Self-synchronizing signal parsing with spiking feature-detection filters

Following an earlier suggestion, the concept of a hierarchical network of feature-detection filters is developed. The individual filters are derived from a localized least-squares approach based on non-generative state space models, which results in simple forward-only recursions for the actual computations. It is demonstrated that such filters can naturally cope with spiking signals, and the use of spiking signals in such networks is advocated. The feasibility of the approach is demonstrated with a four-layer network that understands Morse code.