A spintronic device measuring just 375 nm across has been used to recognize human speech. The device is a spintronic oscillator, which behaves much like a neuron in the brain. Created by physicists in France, Japan and the US, the system is described as the first neuromorphic computer that is based on a nanoscale device.
Neuromorphic computers try to emulate the human brain. As well as having the potential to be faster and more energy efficient than conventional computers, they could also excel at learning how to perform certain tasks – rather than being pre-programmed to do so.
A spintronic oscillator comprises a non-magnetic layer of material sandwiched between two ferromagnetic layers – with each ferromagnetic layer being magnetized in a different direction. A voltage is applied to the device, causing a spin-polarized current to flow from one magnetic layer, across the non-magnetic layer, and into the second magnetic layer. This exerts a torque on the second magnetic layer, causing its magnetization to precess at microwave frequencies. This precession is monitored in terms of an oscillating voltage that develops across the device.
Nonlinear response
A minimum current is required for these oscillations to occur. As the current rises above this threshold, the amplitude of the oscillating voltage increases as the square root of the current. This current threshold and nonlinear response is similar to the behaviour of neurons, which is one reason why spintronic oscillators show promise for making neuromorphic computers.
The speech-recognition system was created by Julie Grollier and colleagues at Université Paris-Sud and Université Paris-Saclay, the National Institute of Advanced Industrial Science and Technology in Tsukuba and the National Institute of Standards and Technology, Gaithersburg, Maryland.
The process begins with a spoken word being captured by a microphone, digitized and then pre-processed to create an electrical current. This current is then fed into a spintronic oscillator, creating an oscillating voltage that is then analysed by a computer running a machine-learning program.
State-of-the-art performance
The team looked at how the system is able to recognize the numbers 0–9 when spoken by several different people. When the input signals were pre-processed using a “nonlinear cochlear filter” – the standard in such applications – the system achieved a recognition rate of 99.6%. Writing in Nature, the team describes this as a “state-of-the-art” performance that is normally achieved using much more complicated systems.
As well as being sub-micron in size, the oscillators can be made using the same fabrication methods as conventional computer chips. This, says the team, could allow one hundred million oscillators to fit on a thumb-sized chip. The researchers also point out that unlike other nanoscale oscillators, spintronic oscillators offer low noise operation, high stability and low energy consumption.
