An Analog Integrate-and-Fire Neuron with Robust Soft Reset Mechanism

Abstract

This paper presents a novel soft reset technique for analog Spiking Neural Networks with minimal circuit complexity. The reset mechanism plays a vital role in SNNs, affecting the inference accuracy directly. Implementing the soft reset mechanism in analog SNNs has traditionally posed challenges due to the difficulties in subtracting static analog values, which is vulnerable to noise and variations. Utilizing the Schmitt trigger, the proposed reset mechanism changes the threshold voltage and the sign of the membrane potential accumulation, rather than implementing an analog subtractor. Compared to the conventional analog neurons, the proposed neuron is immune to error in passing the residual charge to next output spike generation.