Mean mutual information reduction and ‘dipolarity’ measures for each algorithm across 13 data sets.

1EEGLAB (sccn.ucsd.edu/eeglab); ICAcentral (tsi.enst.fr/icacentral); ICALAB (bsp.brain.riken.go.jp/ICALAB); ICA DTU Toolbox (mole.imm.dtu.dk.toolbox/ica). Numbers in parentheses to the right of the ICAcentral source label indicate the entry in the ICACentral.org database; other numbers in this position give the toolbox version used.
2A symmetric approach to optimizing the FastICA weights returned similar results.
3By default not using pre-whitening.
4The time lag used was 100 samples, which is supposed to be optimal for EEG data.
The leftmost column gives the algorithm used (and when ambiguous, the MATLAB function in parentheses). The second column (Mutual Information Reduction, MIR, in kilobits per second) indicates the excess mutual information remaining among the component time courses, compared to the component time courses of the most efficient algorithm tested (AMICA). The third column (near-dipolar percentage, ND10%) indicates the percentage of returned components whose scalp maps had less than 10% residual variance from the scalp projection of the best-fitting single equivalent dipole model. The fourth column (Origin) indicates the online source repository from which the MATLAB source code was obtained (see footnotes).