1 / 43

Multidimensional Data Analysis : the Blind Source Separation problem.

Multidimensional Data Analysis : the Blind Source Separation problem. Outline : Blind Source Separation Linear mixture model Principal Component Analysis : Karhunen-Loeve Basis Gaussianity Mainstream Independent Component Analysis Spectral Matching ICA Diversity and separability

booth
Download Presentation

Multidimensional Data Analysis : the Blind Source Separation problem.

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript


  1. Multidimensional Data Analysis :the Blind Source Separation problem. Outline : • Blind Source Separation • Linear mixture model • Principal Component Analysis : Karhunen-Loeve Basis • Gaussianity • Mainstream Independent Component Analysis • Spectral Matching ICA • Diversity and separability • Multiscale representations and sparsity • Applications : MCA, MMCA, ICAMCA, denoising ICA, etc.

  2. Blind Source Separation Examples : • The cocktail party problem

  3. Blind Source Separation Examples : • The cocktail party problem • Foregrounds in Cosmic Microwave Background imaging • Detector noise • Galactic dust • Synchrotron • Free - Free • Point Sources • Thermal SZ • …

  4. Blind Source Separation Examples : • The cocktail party problem • Foregrounds in Cosmic Microwave Background imaging • Fetal heartbeat estimation

  5. Blind Source Separation Examples : • The cocktail party problem • Foregrounds in Cosmic Microwave Background imaging • Fetal heartbeat estimation • Hyperspectral imaging for remote sensing

  6. Blind Source Separation Examples : • The cocktail party problem • Foregrounds in Cosmic Microwave Background imaging • Fetal heartbeat estimation • Hyperspectral imaging for remote sensing Differentes views on a single scene obtained with different instruments. COHERENT DATA PROCESSING IS REQUIRED

  7. A simple mixture model • linear, static and instantaneous mixture • map observed in a given frequency band sums the contributions of various components • the contributions of a component to two different detectors differ only in intensity. • additive noise • all maps are at the same resolution

  8. 23 GHz Synchrotron Free-free CMB 33 GHz 41 GHz 94 GHz A simple mixture model

  9. 23 GHz Synchrotron Free-free CMB 33 GHz 41 GHz 94 GHz A simple mixture model

  10. Foreground removal as a source separation problem. • Processing coherent observations. • All component maps are valuable ! • How much prior knowledge? • Mixing matrix, component and noise spatial covariances known [Tegmark96, Hobson98]. • Subsets of parameters known. • Blind Source Separation assuming independent processes [Baccigalupi00, Maino92, Kuruoglu03, Delabrouille03]. • Huge amounts of high dimensional data.

  11. Simplifying assumption • The noiseless case : X = AS • Can A and S be estimated (up to sign, scale and permutation indeterminacy) from an observed sample of X? Solution is not unique. Prior information is needed.

  12. Karhunen-Loeve Transform a.k.a.Principal Component Analysis • Estimation of A and S assuming uncorrelated sources and orthogonal mixing matrix. • Practically : • Compute the covariance matrix of the data X • Perform an SVD : the estimated eigenvectors are the lines of A • Adaptive approximation. • Dimensionality reduction; • Used for compression and noise reduction.

  13. Application : original image

  14. Application : KL basis estimated from 12 by 12 patches.

  15. Application : keeping 10 percent of basis elementscompression

  16. Application : keeping 50 percent of basis elementscompression

  17. Independent Component Analysis • Still in the noiseless case : X = AS • Goal is to find a matrix B such that the entries of Y = BX are independent. • Existence/Uniqueness : • Most often, such a decomposition does not exist. • Theorem [Darmois, Linnik 1950] : Let S be a random vector with independent entries of which at most when is Gaussian, and C an invertible matrix. If the entries of Y = CS are independent, then C is almost the identity. Goal will be to find a matrix B such that the entries of Y = BX are as independent as possible.

  18. Independence or De-correlation? • Independence : • Mutual information or an approximation is used to assess statistical independence: • De-correlation measures linear independence: E( (x - E(x)) (y - E(y)) ) = 0

  19. ICA for gaussian processes? Two linearly mixed independent Gaussian processes do not separate uniquely into two independent components. Need diversity from elsewhere. Different classes of ICA methods • Algorithms based on non-gaussianity i.e. higher order statistics. Most mainstream ICA techniques: fastICA, Jade, Infomax, etc. • Techniques based on the diversity (non proportionality) of variance (energy) profiles in a given representation such as in time, space, Fourier, wavelet : joint diagonalization of covariance matrices, SMICA, etc.

  20. Example : Jade • First, whitening : remove linear dependencies

  21. Example : Jade • First, whitening : removes linear dependencies, and normalizes the variance of the projected data. • Then, minimize a measure of statistical dependence ie by finding a rotation that minimizes fourth order cross-cumulants.

  22. Example : Jade • First, whitening : removes linear dependencies, and normalizes the variance of the projected data. • Then, minimize a measure of statistical dependence ie by finding a rotation that minimizes fourth order cross-cumulants.

  23. Application to SZ cluster map restoration Input maps

  24. Application to SZ cluster map restoration Simulated map Output map

  25. ICA of a mixture of independent Gaussian, stationary, colored processes (1) • Same mixture model in Fourier Space : X(l) = A S(l) • Parameters : the n*n mixing matrix A and the n • power spectra DSi(l), for l = 1 to T . • Whittle approximation to the likelihood :

  26. ICA of a mixture of independent Gaussian, stationary, colored processes (2) • Assuming the spectra are constant over symmetric subintervals, maximizing the likelihood is the same as minimizing: • D turns out to be the Kulback Leibler divergence between two gaussian distributions. • After minimizing with respect to the source spectra, we are left with a joint diagonality criterion to be minimized wrt B = A-1. There are fast algorithms.

  27. A Gaussian stationary ICA approach to the foreground removal problem? • CMB is modeled as a Gaussian Stationary process over the sky. • Working with covariance matrices achieves massive data reduction, and the spatial power spectra are the main parameters of interest. • Easy extension to the case of noisy mixtures. • Connection with Maximum-Likelihood guarantees some kind of optimality. • Suggests using EM algorithm. • [Snoussi2004, Cardoso2002, Delabrouille2003]

  28. Spectral Matching ICA • Back to the case of noisy mixtures : • Fit the model spectral covariances : • to estimated spectral covariances : • by minimizing : • with respect to :

  29. Algorithm • EM can become very slow, especially as noise parameters become small. • Speed up convergence with a few BFGS steps. • All /part of the parameters can be estimated.

  30. Estimating component maps • Wiener filtering in each frequency band : • In case of high SNR, use pseudo inverse :

  31. A variation on SMICA using wavelets • Motivations : • Some components, and noise are a priori not stationary. • Galactic components appear correlated. • Emission laws may not be constant over the whole sky. • Incomplete sky coverage.

  32. A variation on SMICA using wavelets • Motivations : • Some components, and noise are a priori not stationary. • Galactic components are spatially correlated. • Emission laws may not be constant over the whole sky. • Incomplete sky coverage.

  33. A variation on SMICA using wavelets • Motivations : • Some components, and noise are a priori not stationary. • Galactic components appear correlated. • Emission laws may not be constant over the whole sky. • Incomplete sky coverage.

  34. A variation on SMICA using wavelets • Motivations : • Some components, and noise are a priori not stationary. • Galactic components appear correlated. • Emission laws may not be constant over the whole sky. • Incomplete sky coverage. Use wavelets to preserve space-scale information.

  35. The à trous wavelet transform • shift invariant transform • coefficient maps are the same size as the original map • isotropic • small compact supports of wavelet and scaling functions (B3 spline) • Fast algorithms

  36. wSMICA Model structure is not changed : Objective function : Same minimization using EM algorithm and BFGS.

  37. Experiments Simulated Planck HFI observations at high galactic latitudes : Component maps Mixing matrix Planck HFI nominal noise levels

  38. Experiments Simulated Planck HFI observations at high galactic latitudes :

  39. Experiments Contribution of each component to each mixture as a function of spatial frequency :

  40. Results Jade wSMICA initial maps

  41. Results Errors on the estimated emission laws :

  42. Results Rejection rates : where

  43. Multiscale transforms and the quest for sparsity

More Related