Biologically Sound Neural Networks for Embedded Systems Using OpenCL

István Fehérvári, Anita Sobe, Wilfried Elmenreich

Abstract: In this paper, we present an OpenCL implementation of a biologically sound spiking neural network with two goals in mind: First, applied neural dynamics should be accurate enough for bio-inspired training methods, thus resulting network data is reproducible in ”in vitro” experiments. The second is that the implementation produces code that runs adequately on up-to-date embedded graphical chips for fast on-board classification applications, e.g., video image processing. We describe the necessary steps required to implement an efficient algorithm using the OpenCL framework and present evaluation results of the execution time compared to traditional serial CPU code. We show that an optimized GPU kernel code can perform sufficiently fast to be used for future embedded neural processing.