A methodology for evolving spiking neural-network topologies on line using partial dynamic reconfiguration

There is no systematic way to define the optimal topology of an artificial neural network for a given task. Heuristic methods, such as genetic algorithms, have been widely used to determine the number of neurons and the connectivity required for specific applications. However, artificial evolution uses to be highly time-consuming, making it unsuitable for on-line execution. Herein we present a methodology to evolve neural topologies on digital hardware systems. Evolution is performed on line thanks to the partial reconfiguration properties of Virtex II FPGAs. The genome encodes the combination of different layers, which, once downloaded to the FPGA, compose a neural network. The genetic algorithm execution time is reduced, since the fitness is computed on hardware and the downloaded configuration streams have a re-