Load sharing in the training set partition algorithm for parallel neural learning
Bernard Girau, Hélène Paugam‐Moisy · 2002
A parallel back-propagation algorithm that partitions the training set on a ring of processors has been introduced. In this paper, we study the performance of this algorithm on MIMD machines and develop a new version, based on a heterogeneous load sharing. Algebraic models allow precise comparisons between the different methods, and show great improvements in case of parallel learning.>