Design and Implementation of Stochastic Neural Networks Using Superconductor Quantum-Flux-Parametron Devices

Stochastic logic has been proven to be a low-cost hardware solution to accelerate neuromorphic computing thanks to its bit-wise operation feature, which removes the most hardware. On the other hand, the superconductive electronic device exhibiting stochastic behaviour and deep-pipelining nature is a perfect candidate for implementing stochastic logic-based neuromorphic computing. The adiabatic quantum-flux-parametron (AQFP) logic family is demonstrated with the highest energy efficiency among its superconducting cousins. In this work, we introduce the recent designs and implementations of the AQFP-based neural network prototypes utilizing the stochastic computing paradigm. We further report the low-temperature measurement results of two different implementations using approximate-parallel-counter (APC) and bitonic-sorter based design approaches. Thanks to its adiabatic switching nature and zero-static power dissipation, the AQFP-based implementation averages 5-6 orders of energy efficiency compared to its CMOS counterpart.