Mapping of multilayer perceptron networks to tree shape parallel neurocomputer
T. Hamalainen, H. Klapuri, J. Saarinen, Kimmo K. Kaski · 2002
Mapping of multilayer perceptron networks are presented for a parallel neurocomputer system called TUTNC (Tampere University of Technology Neural Computer). The tree shape architecture of the system offers implementation possibilities at several levels of parallelism for both forward mode of operation and network training. An example mapping is given for error backpropagation training utilizing combined neuron and weight parallel schemes. The performance is measured for some example network sizes, and estimations are given for a larger system.