A Preprocessing Layer in Spiking Neural Networks – Structure, Parameters, Performance Criteria

The subject of this article is the section of spiking neural network (SNN) closest to the source of input signals. The purpose of this layer is to provide input signal preprocessing and the extraction of primary informative features, promoting its higher-level analysis by subsequent SNN layers. The 1-layer and 2-layer WTA (winner takes all) architectures of this section are explored. Two independent criteria used to evaluate its necessity and efficiency in the given task are discussed. A variant of the STDP plasticity rule specially designed for the WTA preprocessing layer is described. An optimization procedure based on genetic algorithm and practical recommendations for its implementation are also included in the paper. The article contains three practical examples similar to real-world problems, which serve as illustrations for ideas presented in this work.