In-Memory Euclidean Distance Computation in a Stacked Memristor Crossbar for Hardware Self-Organizing Maps

Abstract

Self-organizing maps (SOMs) are widely used for visualizing and clustering high-dimensional data, yet their efficient hardware realization has remained challenging because Euclidean distance (ED) computations are difficult to implement directly within memory arrays. Conventional SOM accelerators typically rely on peripheral digital processors or approximate similarity metrics, which increase latency and energy consumption while limiting scalability. Here, we present a SOM learning architecture based on a 2 x 32 x 32 stacked memristor crossbar array that enables in-situ, fully parallel ED evaluation by in-memory computing. By exploiting the middle electrode current of the stacked crossbar structure, the system inherently encodes the ED between input and weight vectors, eliminating external arithmetic units and preserving full analog parallelism. We further develop a complete unsupervised learning pipeline mapped directly onto the crossbar, encompassing distance evaluation, competition, and weight adaptation, demonstrating seamless hardware integration of SOM operations. Experimental demonstrations on diverse tasks, including the traveling salesman problem, image clustering, and color quantization, validate the functional correctness and high hardware efficiency of this approach, enabled by the massively parallel in-memory distance computation. This work advances the hardware realization of SOMs and establishes a general methodology for mapping distance-driven unsupervised learning algorithms onto memristive computing-in-memory systems, enabling scalable, energy-efficient neuromorphic accelerators.