Published: January 28, 2018
Michael L. Schneider, Christine A. Donnelly, Stephen E. Russek, Burm Baek, Matthew R. Pufall, Peter F. Hopkins, Paul D. Dresselhaus, Samuel P. Benz, William H. Rippard
Neuromorphic computing is a promising avenue to dramatically improve the efficiency of certain computational tasks, such as perception and decision making. Neuromorphic systems are currently being developed for critical applications ranging from self-driving cars to cancer diagnoses1-3. An ultra-high-speed, low-power neuromorphic hardware platform that is scalable to levels of complexity matching the human brain, will greatly advance the power and range of applications of neuromorphic systems beyond what is presently achievable in software and custom CMOS systems. While software implementations of deep learning neural networks have made tremendous accomplishments4-6 they are still many orders of magnitude less efficient than the human brain1. Here we demonstrate a new form of artificial synapses based on dynamically reconfigurable superconducting Josephson junctions. Each artificial synapse is composed of two superconducting Nb electrodes with an interposing barrier of nominally amorphous-Si containing Mn nanoclusters. We define an artificial synapse as a component that can take a series of input spikes and non-linearly process them into an outgoing series of spikes based on an internal synaptic configuration that can be trained. The critical current of each synapse junction can be tuned by input voltage spikes that change the spin alignment of Mn nanoclusters in a manner consistent with Hebbian learning7. These synapses are technologically compatible with Josephson-junction single-flux-quantum-based neurons8-11 with low-dissipation superconducting transmission lines forming efficient axons. Together, these elements comprise a highly scalable technology platform for signal processing with naturally spiking signals, embedded memory, and tunable stochasticity. The spiking rate (> 1 GHz), spike energy ( < 10-20 J), small size (< 100 nm), and 3D stacking of these neurons provide the means for a neuromorphic platform can attain far greater complexity than before.
Citation: Science Advances
Pub Type: Journals
Neuromorphic computing, Artificial Synapse, Neuronal Network Dynamics, Magnetic Josephson Junction
Created January 28, 2018, Updated November 10, 2018