Approximate Fully Connected Neural Network Generation

Abstract

Approximate computing is exploited in implementation of fully connected networks for classification problems. A multiplier structure whose area is scalable over accuracy through approximate computing is proposed. In order to employ the multipliers in a network, an area reduction algorithm is formed. It can adjust the approximation level of multipliers while still maintaining the target classification performance, without prior information on the value of network weights. Implementing on a Spartan6 FPGA, up to 79% area saving is recorded for various performance targets.