Algorithmic mapping of MLP network on Neural DF KPI architecture

Artificial neural networks gain increasingly higher popularity in application fields with stress laid on acquiring information in real-time. Although there are various implementations of neural networks available for sequential computer systems, lots of these models require an enormous amount of time for the training phase of neural network in case of large network. Therefore the new time saving concepts were developed especially for time reduction of network training thus including modification of original models and learning algorithms as well as implementation of these models on parallel computer system. Conventional formulation of fundamental neural algorithms make implementation of neural networks on existing parallel hardware more difficult, therefore a solution for effective network parallelisation in the form of algorithmic mapping of multilayer feedforward neural network with backpropagation learning (FFBP) on massively parallel system neural DF KPI was searched in the framework of VEGA 1/1064/04 project performed on Department of Computers and Informatics at Technical University of Kosice.