Trace regulation techniques for feature extraction

The trace neural network (TNN) and the sparse trace neural network (STNN) have been explored as good spatial-temporal invariance extractors. However, it is recognized that the overlapping of traces for rapidly varying input sample sequences will result in poor performance of the network. Here we propose trace regulation (TR) techniques to adaptively adjust the distances between traces and to adaptively cluster patterns in the volume-increased representation space. Preliminary simulation results indicate the advantages of the TRs.