Pulse-domain signal parsing and neural computation

We propose a new model of pulse-based computation based on inner-product filters with linear-system kernels. Each inner-product filter looks for some pulse pattern in its multichannel-input signal by projecting the input signal into a one-dimensional subspace; an output pulse is generated if this projection exceeds some threshold. A layered network of such filters can be used for self-synchronizing multiscale signal parsing. Such a network can be built with computational units that are biologically plausible neurons. The feasibility of the proposed approach is demonstrated with a network that understands Morse code.