Fully memristive spiking neural network for energy-efficient graph learning

• DOI: 10.1126/sciadv.adv2312
• Publication Date: 2025-05-07T17:58:56Z
• Authors: Tuo Shi; Lili Gao; Ruixi Zhou; Yang Tian; Pei Chen; Yanting Ding; Shuangzhu Tang; Huiqin Ma; Jian Lu; Hui Zhang; Zhanfeng Wang; Bo Lyu; Xumeng Zhang; Xiaobing Yan; Qi Liu
• Keywords: Memristor; Neuromorphic engineering

Parallel and energy-efficient searching of the shortest paths on a large graph is challenging. Conventional methods commonly used are sequential computing intensive, rendering them inadequate for addressing large-scale real-time situations. Here, we propose highly parallel, computation- approach to path–based learning based an emerging memristor spiking neural network via algorithm-device codesign. The path obtained parallelly in nature using simultaneous spike traveling instead arithmetic calculation, achieving extremely low time space complexity. A nonlinear weight mapping proposed counterbalance neuron intrinsic nonlinearity guarantee accuracy support graphs. hardware capability experimentally demonstrated unsupervised supervised classification tasks. estimated energy efficiency 517.82 giga-traversal edges per second watt outperforms field programmable gate arrays by three four orders magnitude, providing pathway toward hardware.