Design of neural networks for solving computational problems

Abstract

Neural network implementation using analog circuits has the advantage that computational problems such as multiplication or addition can be realized with simple circuits. In addition, analog circuits are faster than digital implementation and occupy a small silicon area. A software program for simulation and realization of artificial neural nets by using the backpropagation algorithm is designed. An analog neural network is implemented for realizing XOR function using D-MOS transistors acting as synaptic weights and bipolar transistors to represent the nonlinear sigmoid function. Computer simulations for this network are performed with the Pspice program. The learning phase is done in a very fast time. Experimental results confirm the theoretical considerations.