Parallel Implementation of Artificial Neural Networks for the Diagnosis of Coronary Artery Disease

摘要

In this paper, we investigate implementations of artificial neural networks on Parallel Virtual Machine (PVM) for the prediction of coronary artery disease. Real data from international medical organizations are used in the training and testing of the basic backpropagation algorithm. To exploit the inherent parallel nature of ANN algorithms, we implement the basic backpropagation algorithm on Parallel Virtual Machine, which allows a heterogeneous collection of workstations and supercomputers to function as a single high-performance parallel machine, with various degree of parallelism. We investigated two modes of parallelization, namely node-partitioning and data-partitioning. Our results indicate that for a large neural network, we gain time-efficiency by partitioning the neural network into several parts on different machines, hence the name node-partitioning. For a small-size neural network with large amount of data, it is more advantageous to partition the data on different machines with each machine keeping a copy of the entire network, hence the name data-partitioning. To ensure robust convergence, we use a weight adaptation algorithm, Delta-Bar-Delta, in which each weight has a different learning rate in data-partitioning. Speed of convergence is somewhat faster if the initial learning rate is chosen according to the network size. From the results of our experiments with node-partitioning and data-partitioning, it shows that we can benefit from parallel implementation if the ratio of computation time versus the communication time between processes is large. Thus the algorithms proposed in this paper are applicable for real world problems with large size of network or large amount of data.