Figure 3:

(a) Model errors in the learned model unmixing matrices versus simulated ground truth, (b) signal-to-interference ratios (SIR) of the decomposed model source activities, and (c) symmetric KL divergence of the learned source probability densities for AMICA decompositions using 1–6 models each averaged across 100 simulations. Red dashed lines indicate the performance of one-model Infomax ICA applied to each of the known data segments (whereas AMICA has to learn the segmentation). Significant differences in unpaired t-tests are shown (* p < 0.01, ** p < 1 × 10⁻⁴, *** p < 1 × 10⁻⁶). Red comparisons between AMICA and ICA models; blue asterisks denote comparisons between AMICA model orders. Overall, model errors were lowest for veridical (3-model) and slightly over-complete (4-model) AMICA decompositions.