Extended Neural Networks for Signal Detection and Classification: An Approach for Simultaneous Optimization of Parameterized Preprocessing and Neural Networks
Astrid Doering, Herbert Witte · Studies in health technology and informatics · 1996
In this contribution, a generic framework for the simultaneous adaptation of a neural classifier and a parameter–controlled preprocessing scheme is presented. Training algorithms are introduced for preprocessing methods with both differentiable and non–differentiable (with respect to the controlling parameters) transfer functions. As an example, an Extended Neural Network that consists of a bank of quadrature filters (preprocessing stage) and a Multi-Layer-Perceptron (classification stage) was used for the segmentation of discontinuous neonatal EEG. Since the spectral characterisation of the segment classes (particularly of the burst onset, [7]) is interindividually different, the “tuning” of the standard filters' mean frequencies with the proposed algorithm improves the segmentation performance significantly.