Showing 1–1 of 1 results for author: Rasoulinezhad, S

MajorityNets: BNNs Utilising Approximate Popcount for Improved Efficiency

Authors: Seyedramin Rasoulinezhad, Sean Fox, Hao Zhou, Lingli Wang, David Boland, Philip H. W. Leong

Abstract: Binarized neural networks (BNNs) have shown exciting potential for utilising neural networks in embedded implementations where area, energy and latency constraints are paramount. With BNNs, multiply-accumulate (MAC) operations can be simplified to XnorPopcount operations, leading to massive reductions in both memory and computation resources. Furthermore, multiple efficient implementations of BNNs… ▽ More Binarized neural networks (BNNs) have shown exciting potential for utilising neural networks in embedded implementations where area, energy and latency constraints are paramount. With BNNs, multiply-accumulate (MAC) operations can be simplified to XnorPopcount operations, leading to massive reductions in both memory and computation resources. Furthermore, multiple efficient implementations of BNNs have been reported on field-programmable gate array (FPGA) implementations. This paper proposes a smaller, faster, more energy-efficient approximate replacement for the XnorPopcountoperation, called XNorMaj, inspired by state-of-the-art FPGAlook-up table schemes which benefit FPGA implementations. Weshow that XNorMaj is up to 2x more resource-efficient than the XnorPopcount operation. While the XNorMaj operation has a minor detrimental impact on accuracy, the resource savings enable us to use larger networks to recover the loss. △ Less

Submitted 26 February, 2020; originally announced February 2020.

Comments: 4 pages

Journal ref: International Conference on Field-Programmable Technology, {FPT} 2019,Tianjin, China, December 9-13, 2019