Performance analysis of Backpropagation Algorithm of Artificial Neural Networks in Verilog

Abstract

Artificial neural networks learn or get trained to execute definite tasks, instead of programmed computational systems through training algorithms such as Backpropagation algorithm. It is the basic tool for pattern classification application of ANN in field of medical diagnosis and remote sensing. This method involves changing the weights in the network using a training set of input output examples. Digital implementation of these neural networks for classification is suitable as it preserves the parallel architecture of the neurons and can be reconfigured by the user with FPGA. This parallelism in neural networks make it potentially fast for computation of tasks. This work implements backpropagation training algorithm in verilog with Modelsim-Altera 6.5b for a feedforward neural network. Since multiplier algorithms determine the operational speed and power consumption; a performance analysis is made based on different multipliers. The work can be further extended to the implementation of artificial neural networks on FPGA and to implement a classification application in the trained network.