Implementing and Mapping Anns on Reconfigurable Mesh Massively Parallel Architectures

Abstract

Artificial neural networks (ANNs) have been used successfully for solving hard problems in many disciplines. Multi-layer feed-forward ANNs with back-propagation learning is one of the most widely used models. Methods for efficiently implementing and mapping this model on general-purpose, massively parallel reconfigurable mesh architectures are presented. IntroductionParallel implementation of artificial neural networks (ANNs) involves mapping of neural network models onto parallel architectures. Due to the large, diverse, and growing number of learning algorithms and architectures in the ANN literature, it is impossible to come up with a single classification scheme that can categorize all the essential features of all these paradigms. Nonetheless, ANN models are frequently characterized by their learning rules (supervised, un-supervised, hybrid), their network topology (feed-forward, recurrent, single layer, multi-layer), and neuron characteristic. Some of the most commonly used and discussed ANN models include multi-layer feed-forward networks with supervised 373 Neural Networks and Systolic Array Design Downloaded from www.worldscientific.com by MONASH UNIVERSITY on 04/23/17. For personal use only.