标题： 在超导突触结构中以皮秒时间尺度学习动力学 显示英文标题

标题： Learning dynamics on the picosecond timescale in a superconducting synapse structure

摘要： 传统人工智能（AI）系统在训练时间和能耗方面遇到了限制。遵循人脑的原理，使用无监督学习训练的尖峰神经网络提供了一种更快、更节能的替代方案。然而，在此类方案中，尖峰、学习和遗忘的动力学变得更加复杂。在这里，我们研究了学习突触的一种超导电子实现，并对其尖峰动力学进行了实验测量。通过用超导神经元脉冲系统，我们展示了超导电感器能够动态保持突触权重，并由于学习和遗忘而更新。通过减慢突触后脉冲到达时间，可以停止学习，这与基于尖峰时间依赖可塑性的范式一致。我们发现电路仿真结果与实验结果非常吻合，并通过拟合脉冲频率的开启时间，确认了 16.1 ± 1 ps 的学习时间。突触学习部分的功耗小于每次学习事件一个阿托焦耳。这使得极快且节能的学习处理器成为可能。 显示英文摘要

摘要： 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.