Combination of adaptive signal processing and neural classification using an extended backpropagation algorithm

Beside the use of purely neural systems, the combination of preprocessing units and neural classifiers has been used for a variety of signal segmentation and classification tasks. Whereas this approach reduces the input dimensionality as well as the complexity of the classification problem, its performance crucially depends on a proper preprocessing scheme, i.e., feature extraction. In this contribution, adaptive preprocessing units (frequency-selective quadrature filters) are proposed that can be adjusted in order to provide optimal features. The mean frequencies of the filters are tuned to minimize the classification error. Both FIR- and IIR-based filters are introduced and compared with respect to their convergence properties and the classification results. Results for the solution of an EEG segmentation task using the combined system are given