CICAAR: Convolutive ICA with an auto-regressive inverse model

Pronunciation: si-'gär

Copyright and Disclaimer

Why CICAAR?

• The parameters of a convolutive mixing model are estimated. Hence, individual sources can be forwarded into the sensor domain.
• Optional model selection on the number of lags in the mixing model using Bayes Information Criterion (BIC, see 6]). In [1] this was used to show that Convolutive ICA was a better model than instantaneous ICA for the EEG data under consideration.
• The CICAAR is a clean generalization of Infomax ICA to convolutive mixtures. For L=0 (the number of lags), the CICAAR reduces to exactly Infomax ICA with estimation of the inverse mixing matrix.

Why not?

• It is computationally heavy. Successful application so far goes up to ~10 dimensions, and ~50 lags. (See the Tips and Tricks section below for application to EEG and Audio).

Download and installation

1. Download the CICAAR binary compiled for your system:

   Platform	file
   Linux on Intel 32bit	cicaar_linux_IA32.tar
   Mac OS X on Intel 64bit	cicaar_osx_EM64T.tar

   and extract it somewhere on your harddrive.
2. Then download this archive of Matlab functions: cicaar_tools.tar (updated October 16, 2007). Make sure the Matlab path is set correspondingly.
3. NOTE!: You have to edit the file CICAAR.M (which is in the cicaar_tools.tar file) to set up the path for the CICAAR binary and temp files!

Tips and Tricks

EEG

• Downsample your data to, say, ~50Hz sampling rate.
• It can be useful to project your data. Actually, the way we propose to analyze EEG, is to first extract instantaneous ICA components (using e.g. Infomax), and then analyze a subset of "interesting" component activations using the CICAAR. [1,2] The scaling is perfect too then.

Audio

• In a non-reverberant environment, what really matters for succesful separation is the transfer function between the microphones. These nice results were computed using only 50 filter lags (L=50): audio demo [3]

General notes

• Remember to include BIC terms for removed PCA/ICA dimensions (if you plan to compare across different subspaces) [6][7]

Cite the toolbox accurately

@Article{pmid17348768,
   Author="Dyrholm, Mads and Makeig, Scott and Hansen, Lars Kai",
   Title="{{M}odel selection for convolutive {I}{C}{A} with an application to spatiotemporal analysis of {E}{E}{G}}",
   Journal="Neural Comput",
   Year="2007",
   Volume="19",
   Number="4",
   Pages="934--955",
   Month="Apr"
 }

References

    [1] Dyrholm, M., Makeig, S., Hansen, L. K., "Model selection for 
        convolutive ICA with an application to spatio-temporal Analysis
        of EEG", Neural Computation, 19(4):934-955, 2007
 
    [2] Dyrholm, M., Makeig, S., Hansen, L. K., "Model structure selection
        in convolutive mixtures", Independent Component Analysis and Blind
        Signal Separation, Springer LNCS vol. 3889, pp. 74-81, 2006
 
    [3] Dyrholm, M., Hansen, L. K., "CICAAR: Convolutive ICA with 
        an Auto-Regressive Inverse Model", Independent Component Analysis
        and Blind Signal Separation, vol. 3195, pp. 594-601, 2004

    [4] Attias, H., Schreiner, C.E., "Blind Source Separation and Deconvolution:
        the Dynamic Component Analysis Algorithm", Neural Computation,
        10(6):1373-1424,1998

    [5] Torkkola, K., "Blind Separation of Convolved Sources Based on
        Information Maximization", In proceedings of the Workshop on
        Neural Networks for Signal Processing, Kyoto, Japan, 1996

    [6] Schwarz, G., 1978. "Estimating the dimension of a model".
        Annals of Statistics 6(2):461-464.

    [7] Hansen, L.K., Larsen, J., Kolenda, T., "Blind Detection of
        Independent Dynamic Components", in Proc. IEEE ICASSP'2001,
        Salt Lake City, SAM-P8.10, vol. 5, 2001.