Few-shot graph learning with robust and energy-efficient memory-augmented graph neural network (MAGNN) based on homogeneous computing-in-memory

Woyu Zhang, Shaocong Wang, Yi Li, Xiaoxin Xu, Danian Dong, Nanjia Jiang, Fei Wang, Zeyu Guo, Renrui Fang, Chunmeng Dou, Kai Ni, Zhongrui Wang, Dashan Shang, Ming Liu

Abstract: Learning graph structured data from limited examples on-the-fly is a key challenge to smart edge devices. Here, we present the first chip-level demonstration of few-shot graph learning which homogeneously implements both the controller and associative memory of a memory-augmented graph neural network using a 1T1R resistive random-access memory (RRAM). Leveraging the in-memory computing paradigm, we validated the high end-to-end accuracy of 78% (GPU baseline 80%) and robustness on node classification of CORA dataset, while achieved 70-fold reduction in latency and 60-fold reduction in energy consumption compared with conventional digital systems.

External IDs:dblp:conf/vlsit/ZhangWLXDJWGFDN22