Facial expression recognition by ICA with selective prior

Permutation ambiguity of the classical ICA may cause problems in feature extraction for pattern classification. To solve that, we include a selective prior for de-mixing coefficients into the classical ICA. Since the prior is constructed upon the classification information from the training data, we refer to the proposed ICA model with a selective prior as a supervised ICA. We formulate the learning rule for the supervised ICA by taking a form of the natural gradient approach, and then investigate the performance of the supervised ICA in facial expression recognition from the aspects of both the correct rate of recognition and the robustness to the number of independent components.