Blind Source Separation by Sparse Decomposition in a Signal Dictionary
read more
Citations
Performance measurement in blind audio source separation
From Sparse Solutions of Systems of Equations to Sparse Modeling of Signals and Images
Online Learning for Matrix Factorization and Sparse Coding
Online Learning for Matrix Factorization and Sparse Coding
Handbook of Blind Source Separation: Independent Component Analysis and Applications
References
Maximum likelihood from incomplete data via the EM algorithm
A wavelet tour of signal processing
Learning the parts of objects by non-negative matrix factorization
Atomic Decomposition by Basis Pursuit
Learning parts of objects by non-negative matrix factorization
Related Papers (5)
Frequently Asked Questions (15)
Q2. What have the authors stated for future works in "Blind source separation by sparse decomposition in a signal dictionary" ?
It would be interesting to compare these possibilities to the other methods presented in this article. In this case, the matrices A and W will have linear filters as an elements, and multiplication by an element corresponds to convolution.
Q3. What is the main difficulty in a maximization problem like equation 3.9?
The main difficulty in a maximization problem like equation 3.9 is the bilinear term AC8, which destroys the convexity of the objective function andmakes convergence unstable when optimization starts far from the solution.
Q4. What is the easiest way to perform sparse decomposition of sources?
The easiest way to perform sparse decomposition of such sources is to compute a spectrogram, the coefficients of a time-windowed discrete Fourier transform.
Q5. What is the definition of a blind source separation problem?
The blind source separation problem is to extract the underlying source signals from a set of linear mixtures, where the mixing matrix is unknown.
Q6. How can the authors estimate the mixing matrix A?
Considering the mixing matrix A as a parameter, the authors can estimate it by maximizing the probability of the observed signal X:max A[ P(X|A) = ∫ P(X|A,C)P(C) dC ]
Q7. What is the method for separating a signal?
The authors suggest a twostage separation process: a priori selection of a possibly overcomplete signal dictionary (for instance, a wavelet frame or a learned dictionary) in which the sources are assumed to be sparsely representable, followed by unmixing the sources by exploiting the their sparse representability.
Q8. What is the way to ensure the nonsingularity of W?
Another possibility for ensuring the nonsingularity of W is to subtract K log |det W| from the objectivemin W,C −K log |det W| + 1 2 ‖C8−WX‖2F + µ ∑ j,k βjh(Cjk) (4.3)which (Bell & Sejnowski, 1995; Pearlmutter & Para, 1996) can be viewed as a maximum likelihood term.
Q9. What is the first approach to limiting the norm of the rows?
The first approach is to force each row Ai of the mixing matrix A to be bounded in norm,‖Ai‖ ≤ 1 i = 1, . . . ,N. (3.10)The second way is to restrict the norm of the rows
Q10. What is the prior pdf of C?
we haveP(X|A,C) ∝ ∏ i,t exp− (Xit − (AC8)it) 2 2σ 2 . (3.6)By the independence of the coefficients Cjk and equation 3.1, the prior pdf of C isP(C) ∝ ∏ j,k exp(−βjh(Cjk)).
Q11. What is the a priori probability of the noise in equation 1.3?
The authors also suppose a priori that the mixing matrix A is uniformly distributed over the range of interest and that the noise ξ(t) in equation 1.3 is a spatially and temporally uncorrelated gaussian process2 with zero mean and variance σ 2.3.1 Maximum A Posteriori Approach.
Q12. What is the definition of a linear combination of a small number of dictionary elements?
By sparsity, the authors mean the ability of the signal to be approximated by a linear combination of a small number of dictionary elements ϕk, as s ≈ cT8.
Q13. What is the simplest way to get a reliable estimate of the source signals?
When the noise is small and the matrix A is far from singular, WX gives a reasonable estimate of the source signals S. Taking into account equation 1.4, the authors obtain a least-squares term ‖C8−WX‖2F, so the separation objective may be written asmin W,C 1 2 ‖C8−WX‖2F + µ ∑ j,k βjh(Cjk).
Q14. What is the way to separate a signal from a mixture?
The authors consider the general case of more sources than mixtures, but also derive a more efficient algorithm in the case of a nonovercomplete dictionary and an equal numbers of sources and mixtures.
Q15. What is the reason why PBM used a clustering algorithm?
For this reason, reliable convergence was achieved only when the search started randomly within 10% to 20% distance to the actual solution (in order to get such an initial guess, one can use a clustering algorithm, as in Pajunen et al., 1996, or Zibulevsky et al., in press).