Challenges in implementation of ANN in embedded system

Abstract

Artificial Neural Networks (ANN) provides a simple and efficient method to implement highly non-linear complex systems due to its “Universal Function Approximation” capabilities. However lack of a simple hardware design that is capable of adopting any changes in operating environment of the system limits the applicability of ANN in automotive and industrial environment. The most challenging task for implementation of ANN in embedded plat-form is realization of non-linear sigmoidal activation function. This paper aims to address various hardware implementation issues of ANN in terms of speed, accuracy and resource utilization. Inverse Definite Minimum Time (IDMT) characteristic has been realized and verified using XILINX Spartan-3AN FPGA with very simple ANN model. Sigmoid activation function played a very crucial role in designing and implementation of ANN. Among various techniques piece wise linear approximation (PLAN) has found to be the most optimized and hardware friendly methods for implementing of sigmoid function on reconfigurable FPGA platform.