An Introduction to Blind Source Separation Kenny Hild Sept. 19, 2001
Problem Statement Communication System Transmitter Medium Receiver 1.Data sent is unknown 2.Transfer function of medium may be unknown 3.Interference
Possible Solutions Beamforming Uses geometric information Steer antenna array to a desired angle of arrival Filtering Separate based on frequency information Blind Source Separation, BSS Statistical beamforming Steer antenna array to directions based on statistics
Beamforming Suppose Direction of arrival is 0 0 azimuth s 1 (n) = s 2 (n) = cos(wn) Transfer functions are pure delays Then y(n) = x 1 (n) + x 2 (n- ), = 0 y(n) = cos(wn) + cos(wn) + cos(wn + ) + cos(wn + ) y(n) = 2cos(wn) + 2cos(( )/2)cos(wn + ( )/2) s 1 (n) s 2 (n) x 2 (n)x 1 (n)
Filtering Suppose Signal is low pass, noise is white Signals are bandpass Then Design LPF to remove high frequencies Frequency (normalized rad/s) noise signal signal #1 signal #2
Assumptions Signals Overlap in time Angle of arrival is unknown – prevents beamforming Overlap in frequency – prevents filtering Blind Source Separation Does not assume knowledge of DOA Does not require signals to be separable in frequency domain
Applications Early diagnosis of pathology in fetus Each EKG sensor contains a mixture of signals Desire is to separate out fetus’ heartbeat Hearing aids Speech discrimination difficult with multiple speakers The observations are the signals at each ear Cellular communications CDMA signals utilize overlapping frequency ranges Additional signals, multi-path deteriorate performance
Types of Mixtures Memory Instantaneous Convolutive Noise Linearity Over/under-determined
Components of Adaptive Filter Topology Instantaneous Convolutive Criterion Optimization method Gradient descent Fixed point
Topology Over-determined, linear mixture N > M H, W are matrices of ARMA filters Types of topologies Frequency-domain Time-domain Feedforward Feedback Lattice
Topology For Instantaneous Mixtures H, W are matrices of constants Often W is broken down into 2-3 operations Dimension reduction, (N x M) matrix D Spatial whitening, (N x N) matrix W Rotations, (N x N) matrix R W = RWD(N x M) x = Hs(M x 1) y = Wx = WHs(N x 1)
Topology Spatial whitening Makes outputs uncorrelated This is insufficient For separation 4 possible rotations s 1 (n) s 2 (n) y 1 (n)y 2 (n)
Criterion Spatial whitening x = Wx E[xx T ] = I N W = x x J = i j (R x (i,j) – I N (i,j)) 2 Rx=Rx=
Criterion Indeterminacies Gain Permutation Rotations Find characteristic of sources that is not true for any mixture
Criterion Nullify correlations Between nonlinear functions of the outputs Nonlinearity can be most any odd function Cubic Hyperbolic tangent Requires source pdf’s to be even-symmetric Non-linear PCA If data is sphered, stable points are ICA solution Minimizes joint entropy of nonlinear functions of outputs
Criterion Cancellation of HOS 4 th -order (kurtosis) is most common If y 1, y 2, y 3, y 4 can be separated into 2 groups that are mutually independent, 4 th -order cumulant is zero Must check all 4 th -order cumulants Statistical properties of cumulant estimators are poor Central limit theorem Sum of independent, non-Gaussian sources approaches Gaussian Maximize (K-L) distance between marginal pdf and Gaussian Must know/estimate the kurtosis for each source
Criterion Maximum Likelihood Must know/assume source distributions Minimize K-L divergence between output pdf’s and known/assumed source pdf’s Sensitive to outliers, model mismatch Maximize the information flow Maximize joint entropy of outputs (of the nonlinearities) Nonlinearities should be source cdf’s Equivalent to maximum likelihood
Criterion Mutual statistical independence Oftentimes sources are independent Uncorrelatedness does not imply independence Canonical criterion Difficult to estimate Solution includes an infinite-limit integral Marginal pdf’s estimated by truncated expansion about Gaussian