ(x )) H x, x X. (3) Jan Reininghaus, Stefan Huber, Roland Kwitt, Ulrich Bauer: A kernel on persistence diagrams Kernels 4 of 25

7 Background on kernels: Examples Given a Vector space V with inner product (, ). k(x, x ) = (x, x ) is a kernel. k(x, x ) = (x, x ) d, with d 0, is a kernel. k(x, x ) = e λ x x 2 is a kernel for all λ > 0. Radial basis function kernel (RBF-kernel). Let k 1,..., k m be kernels on a set X. m k=1 λ ik i, with λ i 0, is a kernel. Jan Reininghaus, Stefan Huber, Roland Kwitt, Ulrich Bauer: A kernel on persistence diagrams Kernels 5 of 25

8 Background on kernels: Optimal assignment kernel Goal: A kernel on the set of finite point sets in R 2. Definition Let X, Y be two finite point sets in R 2. The optimal assignment kernel is given as {max π : X Y x X k(x, π(x)) for X Y k A (X, Y ) = max π : Y X y Y k(π(y), y) otherwise, (4) with k(, ) being a positive definite kernel for points in R 2. The optimal assignment kernel with k(x, y) = e λ x y 2 positive definite, i.e., not a kernel. and λ > 0 is not Jan Reininghaus, Stefan Huber, Roland Kwitt, Ulrich Bauer: A kernel on persistence diagrams Kernels 6 of 25

9 Hilbert-space embedding Goal: A kernel on persistence diagrams, i.e., on finite multi-sets of points in R 2. Jan Reininghaus, Stefan Huber, Roland Kwitt, Ulrich Bauer: A kernel on persistence diagrams The kernel / Hilbert space 7 of 25

10 Hilbert-space embedding Goal: A kernel on persistence diagrams, i.e., on finite multi-sets of points in R 2. Goal: An embedding Ψ: D H from the set D of persistence diagrams into a Hilbert space H. The inner product of H will serve as kernel: with F, G D. k(f, G) = (Ψ(F ), Ψ(G)) H Or we use the RBF-kernel on H, or a linear-combination of kernels, or... Jan Reininghaus, Stefan Huber, Roland Kwitt, Ulrich Bauer: A kernel on persistence diagrams The kernel / Hilbert space 7 of 25

11 Embedding a multi-set of points Given a persistence diagram D: Replace each point p D with a Dirac delta δ p centered at p: D H 2 (R 2 ): D p D δ p However, the induced metric in H 2 (R 2 ) is not stable. We apply heat diffusion with p D δ p as initial condition and Dirichlet boundary conditions at the diagonal. Solutions are in L2(R 2 ). Claim: Solutions are stable. Jan Reininghaus, Stefan Huber, Roland Kwitt, Ulrich Bauer: A kernel on persistence diagrams The kernel / Hilbert space 8 of 25

12 Heat diffusion: Definition Let Ω = {(x, y) R 2 : y x}. Heat-diffusion PDE For a given diagram D we consider the solution u : Ω R 0 R of xx u + yy u = t u in Ω R >0, (5) u = 0 on Ω R 0, (6) u = δ p p D on Ω {0}. (7) At scale σ > 0, the embedding of diagrams is given as: Ψ σ (D) is in L 2 (Ω), that is, Ψ σ is an L 2 -embedding. Ψ σ (D) = u(,, σ) (8) Jan Reininghaus, Stefan Huber, Roland Kwitt, Ulrich Bauer: A kernel on persistence diagrams The kernel / Heat diffusion 9 of 25

13 Heat diffusion: The solution Heat-diffusion PDE For a given diagram D we consider the solution u : Ω R 0 R of xx u + yy u = t u in Ω R >0, u = 0 on Ω R 0, u = δ p on Ω {0}. p D Jan Reininghaus, Stefan Huber, Roland Kwitt, Ulrich Bauer: A kernel on persistence diagrams The kernel / Heat diffusion 10 of 25

14 Heat diffusion: The solution Heat-diffusion PDE For a given diagram D we consider the solution u : R 2 R 0 R of xx u + yy u = t u in R 2 R >0, (9) u = p D δ p δ p on R 2 {0}. (10) Extend Ω to R 2 : Mirror and negate each δ p at diagonal. For p = (a, b), we set p = (b, a). Restricting its solution to Ω solves the original problem. Solution is given by convolution with Gaussian kernel. u(x, y, t) = 1 4πt p D e (x,y) p 2 4t e (x,y) p 2 4t (11) Jan Reininghaus, Stefan Huber, Roland Kwitt, Ulrich Bauer: A kernel on persistence diagrams The kernel / Heat diffusion 10 of 25

15 Heat diffusion: The solution The L 2 -embedding at scale σ results in: Ψ σ (D): Ω R: (x, y) 1 4πσ p D e (x,y) p 2 4σ e (x,y) p 2 4σ (12) Jan Reininghaus, Stefan Huber, Roland Kwitt, Ulrich Bauer: A kernel on persistence diagrams The kernel / Heat diffusion 11 of 25

16 Heat diffusion: The solution The L 2 -embedding at scale σ results in: Ψ σ (D): Ω R: (x, y) 1 4πσ p D e (x,y) p 2 4σ e (x,y) p 2 4σ (12) Jan Reininghaus, Stefan Huber, Roland Kwitt, Ulrich Bauer: A kernel on persistence diagrams The kernel / Heat diffusion 11 of 25

17 Heat diffusion: The solution The L 2 -embedding at scale σ results in: Ψ σ (D): Ω R: (x, y) 1 4πσ p D e (x,y) p 2 4σ e (x,y) p 2 4σ (12) Jan Reininghaus, Stefan Huber, Roland Kwitt, Ulrich Bauer: A kernel on persistence diagrams The kernel / Heat diffusion 11 of 25

18 Heat diffusion: The solution The L 2 -embedding at scale σ results in: Ψ σ (D): Ω R: (x, y) 1 4πσ p D e (x,y) p 2 4σ e (x,y) p 2 4σ (12) Jan Reininghaus, Stefan Huber, Roland Kwitt, Ulrich Bauer: A kernel on persistence diagrams The kernel / Heat diffusion 11 of 25

19 Heat diffusion: The solution The L 2 -embedding at scale σ results in: Ψ σ (D): Ω R: (x, y) 1 4πσ p D e (x,y) p 2 4σ e (x,y) p 2 4σ (12) Jan Reininghaus, Stefan Huber, Roland Kwitt, Ulrich Bauer: A kernel on persistence diagrams The kernel / Heat diffusion 11 of 25

20 The inner product Given two diagrams F and G. We compute the inner product explicitly: (Ψ σ (F ), Ψ σ (G)) = Ψ σ (F )Ψ σ (G) Ω = 1 8πσ p F q G e p q 2 8σ e p q 2 8σ (13) O( F G ) time, no approximation. Jan Reininghaus, Stefan Huber, Roland Kwitt, Ulrich Bauer: A kernel on persistence diagrams The kernel / Heat diffusion 12 of 25

21 Robustness Goal: We want to bound Ψ σ (F ) Ψ σ (G) L2 by W q (F, G). Observation A linear embedding Ψ: D V, with a normed vector space V, cannot be stable w.r.t W q for q {2,..., }. By linear we mean Ψ(F + G) = Ψ(F ) + Ψ(G) and Ψ( ) = 0. Set G = and F n = {u,..., u}. }{{} n times We have W q (F n, G) = min η : F n G q (u,v) η But Ψ(F n ) Ψ(G) V = n Ψ(F 1 ) V. u v q = q n W q (F 1, G). Jan Reininghaus, Stefan Huber, Roland Kwitt, Ulrich Bauer: A kernel on persistence diagrams The kernel / Robustness 13 of 25

22 Robustness Theorem For any two persistence diagrams F and G, and any σ > 0 we have Ψ σ (F ) Ψ σ (G) L2(Ω) 8 σ π W 1(F, G). (14) Multi-scale and noise: Noise on the input data causes W 1 (F, G) to increase. We can counteract by increasing σ. Jan Reininghaus, Stefan Huber, Roland Kwitt, Ulrich Bauer: A kernel on persistence diagrams The kernel / Robustness 14 of 25

23 Robustness: Proof I Proof. Let η = arg min η : F G (u,v) η u v. Augmenting D with diagonal points leaves Ψσ(D) unchanged. We denote by N u (x) = 1 4πσ x u 2 e 4σ. Ψ σ (F ) Ψ σ (G) L2(Ω) = (N u N u ) (N v N v ) (u,v) η 2 N u N v L2(Ω) (u,v) η L2(Ω) Jan Reininghaus, Stefan Huber, Roland Kwitt, Ulrich Bauer: A kernel on persistence diagrams The kernel / Robustness 15 of 25

24 Robustness: Proof II Using N u N v L2(Ω) = 1 4πσ 1 e u v 2 2 8σ Ψ σ (F ) Ψ σ (G) L2(Ω) 1 πσ we have (u,v) η 1 e u v 2 2 8σ. By the convexity of x e x, we have e λx 1 λx, and hence Ψ σ (F ) Ψ σ (G) L2(Ω) 1 πσ 1 σ 8π (u,v) η 1 σ 8π W 1(F, G) 1 8σ u v 2 2 (u,v) η u v Jan Reininghaus, Stefan Huber, Roland Kwitt, Ulrich Bauer: A kernel on persistence diagrams The kernel / Robustness 16 of 25

25 Persistence landscapes: Definition Introduced by Bubenik [Bubenik, 2013]. death birth Jan Reininghaus, Stefan Huber, Roland Kwitt, Ulrich Bauer: A kernel on persistence diagrams Persistence landscapes 17 of 25

26 Persistence landscapes: Definition Introduced by Bubenik [Bubenik, 2013]. death birth Jan Reininghaus, Stefan Huber, Roland Kwitt, Ulrich Bauer: A kernel on persistence diagrams Persistence landscapes 17 of 25

27 Persistence landscapes: Definition Introduced by Bubenik [Bubenik, 2013]. λ k (x) death (a, b) x a b birth Jan Reininghaus, Stefan Huber, Roland Kwitt, Ulrich Bauer: A kernel on persistence diagrams Persistence landscapes 17 of 25

28 Persistence landscapes: Definition Introduced by Bubenik [Bubenik, 2013]. death (a, b) λ k (x) a b x Jan Reininghaus, Stefan Huber, Roland Kwitt, Ulrich Bauer: A kernel on persistence diagrams Persistence landscapes 17 of 25

29 Persistence landscapes: Definition Introduced by Bubenik [Bubenik, 2013]. death (a, b) λ k (x) λ 1 a b x Jan Reininghaus, Stefan Huber, Roland Kwitt, Ulrich Bauer: A kernel on persistence diagrams Persistence landscapes 17 of 25

30 Persistence landscapes: Definition Introduced by Bubenik [Bubenik, 2013]. death (a, b) λ k (x) λ 2 λ 1 a b x Jan Reininghaus, Stefan Huber, Roland Kwitt, Ulrich Bauer: A kernel on persistence diagrams Persistence landscapes 17 of 25

31 Persistence landscapes: Definition Introduced by Bubenik [Bubenik, 2013]. death (a, b) λ k (x) λ 2 λ 1 a λ 3 b x Jan Reininghaus, Stefan Huber, Roland Kwitt, Ulrich Bauer: A kernel on persistence diagrams Persistence landscapes 17 of 25

32 Persistence landscapes: Definition The persistence landscape of a persistence diagram can be interpreted as a function in L p (R 2 ). Let Ψ L : D L p (R 2 ): D λ(, ) denote this embedding. λ(k, x) 0 x 1 2 k... Jan Reininghaus, Stefan Huber, Roland Kwitt, Ulrich Bauer: A kernel on persistence diagrams Persistence landscapes 18 of 25

33 Persistence landscapes: Stability Lemma (Bottleneck stability) For two persistence diagrams F and G it holds that Ψ L (F ) Ψ L (G) L (R 2 ) W (F, G) (15) Lemma (Weighted Wasserstein stability) For two persistence diagrams F and G it holds that Ψ L (F ) Ψ L (G) Lq(R 2 ) q min η : F G (u,v) η pers(u) u v q + 2 u v q+1 (16) q + 1 Ψ L : D L q (R 2 ) is not a linear embedding. Jan Reininghaus, Stefan Huber, Roland Kwitt, Ulrich Bauer: A kernel on persistence diagrams Persistence landscapes 19 of 25

34 Stability comparison: Experiment 1 Let G = and let F t = {( t, t)}, with t R 0. Ψ L (F t ) Ψ L (G) Lq t 1+ 1 q W q (F t, G) t 1 4πσ Ψ σ (F t ) Ψ σ (G) L2 1 e t2 8σ t Jan Reininghaus, Stefan Huber, Roland Kwitt, Ulrich Bauer: A kernel on persistence diagrams Persistence landscapes 20 of 25

35 Stability comparison: Experiment 2 Let F t = {( t, t)} and G t = {( t + 1, t + 1)}, with t R 0. Ψ L (F t ) Ψ L (G t ) Lq t 1 q 1 1 4πσ 1 e 2 8σ W q (F t, G t ) O(1) Ψ σ (F t ) Ψ σ (G t ) L2 1 t Jan Reininghaus, Stefan Huber, Roland Kwitt, Ulrich Bauer: A kernel on persistence diagrams Persistence landscapes 21 of 25

36 Stability comparison: Experiment 3 Persistence landscapes emphasize high-persistence features. Noise at high-persistence points can hide significant mid-range features. It becomes hard to distinguish between Class 1 and Class 2. Class 1 Class 2 t Jan Reininghaus, Stefan Huber, Roland Kwitt, Ulrich Bauer: A kernel on persistence diagrams Persistence landscapes 22 of 25

37 Experiments: PSS- vs. Landscape-kernel SHREC 2014 dataset of shapes: 300 meshes, 20 classes (poses). Morse-filtration of heat-signature scalar field on each mesh. An additional scale parameter ( frequency ). 2 Tasks: Retrieval: Get the nearest neighbor in Hilbert space. Classification: Train a SVM with the kernel. Jan Reininghaus, Stefan Huber, Roland Kwitt, Ulrich Bauer: A kernel on persistence diagrams Applications 23 of 25

38 Experiments: Retrieval 80 Frequency=6, Dimension= NN PSS Landscape e e e e e e e e-01 Time Jan Reininghaus, Stefan Huber, Roland Kwitt, Ulrich Bauer: A kernel on persistence diagrams Applications 24 of 25

39 Experiments: Classification Freq. PSS [%] Landscapes [%] Diff ± ± ± ± ± ± ± ± ± ± ± ± ± ± ± ± ± ± ± ± ± ± ± ± ± ± ± ± ± ± ± ± ± ± ± ± ± ± ± ± Table : Comparison of the PSS- vs. Landscape-kernel using 5-fold cross-validation on the SHREC2014 (Synthetic) dataset for dimension 1. Jan Reininghaus, Stefan Huber, Roland Kwitt, Ulrich Bauer: A kernel on persistence diagrams Applications 25 of 25

40 Bibliography I Bubenik, P. (2013). Statistical topological data analysis using persistence landscapes. arxiv, available at Jan Reininghaus, Stefan Huber, Roland Kwitt, Ulrich Bauer: A kernel on persistence diagrams Applications 26 of 25