Efficient Finite Word Length Determination For Neural Networks Implementation

Most of the artificial neural networks (ANN) based applications are implemented on FPGAs using fixed-point arithmetic. The problem is to achieve a balance between the need for numerical precision, which is important for network accuracy, and the cost of logic areas, i.e. FPGA resources. In this paper we propose a genetic algorithm based methodology permitting the optimization of the FPGA resources needed for the implementation of a pipelined recurrent neural network (PRNN) while respecting the precision constraints. The quality of our methodology would be evaluated through experiment on a PRNN based WCDMA receiver. Our methodology is not restricted to this class of ANNs and can be used for any complex with variable dimensions system