Neural network computation in a parallel multiprocessor architecture
Petri Kotilainen, Jukka P. P. Saarinen, Kimmo K. Kaski · 2005
A parallel multiprocessor architecture for general-purpose neurocomputing applications is introduced. Methods to map the multilayer perceptron, Kohonen's self-organising feature map and Kanerva's sparse distributed memory to the suggested architecture are discussed. The mapping examples include both the forward operation and training phase of the networks. The computational performance of the architecture is estimated for these three example cases.