Properties of processing element of artificial neural network
Answers
(a) Properties of a processing element of a traditional artificial neural network versus properties of a processing element of a biologically realistic Dynamic Synapse neural network. (b) Conceptual representation of speaker-independent word recognition identification by a Dynamic Synapse neural network. Inputs to the network are digitized speech waveforms from different speakers for the same word, which have little similarity (low cross-correlation) because of differences in speaker vocalization. The two networks shown are intended to represent the same network on two different training or testing trials; in a real case, one network is trained with both (or more) speech waveforms. On any given trial, each speech waveform constitutes the input for all five of the input units shown in the first layer. Each unit in the first layer of the network generates a different pulse-train encoding of the speech waveform ("integrate and fire neurons" with different parameter values). The output of each synapse (arrows) to the second layer of the network is governed by four dynamic processes [see (a)], with two of those processes representing second-order nonlinearities; thus, the output to the second layer neurons depends on the time since prior input events. A "dynamic learning rule" modifies the relative contribution of each dynamic process until the output neurons converge on a common temporal pattern in response to different input speech signals (i.e., high cross-correlation between the output patterns).