Applying ICA in EEG: Choice of the Window Length and of the Decorrelation Method - Biomedical Engineering Systems and Technologies - page 257

(a) (b)

(c)

(d) (e)
Fig. 7. Results on the random data set: performance index ( PI ) curves vs data length for (a) 8
channels, (b) 16 channels, (c) 24 channels, (d) 32 channels and (e) 48 channels. Initialization with
whitening (solid lines) and sphering (dashed lines)

Note on Time Convergence: Due to the large amount of parameters to be estimated by
AMICA, it might be important to notice that this method is extremely time consuming
compared to JADER and FastICA, and also much slower than Extended InfoMax. To
give an idea: while FastICA is taking less than 1s on 32 channels and 20 s data length
(mean time observed for 50 iterations on random data), JADER requires no more than
3s, Extended InfoMax needs up to 3 minutes, and AMICA requires almost 4 minutes.
On the other hand, AMICA is per se much more flexible than Extended InfoMax which
only try to fit a single generalized Gaussian distribution on each source, explaining the
better performances of AMICA when compared to Extended InfoMax.

5.2 Plausible Data Set

Minimum Length Rule Validation. As it can be observed by comparing the computed
(normalized) Riemannian distance, covariance estimations on this data set (figure 8)