ANP-L: A 22nm 0.72pJ/SOP Asynchronous SNN-based Edge-AI Processor Enabling One/Few-shot On-chip Incremental Learning

Processing multisensory data usually requires substantial computational resources, which is challenging for power-constrained AI-edge processors. Moreover, AI-edge processors deployed in constantly changing environments are expected to adapt to the dynamic environment automatically to maintain performance even when data sources are disrupted. To address these challenges, we introduce a 22nm asynchronous SNN-based processor enabling one/few-shot on-chip lifelong learning (ANP-L), which has three key features: 1) A dynamic SNN model with STDP of adaptive learning rate and lateral inhibition effect on synapse delays, enabling support for one/few-shot learning. 2) A hardware framework with on-chip incremental learning, which enables maintaining accuracy to overcome data disturbances caused by environmental changes. 3) The event-driven clock-free asynchronous circuit techniques, saving the dynamic power of neuron circuits, memory and the training time by 81.3%, 23.9~52.9% and 65% respectively. These key features enable ANP-L to obtain 94.78% average inference accuracy with on-chip learning in recognition tasks including hand gesture recognition with vision, keyword spotting with hearing, object recognition with touch, and gas identification with olfaction, and a peak energy efficiency of 0.72pJ per synaptic operation (SOP) at 0.5V.