Reservoir Computing Leveraging the Transient Non-linear Dynamics of Spin-Torque Nano-Oscillators

Present artificial intelligence algorithms require extensive computations to emulate the behavior of large neural networks, operating current computers near their limits, which leads to high energy costs. A possible solution to this problem is the development of new computing architectures, with nanoscale hardware components that use their physical properties to emulate the behavior of neurons. In spite of multiple theoretical proposals, there have been only a limited number of experimental demonstrations of brain-inspired computing with nanoscale neurons. Here we describe such demonstrations using nanoscale spin-torque oscillators, which exhibit key features of neurons, in a reservoir computing approach. This approach offers an interesting platform to test these components, because a single component can emulate a whole neural network. Using this method, we classify sine and square waveforms perfectly and achieve spoken-digit recognition with state of the art results. We illustrate optimization of the oscillator’s operating regime with sine/square classification.