Spatio-Chromatic Adaptation via Higher-Order Canonical Correlation Analysis of Natural Images

Independent component and canonical correlation analysis are twogeneral-purpose statistical methods with wide applicability. Inneuroscience, independent component analysis of chromatic naturalimages explains the spatio-chromatic structure of primary corticalreceptive fields in terms of properties of the visual environment.Canonical correlation analysis explains similarly chromatic adaptationto different illuminations. But, as we show in this paper, neither ofthe two methods generalizes well to explain both spatio-chromaticprocessing and adaptation at the same time. We propose a statisticalmethod which combines the desirable properties of independent componentand canonical correlation analysis: It finds independent components ineach data set which, across the two data sets, are related to eachother via linear or higher-order correlations. The new method is aswidely applicable as canonical correlation analysis, and also to morethan two data sets. We call it higher-order canonical correlationanalysis. When applied to chromatic natural images, we found that itprovides a single (unified) statistical framework which accounts forboth spatio-chromatic processing and adaptation. Filters withspatio-chromatic tuning properties as in the primary visual cortexemerged and corresponding-colors psychophysics was reproducedreasonably well. We used the new method to make a theory-driventestable prediction on how the neural response to colored patternsshould change when the illumination changes. We predict shifts in theresponses which are comparable to the shifts reported for chromaticcontrast habituation.