XNOR-Nets with SETs: Proposal for a binarised convolution processing elements with Single-Electron Transistors

Abstract

AbstractDeep neural network (DNN) and Convolution neural network (CNN) algorithms have significantly increased the accuracies in cutting-edge large-scale image recognition and natural-language processing tasks. Generally, such neural nets are implemented on power-hungry GPUs, beyond the reach of low-power edge-devices. The binary neural nets have been proposed recently, where both the input activations and weights are constrained to $$+$$ +  1 and − 1 to address this challenge. Here in the present proof-of-concept study, we propose a simple class of mixed-signal circuits composed of single-electron devices and exploit the nonlinear Coulomb staircase phenomena to alleviate the challenges of binarised deep learning hardware accelerators. In particular, through SPICE modeling, we demonstrate the realisation of space-time-energy efficient XNOR-Accumulation (XAC) operation, reconfigurabilty of XAC circuit to perform 1D convolution and a busbar design to augment a contemporary accelerator. These nanoscale circuits could be readily fabricated and may potentially be deployed in low-power deep-learning systems.