Supervised Independent Component Analysis with Class Information

Independent Component Analysis (ICA) is a method to transform from mixed signals into independent components. ICA has been so far applied to blind signal separation problems such as sound, speech, images, and biological signals. Recently, ICA is applied to fea- ture extraction for face, speech, and image recognitions. Since ICA is an unsupervised learning, extracted independent components are not al- ways useful for recognition purposes. In this paper, we propose a new supervised learning approach to ICA using class information to enhance the separability of features. The proposed method is implemented by a three-layered feedforward network in which target signals are given to the output units. The defined objective function is composed of the following two terms: one is for evaluating independency of hidden out- puts and the other is for evaluating errors between output signals and their targets. Simulations are performed for some datasets in the UCI repository to evaluate the effectiveness of the proposed method. In the proposed method, we obtain higher recognition accuracies as compared with a conventional unsupervised ICA algorithm.