Evolving Spike-Train Processors

The research described in this paper was motivated by the idea to process purposefully given spike-trains using a cellular automaton (CA). CAs have three attractive features, namely massive parallelism, locality of cellular interactions, and simplicity of basic components (cells). However, the difficulty of designing a CA for a specific behavior causes limited interest in this computational paradigm. Automating the design process would substantially enhance the viability of CAs. Evolving CAs for purposeful computation is a scientific challenge undertaken to date by, among others, Mitchell et al. [1], Sipper et al [2] and de Garis et al [3]. In previous work [4], we designed a special 2-dimensional cellular automaton, called qCA, which can be used for purposeful manipulations on binary time-series (spike-trains). The proposed procedure of qCA synthesis is to decompose the problem into logic synthesis and routing. As for logic synthesis, we evolve a graph as a halfproduct for the qCA synthesis, called Pulsed Para-Neural Networks (PPNN). PPNN consists of functional nodes and directed edges representing pure delays, in which each node returns a pulse at clock t if it received one and only one pulse at clock t-1. The routing is a conversion of the obtained PPNN into an initial state of a qCA using a relatively trivial heuristics. The desired PPNN is not assumed to be evolved from scratch. In this paper, we described an evolutionary method to find unknown delays of a given scheme so that the network could produce desired spike-trains. We investigated the scheme shown in Fig.1(a) as a test-bed of our method. The chromosomes of GA are represented as