Spike-Predictable Neuron Circuits with Adaptive Threshold for Low-Power SNN Systems

Abstract

This paper proposes an output spike-predictable comparator based on an adaptive threshold value method (ATVM) for obtaining a low-power neuron circuit. The proposed comparator operates during the predicted time at which the membrane voltage and threshold voltage coincide. This prediction-based power-gating method can help decrease the static power consumption of the comparator. In addition, the ATVM increases the threshold in proportion to the number of output spikes, and thus, the reduced use of the main comparator further decreases the power consumption. With the 28 nm complementary metal-oxide-semiconductor process, a framework with 144 input layers, 25 hidden layers, and 10 output layers was trained using MATLAB((R)). Modified National Institute of Standards and Technology (MNIST) classification operations were conducted using 250 synapses and 10 neurons. Using the proposed comparator and ATVM, the total power consumption of the comparator could be reduced by 90.37% with a supply voltage of 1.8 V. The accuracy of the MNIST classification using the ATVM was 95.02%.