Exploiting Stochasticity in Superconducting Flux-Quanta Memory for Energy-Efficient Neuromorphic Computing

Abstract Employing superconducting memory to construct neuromorphic hardware offers a promising route to low-energy and high-speed computation while mitigating the data-transfer bottleneck between superconducting functional modules inherent to the conventional von Neumann architecture. In this work, we develop an in-memory computing architecture based on a superconducting multi-fluxon storage device made of nanoscale JJs. It is capable of implementing fundamental neural network operations, and supports 4 discrete synaptic weight storage and probabilistic weighted computation of inputs, achieving clear classification performance and 100% accuracy on a9-pixell image recognition task. By increasing the discretization levels of synaptic weights, the architecture is expected to handle more complex information-processing tasks, such as recognition of the Modified National Institute of Standard and Technology handwritten digit dataset. These results highlight the potential of superconducting memory devices as scalable and powerful building blocks for emerging computing architectures toward general artificial intelligence.