Implementation of Graph Convolution Network Based on Analog Rram

In this work, the implementation of Graph Convolutional Network (GCN) based on resistive switching memory is demonstrated through simulation. After training, the RRAM-based GCN can process a semi-supervised graph classification task. Further, the impacts of read noises and circuit bit-precision on the performance of GCN are analyzed. Results show the proposed GCN can reach high accuracy when bit-precisions; 4-bit. Moreover, read noise can severely affect accuracy.