[20a-M206-1~10] FS.1 Focused Session "AI Electronics"

[20a-M206-4] Small-area CiM for Event-based Vision Sensor Co-optimized by ReRAM device & SNN Algorithm

〇Shunsuke Koshino¹, Naoko Misawa¹, Chihiro Matsui¹, Ken Takeuchi¹ (1.Univ. Tokyo)

Keywords:Spiking Neural Network

This work proposes a compact Computation-in-Memory (CiM) co-optimized by multi-level ReRAM and Spiking Neural Network (SNN) algorithm for Event-based Vision Sensor (EVS) of edge AI. Proposed SNN removes frequent weight update during training by adopting random weights, which reduces the endurance of ReRAM. Excellent 94.7% recognition accuracy for EVS gesture dataset is achieved. With ReRAM cell error analysis, proposed SNN with 2-bit protection and 0.1% errors achieves 90.1% accuracy. By co-optimizing ReRAM device and SNN algorithm, proposed error-tolerant Heterogeneous Multi-level ReRAM CiM decreases the number of required ReRAM cells by 75%.