Multilayer spintronic neural networks with radiofrequency connections

Area of Science:

• Physics, Materials Science, Computer Science
• Spintronics
• Artificial Intelligence

Background:

• Spintronic nano-synapses and nano-neurons offer high accuracy for neural network operations.
• Scalable multilayer connectivity for these devices remains a challenge for advanced deep neural networks.

Purpose of the Study:

• To demonstrate a scalable method for connecting spintronic nano-devices into multilayer neural networks.
• To enable spintronic neural networks to process radiofrequency (RF) signals natively.

Main Methods:

• Utilizing magnetic tunnel junctions (MTJs) as both spintronic neurons and synapses.
• Connecting MTJs into a two-layer hardware spintronic neural network.
• Classifying nonlinearly separable RF inputs using the developed network.

Main Results:

• A two-layer spintronic neural network composed of nine MTJs was constructed.
• The network achieved 97.7% accuracy in classifying RF inputs.
• Simulations show potential for state-of-the-art drone identification with milliwatt power consumption.

Conclusions:

• Spintronic magnetic tunnel junctions can be interconnected in multilayer configurations for neural network applications.
• This approach facilitates native processing of RF signals, paving the way for deep, dynamical spintronic neural networks.
• The technology offers significant power savings compared to existing methods for RF signal identification.