Feedforward neural filter with learning in features space. Preliminary results

An adaptive filtering technique using a multilayer perceptron neural network designed as a transversal filter, based on the information extracted from the second order statistics of the signal, is presented. The statistics are extracted with a principal component analysis (PCA) hierarchical network. The training procedure uses an error signal computed as the difference between the desired and actual largest eigenvalues. Some advantages of the proposed method are illustrated by preliminary experiments on electrocardiographic (ECG) signal filtering.