EM algorithm for convolutive independent component analysis (CICA)

• US 20030046038A1
• Filed: 05/14/2001
• Published: 03/06/2003
• Est. Priority Date: 05/14/2001
• Status: Active Grant

• Current Assignee: International Business Machines Corporation
• Original Assignee: International Business Machines Corporation
• Inventors: Deligne, Sabine, Gopinath, Ramesh A.
• Granted Patent: US 6,622,117 B2
• Time in Patent Office: Days
• US Class Current: 702/190
• CPC Class Codes: G06F 18/2134 based on separation criteria, e.g. independent component analysis

Abstract

In connection with blind source separation, proposed herein, inter alia, are: expectation-maximization equations to iteratively estimate unmixing filters and source density parameters in the context of convolutive independent component analysis where the sources are modeled with mixtures of Gaussians; a scheme to estimate the length of unmixing filters; and two alternative initialization schemes.

Claims

1. A method of facilitating blind source separation in convolutive independent component analysis, said method comprising the steps of:

initializing at least one parameter;

obtaining a set comprising at least one observation;

computing the likelihood of the at least one observation and of an auxiliary function;

maximizing the auxiliary function, and

ascertaining whether the likelihood of the at least one observation converges.

2. The method according to claim 1, wherein said ascertaining step comprises ascertaining whether the likelihood of the at least one observation converges towards a local maximum.

3. The method according to claim 1, further comprising the step of repeating said computing, maximizing and ascertain steps if the likelihood of the at least one observation is not ascertained as converging.

4. The method according to claim 3, wherein said repeating step comprises successively repeating said computing, maximizing and ascertaining steps until the likelihood of the at least one observation is ascertained as converging.

5. The method according to claim 1, wherein said maximizing step comprises reestimating at least one parameter.

6. The method according to claim 5, wherein said reestimating step comprises at least one of the following steps:

reestimating priors and variances;

updating at least one unmixing filter; and

reestimating means and variances.

7. The method according to claim 6, wherein said reestimating step comprises all of said steps of:

reestimating priors and variances;

updating at least one unmixing filter; and

reestimating means and variances.

8. The method according to claim 7, wherein said step of reestimating priors and variances comprises updating priors and variances of densities as at least one unmixing filter is fixed.

9. The method according to claim 7, wherein said step of updating at least one unmixing filter comprises updating at least one unmixing filter row by row as density parameters are fixed.

10. The method according to claim 7, wherein said step of reestimating means and variances comprises updating means and variances of densities as at least one unmixing filter is fixed.

11. The method according to claim 1, wherein said initializing step comprises initializing source density parameters and at least one unmixing filter.

12. The method according to claim 1, wherein said initializing step comprises time-decorrelating the at least one observation via computing at least one linear prediction coefficient.

13. The method according to claim 12, wherein said initializing step further comprises spatially decorrelating the at least one observation via diagonalizing a covariance matrix.

14. The method according to claim 1, wherein said initializing step comprises time-decorrelating the at least one observation via computing error prediction data with a reflection coefficient estimation method.

15. The method according to claim 14, wherein said initializing step further comprises spatially decorrelating the at least one observation via diagonalizing a covariance matrix.

16. The method according to claim 1, wherein said initializing step comprises the step of estimating the length of at least one unmixing filter.

17. An apparatus for facilitating blind source separation in convolutive independent component analysis, said apparatus comprising:

an initializing arrangement which initializes at least one parameter;

an input arrangement which obtains a set comprising at least one observation;

a likelihood computing arrangement which computes the likelihood of the at least one observation and of an auxiliary function;

a maximizing arrangement which maximizes the auxiliary function, and

an ascertaining arrangement which ascertains whether the likelihood of the at least one observation converges.

18. The apparatus according to claim 17, wherein said ascertaining arrangement is adapted to ascertain whether the likelihood of the at least one observation converges towards a local maximum.

19. The apparatus according to claim 17, wherein said maximizing arrangement is adapted to reestimate at least one parameter.

20. The apparatus according to claim 19, wherein said maximizing arrangement is adapted to perform at least one of the following steps:

reestimating priors and variances;

updating at least one unmixing filter; and

reestimating means and variances.

21. The apparatus according to claim 20, wherein said maximizing arrangement is adapted to perform all of the following steps:

reestimating priors and variances;

updating at least one unmixing filter; and

reestimating means and variances.

22. The apparatus according to claim 21, wherein said maximizing arrangement is, in reestimating priors and variances, adapted to update priors and variances of densities as at least one unmixing filter is fixed.

23. The apparatus according to claim 21, wherein said maximizing arrangement is, in updating at least one unmixing filter, adapted to update at least one unmixing filter row by row as density parameters are fixed.

24. The apparatus according to claim 21, wherein said maximizing arrangement is, in reestimating means and variances, adapted to update means and variances of densities as at least one unmixing filter is fixed.

25. The apparatus according to claim 17, wherein said initializing arrangement is adapted to initialize source density parameters and at least one unmixing filter.

26. The apparatus according to claim 17, wherein said initializing arrangement is adapted time-decorrelate the at least one observation via computing at least one linear prediction coefficient.

27. The apparatus according to claim 26, wherein said initializing arrangement is adapted to spatially decorrelate the at least one observation via diagonalizing a covariance matrix.

28. The apparatus according to claim 17, wherein said initializing arrangement is adapted to time-decorrelate the at least one observation via computing error prediction data with a reflection coefficient estimation method.

29. The apparatus according to claim 28, wherein said initializing arrangement is adapted to spatially decorrelate the at least one observation via diagonalizing a covariance matrix.

30. The apparatus according to claim 17, wherein said initializing arrangement is adapted to estimate the length of at least one unmixing filter.

31. A program storage device readable by machine, tangibly embodying a program of instructions executable by the machine to perform method steps for facilitating blind source separation in convolutive independent component analysis, said method comprising the steps of:

initializing at least one parameter;

obtaining a set comprising at least one observation;

computing the likelihood of the at least one observation and of an auxiliary function;

maximizing the auxiliary function; and

ascertaining whether the likelihood of the at least one observation converges.