Simulation of sparse neural networks on a CNAPS SIMD neurocomputer
P. Paschke, Ralf Möller · 1997
Neuroanatomical aspects of the mammalian cerebral cortex can be modeled by neural networks with a sparse and random connection scheme. This paper presents such sparse network models and appropriate algorithms, data structures and optimization for an efficient parallel simulation on a CNAPS SIMD neurocomputer. Using these methods a considerable speedup in comparison to sequential computation is achieved. INTRODUCTION The complete (or at least regular) interconnection between or inside groups of neurons is characteristic for the vast majority of artificial neural networks. Present general purpose, massively parallel computers including CNAPS and other neurocomputers are especially suited for the implementation of these artificial neural networks, but work much less effective when applied to sparse random networks like those modeling neuroanatomical aspects of the mammalian cerebral cortex. Nevertheless, using special simulation algorithms and data structures including preparatory optimiz...