Efficient Design of Artificial Neural Networks using Approximate Compressors and Multipliers

Abstract

Nowadays, Artificial Neural Networks (ANNs) secured impressive results with multiple applications and approaches in various research fields, as well as image processing, face recognition and semantic segmentation. Here, the focus is to minimize the complexity of ANN hardware in keeping accuracy as a major concern. ANN is a subsystem that is approximate, in machine learning where it trains the neurons to get the relevant output according to the target value. By using this ANN, interfacing can be possible between approximate arithmetic circuits. 3:2, 4:2 compressors are designed with unique error positions, usually gives better power area and delay constraints in between 5 to 25%. The designed approximate ANN gains the design constraints in the range of 3 to 30%. The simulation results were done by using synopsys design compiler at 90nm Technology.