A generalized canonical correlation analysis based method for blind source separation from related data sets

Combining PCA and multiset CCA for dimension reduction when group ICA is applied to decompose naturalistic fMRI data

“…Reduced complexity of sources also explains faster convergence of ICA decomposition observed for the extended GICA model. In favour of this argument, Karhunen et al [12] indicated that CCA alone already provides separation of sources at some degree using second-order statistics. Subsequently, in the next stage, partly separated sources form the new mixture of less complex sources in turn is decomposed further by ICA using higher-order statistics.…”

“…Therefore, finding correlated subspace in the data before applying ICA is well justified for our purposes. The extension of blind source separation (BSS) by first finding correlated subspace was proposed in [12] and was reported to be significantly better than ICA alone in separating simulated mixtures. It should be noted, however, that the model we introduce here is not identical with the one applied in [12].…”

“…The main difference is in the ICA approaches employed. We selected group ICA with temporal concatenation strategy, whereas in [12] ICA was applied on each dataset separately. Furthermore, MCCA implementation in the present study is from [13], which in turn follows the original publication [11].…”

“…Furthermore, MCCA implementation in the present study is from [13], which in turn follows the original publication [11]. On the other hand, authors in [12] offered their own implementation of MCCA (or generalized CCA, as referred in the article). Nevertheless, the main principle of finding the correlated subspace in the data prior subjecting to ICA decomposition is the same.…”

“…Nevertheless, the main principle of finding the correlated subspace in the data prior subjecting to ICA decomposition is the same. Interestingly, even though Karhunen et al also included experiments on real fMRI data in [12], only two-set CCA was applied as source separation method rather than as a preprocessing of ICA. In this study, the developed GICA extension was applied on challenging fMRI data acquired during task-free continuous music listening experiment and the results were compared with outcome of conventional GICA.…”