Multimodal independent component analysis—A method of feature extraction from multiple information sources
Shotaro Akaho, Shinji Umeyama · Electronics and Communications in Japan (Part III Fundamental Electronic Science) · 2001
Abstract We propose a method to extract features from a pair of multivariate information sources. CCA (canonical correlation analysis) is a traditional method for this purpose, but it does not always succeed in producing features that are nonlinearly related, because of a Gaussian assumption. We extend the framework of CCA by introducing a criterion used in ICA (independent component analysis). The proposed framework MICA (multimodal ICA) maximizes mutual information between a pair of extracted features from the two modalities, with the features extracted from each modality being statistically independent. The cost function is given by a weighted sum of the two criteria, and it is approximated by Gram–Charlier expansion. The gradient descent learning algorithm is derived to optimize this approximated function by taking a natural gradient. © 2001 Scripta Technica, Electron Comm Jpn Pt 3, 84(11): 21–28, 2001