부산대학교 도서관

There are some studies about applications of independent component analysis (ICA) to feature extractions in pattern recognition. Since the ICA is an unsupervised learning, independent components are not always useful features for recognition. Therefore, we propose an ICA using a category information. The proposed ICA is realized by three-layered neural networks of which learning algorithm is combination of the error back-propagated algorithm and the fast ICA algorithm. Simulations are performed for three databases from the UCI database to evaluate the effectiveness of the proposed algorithm. We obtain higher recognition accuracy than the original ICA.