Neuromorphic computing with STT-MRAM devices as processing units capable of learning pattern recognition

A promising approach in neuromorphic computing seeks to perform cognitive tasks directly in hardware leveraging the physics of well-established, efficient nano-devices. Spintronic devices, such as spin torque oscillators and superparamagnetic tunnel junctions, are strong candidates for neuromorphic networks. However, the most mature and widely adopted spintronic technology is spin-transfer torque magnetic random-access memory (STT-MRAM). While STT-MRAM devices have primarily been used as synaptic (memory) elements, they have not yet been experimentally demonstrated as processing units capable of learning pattern recognition. In this work, we present a reconfigurable classifier based on a network of magnetic tunnel junctions that learns to categorize data encoded in the amplitude of input currents through spin torque-driven magnetization switching. We experimentally show that a network of three junctions can classify seven spoken vowels with 96% accuracy, outperforming software neural networks with the same number of trained parameters. This represents a significant step toward the simple and fast large-scale integration of neuromorphic hardware.