Real and Complex Independent Subspace Analysis by Generalized Variance

Transcription

1 Real and Complex Independent Subspace Analysis by Generalized Variance Neural Information Processing Group, Department of Information Systems, Eötvös Loránd University, Budapest, Hungary ICA Research Network 2006

2 K-Independent Subspace Analysis (K-ISA) Cocktail party problem: groups of people / music bands ISA nicknames: MICA, group ICA, IVA K {R,C }

3 K-ISA Equations Observation z is mixture of independent components: z(t) = As(t), s(t) = [s 1 (t);...;s M (t)], where s m (t) K dm are i.i.d. sampled random variables in time, s i is independent of s j, if i j, mixing matrix A R D D is invertible, with D := dim(s). Goal: ŝ. Specially for d m = 1: R-ICA, C-ICA.

4 Ambiguities of the K-ISA Model Hidden components can (only) be determined up to permutation, and invertible linear transformation within subspaces. Whitening assumption: E[s] = 0, cov [s] = I D, E[z] = 0, cov [z] = I D. }{{} Lessened ambiguities: invertible orthogonal (R) / unitary(c).

5 f-uncorrelatedness Whitening second order uncorrelatedness Our approach: independence is approximated as uncorrelatedness for a lot of functions (for f F) Formally, 1 Estimation of the hidden source in feedforward architecture: y(t) = Wz(t). 2 f-covariance matrix is estimated empirically, after applying a function f( F): C(f, T) = ĉov[f(y), f(y)], 3 Cost function for W to make the blocks zero in C(f, T) out of the block-diagonal, for all f F.

6 Cost for ISA Based on Joint f-decorrelation Cost function for ISA: J(F, T, W) := f F M C(f, T, W) 2 min W:orthogonal, where M picks out the f-covariance of different subspaces. J should be minimized to solve the ISA problem. Too difficult: optimization on the Stiefel / flag manifold [Nishimori et al., 2006]! Reduce the task further.

7 The K-ISA Separation Theorem Essence: K-ISA = K-ICA + permutation search (with K-ISA cost). Formally, Theorem Let H denote Shannon s differential entropy. Let us suppose that the u := s m components in K-ISA satisfy ( d ) H w i u i i=1 d w i 2 H(u i ), w K = 1 i=1 then W ISA = PW ICA with a P R D D permutation matrix. References: [Cardoso, 1998], [Szabó et al., 2006a].

8 Simulations Test databases (can be scaled in d) d-geom (M = 4): d-spherical (M = 3): Performance measure: normalized Amari-error r [0, 1]. Measures block permutation property of WA.

9 Results-1: d-geom Amari error (r) 10 0 d geom 10 1 d=10 (D=40) d=20 (D=80) d=30 (D=120) d=40 (D=160) d=50 (D=200) d=60 (D=240) d=70 (D=280) d=80 (D=320) Number of samples (T) x10 3

10 Results-2: d-spherical Amari error (r) 10 0 d spherical 10 1 d=20 (D=60) d=30 (D=90) d=40 (D=120) d=50 (D=150) d=60 (D=180) d=70 (D=210) d=80 (D=240) d=90 (D=270) d=100 (D=300) d=110 (D=330) Number of samples (T) x10 3

11 Connection to Other Techniques (d = 1) Alike to KC (kernel covariance) we use a function set F [Gretton et al., 2003]. For F = {f}: f-decorrelation is equivalent to minimization the cost { } 0 Q W (f, T) := 1 2 log det[c(f, T)] M. m=1 det[cm,m (f, T)] This is right the cost function of KGV (kernel generalized variance) [Bach and Jordan, 2002]. With RNN architecture (instead of feedforward), K = R gives rise to self-organization ( gradient) [Meyer-Bäse et al., 2006].

12 References-1 Nishimori, Y., Akaho, S., and Plumbley, M. D. (2006). Riemannian optimization method on the flag manifold for independent subspace analysis. In Independent Component Analysis and Blind Signal Separation (ICA 2006), volume 3889 of LNCS, pages Springer. Bach, F. R. and Jordan, M. I. (2002). Kernel independent component analysis. JMLR, 3:1 48. Gretton, A., Herbrich, R., and Smola, A. (2003). The kernel mutual information. In IEEE ICASSP, volume 4, pages Meyer-Bäse, A., Gruber, P., Theis, F., and Foo, S. (2006). Blind source separation based on self-organizing neural network. Engineering Applications of AI, 19:

13 References-2 Cardoso, J. (1998). Multidimensional independent component analysis. In Proceedings of ICASSP 98, Seattle, WA. Szabó, Z., Póczos, B., and Lőrincz, A. (2006a). Cross-entropy optimization for independent process analysis. In Independent Component Analysis and Blind Signal Separation (ICA 2006), LNCS 3889, pages Springer. Szabó, Z., Póczos, B., and Lőrincz, A. (2006b). Separation theorem for K-independent subspace analysis with sufficient conditions. Technical report, Eötvös Loránd University, Budapest.

14 Summary Presented K-ISA method: joint decorrelation on function set F: with feedforward architecture: Reduction using the K-ISA Separation Theorem K-ISA = K-ICA + permutation search (with K-ISA cost) KGV for F = {f}. with recurrent architecture: self-organization. First step toward large scale problems: few hundred dimensions, power law decrease of estimation error.

15 Thank you for the paying attention!