Physically implemented analog in-memory distance computing with InGaZnO transistor-based capacitive unit

Abstract

Modern edge devices increasingly require real-time adaptation to their environment without relying on cloud-based updates, which can introduce latency and security risks. To meet these demands, memory-augmented neural networks (MANNs) have gained traction for enabling adaptive on-device learning. Hardware implementations of MANNs commonly use non-volatile memory-based ternary content-addressable memory (TCAM), but their discrete outputs and write-verify steps limit compatibility with gradient-based learning. This work introduces analog in-memory distance computing (AIMDC), a unified architecture based on indium gallium zinc oxide (IGZO) thin-film transistors that performs both embedding extraction and similarity search using analog capacitive units (ACUs). By producing continuous, differentiable outputs directly from analog embeddings, AIMDC enables hardware-in-the-loop on-device representation learning without additional processing. We demonstrate energy-efficient one-shot learning accuracy comparable to a graphics processing unit but with up to a 576× improvement in energy efficiency. The high retention and endurance of the IGZO-based ACUs establish AIMDC as a scalable and robust solution for high-throughput, low-energy edge learning.