Simulation of cortex-like neural networks on a CNAPS SIMD neurocomputer

Abstract

Neurons of the cortical tissue in a mammalian brain are connected in an extremely sparse and random fashion. The paper presents efficient methods for a parallel simulation of neural networks modeling this connection scheme on a CNAPS SIMD neurocomputer. Appropriate algorithms and data structures are introduced that allow for a minimal loss of parallelism during the computation of input scalar products. A ‘greedy’ optimization procedure applied to the neuron-processor assignment is shown to gain a further reduction of computation time getting close to the lower bound. Using these methods, a considerable speedup in comparison to sequential computation is achieved.