ANP-G: A 28-nm 1.04-pJ/SOP Sub-mm² Asynchronous Hybrid Neural Network Olfactory Processor Enabling Few-Shot Class-Incremental On-Chip Learning

Abstract

The changing of sensitivity and selectivity profiles of gas sensors caused by sensor drifting and surroundings, leads to a serious decline in accuracy and even invalidation of the electronic nose (e-nose). To address this issue, we propose an asynchronous spiking and backpropagation hybrid neural network olfactory processor, asynchronous neuromorphic processor-gas recognition (ANP-G), which is the first gas recognition processor enabling on-chip class-incremental learning with few shot and concentration estimation. Its key features include: 1) a bio-inspired spiking neural network (SNN), approximation to the mammalian olfactory system that enables rapid retraining over nine types of gas with no accuracy loss to overcome the disturbance caused by sensor drift; 2) spike-timing-dependent plasticity (STDP) with lateral inhibition supporting few shot that achieves over 95% accuracy with less than five training samples per gas; and 3) asynchronous SNN circuits with a self-skipping epoch learning (SSEL) mechanism to skip unnecessary weight update reducing over 84% training costs. These features allow the processor to recognize more than nine kinds of gases with 97.85% accuracy, and estimate gas concentration with a 1.83% error rate. The chip is fabricated in a 28-nm CMOS process with a peak efficiency of 1.04 pJ per synaptic operation (SOP) at 0.55 V. The olfactory processor with on-chip learning provides a promising solution for low-power portable intelligent e-nose systems.