Monolithically 3D Integrated Memristive Bayesian Neural Network for Intelligent Motion Planning

作者：Linbo Shan, Lindong Wu, Zongwei Wang, Rui Hua Xie, Chaoyi Ban, Gaoqi Yang, Qishen Wang, Li Yuan, He Ma, Lin Bao, Ling Liang, Yuan Wang, Yimao Cai, Ru Huang · 年份：2024 · DOI：10.1109/iedm50854.2024.10873381 · 被引用次数：7 · 研究领域：Advanced Vision and Imaging、Robotics and Sensor-Based Localization、Human Pose and Action Recognition

In this paper, we demonstrate a novel$\text{VO}_{2}$-based standard normal distribution random number generator (SD-RNG) unit and a memristive ReLU activation (MRA) unit. These units are monolithically 3D (M3D) integrated on a 40 nm 1Mb RRAM array chip to realize a fully memristive Bayesian neural network (BNN), significantly increasing the bandwidth and reducing$2.41\times$hardware cost. Intelligent motion planning tasks in deterministic and non-deterministic environments are implemented based on the M3D-BNN chip, demonstrating the software-level accuracy and$2.47\times$improvement of uncertainty predication ability compared with the deep neural network. Moreover, the M3D-BNN chip consumes$19.9\times$less energy and runs$2.1\times$faster than 2D counterpart. These results prove the great potential of the M3D-BNN chip in intelligent scenarios.