Superconducting artificial synapses integrated into a self-training neuromorphic architecture

Superconducting electronics is a compelling technology for designing neuromorphic computing systems [1]. In this work, we propose a new superconducting obtained with a SQUID-based circuit using a flux storage loop for the synaptic weight [2]. Using SPICE simulations, we demonstrate that the synaptic circuit can be tuned using digital single flux quantum pulses. In the second part of this presentation, we demonstrate the implementation of the synapses in a neural network. We integrate the artificial synapses into a small-scale self-training neural network architecture using SPICE simulations [2]. The network follows reinforcement learning rules that update local weights internally. This property allows the network to learn new functions by changing the target output for a given set of inputs without needing any external adjustments. Finally, using the same constraint as for the network mentioned above, we extend our simulations in Python to a larger problem to classify the MNIST dataset and show that the MNIST network can be trained in about 100 ms [2]. [1] Schneider et al, (2022) Supercond. Sci. Technol. 35 053001 [2] Schneider, Jué et al., arXiv:2404.18774