Independent component analysis (ICA) is a powerful technique for the multivariate, non-inferential, data-driven analysis of functional magnetic resonance imaging (fMRI) data-sets. The non-inferential nature of ICA makes this a suitable technique for the study of complex mental states whose temporal evolution would be difficult to describe analytically in terms of classical statistical regressors. Taking advantage of this feature, ICA can extract a number of functional connectivity patterns regardless of the task executed by the subject. The technique is so powerful that functional connectivity patterns can be derived even when the subject is just resting in the scanner, opening the opportunity for functional investigation of the human mind at its basal "default" state, which has been proposed to be altered in several brain disorders. However, one major drawback of ICA consists in the difficulty of managing its results, which are not represented by a single functional image as in inferential studies. This produces the need for a classification of ICA results and exacerbates the difficulty of obtaining group "averaged" functional connectivity patterns, while preserving the interpretation of individual differences. Addressing the subject-level variability in the very same framework of "grouping" appears to be a favourable approach towards the clinical evaluation and application of ICA-based methodologies. Here we present a novel strategy for group-level ICA analyses, namely the self-organizing group-level ICA (sog-ICA), which is used on visual activation fMRI data from a block-design experiment repeated on six subjects. We propose the sog-ICA as a multi-subject analysis tool for grouping ICA data while assessing the similarity and variability of the fMRI results of individual subject decompositions.