Design Of An Integrated Neuro-Genetic Processor For Pattern Recognition Applications

Abstract

One of the key concerns of modern VLSI design is realization of massively parallel architecture for high throughput and fast data processing. Artificial neural network are best example of such architecture whose implementation suffers from various bottlenecks including slow training and massive consumption of computational resources. This thesis is an effort to overcome the limitation of backpropagation trained artificial neural network by adopting a hybrid Neuro-Genetic processing and introducing novel genetic operators. A new mutation operator derived from statistical analysis has been proposed and pruning of redundant weights has been accomplished.

Furthermore, an improved mutate-discard-crossover scheme has been implemented which retains the fittest weights. Simulation experiments confirm the efficacy of the proposed techniques. With the same amount of error being obtained in less number of epochs and lesser computational burden.