Fig. 1: Random memristor based dynamic graph CNN (RDGCNN) for efficient point cloud learning.

a Point cloud data, represented by a series of discrete points, is widely applied in everyday life with the popularization of 3D sensors in devices like mixed reality headset, automobiles, drones, and smartphone cameras. b Traditional machine learning approaches for handling point clouds: Left: Method of voxelizing point clouds and using 3D convolutional operations for processing; Right: Method of representing point clouds as point sets and directly processing them through neural networks. c The traditional von Neumann architecture of processors, with separate computation and storage units, introduces significant data transfer overhead. d We convert the point cloud into a graph representation, where each point serves as a vertex in the graph, and the edges are determined by the features of the points. e Our random dynamic graph CNN method. Left: The network dynamically updates the vertex features and connections, extracting hierarchical information. Right: The random EdgeConv operation. The central vertex features and neighboring vertex features are concatenated and passed through a random CNN to obtain edge features. The updated central node features are calculated by aggregating the edge features associated with edges emanating from all neighbouring vertices, and new graph connections are computed based on the updated features. f Our memristor-based Computing-In-Memory (CIM) system. Leveraging the in-memory computing capability of memristor, it reduces data transfer overhead and utilizes the intrinsic stochasticity to obtain a random conductance matrix for the physical implementation of random CNN weights.