The stochastic switching behavior of magnetic tunnel junctions (MTJs) has been proven to be suitable for low-power neuromorphic computing. In this work, we demonstrate that the microwave-modulated stochastic switching behavior in a stochastic MTJ can be used to mimic sigmoid neurons in an artificial neural network to recognize handwritten digits in the Mixed National Institute of Standards and Technology database. The learning speed could be enhanced by the microwave and the flexibility of the neurons could be modulated by the power of rf signal. By modulating the neurons, the training speed of the system could be quicker than a conventional software-based sigmoid neural network, and the recognition accuracy reaches 87% within 30 epochs (while 80 epochs for the latter). Our work provides inspiration for alternative spintronic neuromorphic computing systems.

Close

PrevNext

View figure in article

• Energy-efficient infrastructure
• Neuromorphic computing
• Spin optoelectronics
• Spintronics
• Stochastic processes
• Artificial neural networks
• Magnetic tunnel junctions
• Microwave techniques