A Spike Neural Network Design Approach to Reduce Parameters for Evolving Signal Classifiers

Full Text PDF PDF
Author(s) Arnab Roy | J. David Schaffer | Craig B. Laramee
Pages 212-228
Volume 4
Issue 4
Date April, 2014
Keywords design approach; spike neural network; signal classifiers; evolution; pattern detectors; machine learning


Spiking neural networks have powerful appeal when it comes to spatio-temporal pattern detection tasks, due to their implicit nature of accepting as well as processing temporally encoded information. However, achieving the dream has proven rather challenging, with at least one factor being that the number of tunable parameters grows significantly with the network size. In this paper, we offer design approaches for two network types, a sequence detector (multi-channels) and a temporal pattern detector (single channel), that detect predefined inter-spike interval (ISI) patterns in spike trains. The network itself does not learn; it has no need to, the network is correct by design. This permits evolution to concentrate on the task of learning the design specification for signal classification tasks, rather than having to adjust myriad network parameters. Here, we present only the design approach, the mathematical basis for it, and pseudo code that implements it.

< Back to April Issue