Parallel Implementation of Rbf Neural Networks Ecole Normale Supérieure De Lyon Parallel Implementation of Rbf Neural Networks

This report presents several parallel implementations, on a MIMD machine, of a learning algorithm called OLS (Orthogonal Least Squares) for RBF (Radial Basis Function) neural networks. The sequential version is rst described, and a straightforward parallel version is proposed. Two variants are developed, one of them reducing the complexity of the algorithm, and the other one improving the load balancing. An alternative is proposed for the storage of initial or intermediate data on local memory and discussed , according to the size of the application. Ce rapport pr esente plusieurs impl ementations parall eles, sur machine MIMD, de l'algorithme d'apprentissage OLS (Orthogonal Least Squares) pour les r eseaux neuronaux RBF (Radial Basis Function). La version s equentielle est rappel ee, puis une pre-mi ere parall elisation est expos ee. Deux variantes sont ensuite propos ees, l'une r eduisant la complexit e de l'algorithme, l'autre am eliorant l' equilibrage des charges. On discute de l'opportunit e de stocker, en m emoire locale, des donn ees initiales ou interm e-diaires, en fonction de la taille de l'application. Mots-cl es: R eseaux RBF, algorithmes parall eles, equilibrage de charge.