Publication Details
Highly Efficient Parallel ANN Implementation for Real-Time Processing
parallel architecture, hypercube, neural networks, performance, software pipelining
The message-passing parallel implementation of a multi-layer perceptron in recall mode requires regular computation and various group communications. The most efficient software pipelining and network topology are investigated.
The recent increased interest in parallel architectures has focused, among other areas, on digital signal processing and artificial neural networks (ANN). The topology ensuring the highest performance of multi-layer back-propagation ANN on the multicomputer is therefore important. Here we address the problem of the most efficient parallel implementation of the recall algorithm for a continuous stream of data on the hypercube topology with message passing. A prototype of a related parallel program has been written in TRANSIM and run in simulated real time. By overlapping comunication and computation processors utilization reached 82.7%. The resulting setup of distributed processing with speedups and efficiencies is presented.
@INPROCEEDINGS{FITPUB5663,
author = "V\'{a}clav Dvo\v{r}\'{a}k and Petr Matou\v{s}ek",
title = "Highly Efficient Parallel ANN Implementation for Real-Time Processing",
pages = "186--191",
booktitle = "Proceedings of conference Computer Engineering and Informatics CE\&I'99",
year = 1999,
location = "Kosice - Herlany, Slovakia, SK",
publisher = "Faculty of Electrical Engineering and Informatics, University of Technology Ko\v{s}ice",
ISBN = "80-88922-05-4",
language = "english",
url = "https://www.fit.vut.cz/research/publication/5663"
}