Learning dynamics on the picosecond timescale in a superconducting synapse structure

Conventional Artificial Intelligence (AI) systems are running into limitations in terms of training time and energy. Following the principles of the human brain, spiking neural networks trained with unsupervised learning offer a faster, more energy-efficient alternative. However, the dynamics of spiking, learning, and forgetting become more complicated in such schemes. Here we study a superconducting electronics implementation of a learning synapse and experimentally measure its spiking dynamics. By pulsing the system with a superconducting neuron, we show that a superconducting inductor can dynamically hold the synaptic weight with updates due to learning and forgetting. Learning can be stopped by slowing down the arrival time of the post-synaptic pulse, in accordance with the Spike-Timing Dependent Plasticity paradigm. We find excellent agreement with circuit simulations, and by fitting the turn-on of the pulsing frequency, we confirm a learning time of 16.1 +/- 1 ps. The power dissipation in the learning part of the synapse is less than one attojoule per learning event. This leads to the possibility of an extremely fast and energy-efficient learning processor.