Transputer-based parallel implementation of neural nets for a class of pattern recognition problems

Abstract

Neural networks are systems that are made up of a number of computing elements connected to each other making use of some of the organizational principles that are thought to be used in the human brain. For a large array of neurons, the calculations involved are prohibitively large for conventional serial machines to handle. So parallel implementation o f the network which helps in faster operation is desirable. Such an implementation is similar to the operations considered to be done in the brain. Hence, one o f the major objectives of this work has been parallel implementation o f neural networks. The theme o f our work is parallel implementation of backpropagation network and Bidirectional Associative Memory(BAM) on various topologies. Different transputerbased parallel architectures have been chosen as the platform for simulation. The BAM has been implemented on transputer-based topologies like hypercube, mesh and linear array. The speedup and utilization of these topologies for varying number of transputers and varying number of neurons in the layers are found ou t. The hypercube topology gives better speedup and utilization factors compared to mesh and linear array. The backpropagation network has been implemented on a linear array and ring of transputers. Speedup is obtained during learning and recalling operations on varying number of transputers. The ring network performs better during the learning operation and the linear array performs better during the recalling operations. The BAM and backpropagation networks have been used in the field of numeral recognition. The concepts of multiple training and dummy data augmentation have been used for increasing the storage capacity of the BAM. A condition for multiple training the BAM for many patterns is discussed. A comparison is made between BAM and backpropagation networks for numeral recognition under noisy conditions. The backpropagation network is found to be a better classifier in the presence of noise. This network has also been used in the field of medicine for diagnosing arthritis and related rheumatic disorders The network is able to identify and classify various arthritis and allied rheumatic disorders.