Formal Aspects of Streaming Recurrent Neural Networks

摘要

Streaming recurrent neural networks with linear and spiral space-time structures are analyzed. Formal aspects of the construction and operation of such networks are considered. Four types of synapses are identified in such networks. One of them is the track synapses, which ensure the advancement of signals over the network. Three other types are synapses of dynamic memory. Additionally to traditional synapse weights the attenuation functions of diverging and converging signals are taken into account. The results of simulation of signal processing by streaming recurrent networks with different structures of their layers are presented.