Non-linear discriminant feature extraction using generalized back-propagation network
Jian‐Hui Jiang, Jihong Wang, Xia Chu, Ru‐Qin Yu · Journal of Chemometrics · 1996
This paper extends linear feature extraction techniques to a wide variety of non-linear cases using a modified neural network method. This neural network method was developed by replacing the mean-square criterion with a discriminant criterion J = trace (Sm−1·Sb). A new learning algorithm, the generalized back-propagation (GBP) algorithm, has been proposed to maximize this criterion at the network outputs. Though working in a supervised manner, the proposed learning algorithm requires no training outputs for network learning. Experimental results show that the proposed neural network method works much better than linear discriminant analysis (LDA) in linearly inseparable cases. Compared with the conventional neural network method, i.e. the standard back-propagation network (SBPN), the learning efficiency, the classification performance with limited hidden nodes and the generalization ability of the proposed network method are also more favourable. The proposed neural network method provides a promising alternative to SBPN in non-linear pattern recognition applications.