Researchers have proposed a new type of artificial neuron called a "straintronic spin neuron" that could serve as the basic unit of artificial neural networks—systems modeled on human brains that have the ability to compute, learn, and adapt. Compared to previous designs, the new artificial neuron is potentially orders of magnitude more energy-efficient, more robust against thermal degradation, and fires at a faster rate.
"Most computers are digital in nature and process information using Boolean logic," Bandyopadhyay told Phys.org. "However, there are certain computational tasks that are better suited for 'neuromorphic computing,' which is based on how the human brain perceives and processes information. This inspired the field of artificial neural networks, which made great progress in the last century but was ultimately stymied by a hardware impasse. The electronics used to implement artificial neurons and synapses employ transistors and operational amplifiers, which dissipate enormous amounts of energy in the form of heat and consume large amounts of space on a chip. These drawbacks make thermal management on the chip extremely difficult and neuromorphic computing less attractive than it should be.
"Fortunately, there are other ways to implement neurons, such as with magnetic devices. It was thought that magnetic devices will dissipate much less heat, but what we found is that they do not necessarily dissipate less heat in all circumstances. The heat dissipation depends on how the magnetic devices are switched to mimic a neuron's operation. If they are switched with current, which is the usual approach, then they do not dissipate that much less heat, and, in some circumstances, may even dissipate more heat than transistors.
"However, there is a way to switch certain types of magnets with mechanical strain generated by an electrical voltage. We found that if magnets are switched with that approach, then the magnetic neurons are indeed much less dissipative than both their transistor-based counterparts and current-switched magnetic counterparts. This is the 'straintronic spin neuron,' and it may provide a boost to neuromorphic information processing hardware."
With these advantages, straintronic spin neurons could have a variety of applications in neural computing.
"What we have studied is a perceptron, which is a mathematical model of the artificial neuron," Atulasimha said. "There are many possible applications of this in neural computing. One area we are interested in is spike-timing-dependent plasticity, which is a form of Hebbian learning. It is widely believed that it underlies learning and information storage in the brain, and there is a vast body of literature dealing with this. Straintronic spin neurons are fired by voltage impulses, and there are clear pathways to adapt them to the spike-timing-dependent plasticity model. We are also interested in character recognition, which employs feed-forward networks and image compression. That does not exclude anything else. Wherever heat dissipation is a spoiler, the straintronic spin neuron may be able to offer a solution."
