Overview of the methods used to determine the matrices and in Figure 3.

The variables and denote the canonical coordinates (feature outputs) of the representations. Higher-order canonical correlation analysis (HOCCA) generalizes canonical correlation analysis (CCA) in terms of the detected dependencies between the canonical coordinates. Moreover, it makes the canonical coordinates sparse which results in an efficient representation of the data. Independent component analysis (ICA) is maximizing the representation efficiency of the individual data sets without taking possible correspondences into account. Whitening by principal component analysis (PCA) is the first processing step in all methods. CCA and HOCCA yield coupled representations. For ICA and whitening, the correspondence between the filter outputs must be determined as part of a postprocessing step. We used mutual information maximization for the matching, see Materials and Methods for details.