Intrinsic and Parallel Performances of the OWE Neural Network Architecture

. The OWE (Orthogonal Weight Estimator) architecture is constituted of a main MLP in which the values of each weight is computed by another MLP (an OWE). The number of OWEs is equal to the number of weights of the main MLP. But the computation of each OWE is done independently. Therefore the training and relaxation phases can straightforward parallelized. We report the implementation of this architecture on an Intel Paragon parallel computer and the comparison with its implementation on a sequential computer. 1 Introduction One of the heaviest problems of ANNs is the execution time, that is often unsuitably high. Parallel implementations of the training and relaxation phases have been studied for many years [3, 2]. One of the most studied parallelization of ANNs is that of the MLP. The main strategy used for its parallelization is to distribute the multiplications of the connection weight matrix W per the neuron output vectors on several processors [1, 4, 10]. Two goals are pursued: t..