Computers that function more like the human brain rather than conventional digital systems have captivated our imagination since computers were invented. Such machines would be based on neuronal-like networks rather than series of binary 1s and 0s and would require much less power compared to digital computers.

To make such a computer, researchers need to develop simple, energy-efficient electronic devices that mimic the brain’s building blocks – neurons and synapses.

Mimicking “integrate-and-fire” behaviour

In a biological neuron, a thin lipid-bilayer membrane separates the electrical charge inside the cell from that outside it. This separation, together with several electrochemical mechanisms, allows the cell to keep its membrane potential steady. The membrane potential can, however, be altered with the arrival of exciting and inhibiting postsynaptic potentials through the dendrites of the neuron. When the neuron is sufficiently excited, it produces an action potential and the neuron is said to fire or spike.

A team led by Evangelos Eleftheriou has now succeeded in mimicking this “integrate-and-fire” behaviour with its nanoscale electronic phase-change memristive device. The device, which “remembers” how much current has flowed through it, can be used to carry out neuron-like computations (see image).

Towards neuron-like networks

Phase change materials are found in a variety of everyday applications, such as rewritable optical discs. Made from chalcogenide alloys, such as GeSb2Te5, they can be made to quickly switch from being purely amorphous to purely crystalline by applying light or electric current to them. Their optical and electrical properties change as a result, and these changes can be used to determine their state, amorphous or crystalline (or 0 and 1) - ideal for storing binary data.

But even more interesting are the conditions that lie between amorphous and crystalline because these could allow the materials to behave like neuronal networks rather than series of just simple binary 1s and 0s. For example, controlling the crystallization with external inputs (such as light or electricity), so that it occurs progressively, eventually leads to a sudden change in the material’s conductance and an output spike. This spike mimics how a neuron behaves in response to many incoming postsynaptic potentials.

Complex computational tasks

If these phase-change neurons were combined with plastic synapses (like those that have already been made by other research groups), they could carry out complex computational tasks, such as analysing parallel data streams – for example those from social media and search engines. In this way, we would be able to predict the spread of infectious diseases for instance, trends in consumer behaviour and perhaps even how the stock market will behave tomorrow.

The “Internet of Things” and connected objects could also benefit from low-power, memristive devices such as these, says team leader and lead author of the study, Tomas Tuma.

Artificial neurons are inherently stochastic

As well their phase-change properties, the artificial neurons are also inherently stochastic, and so resemble biological neurons in this respect too, he adds.

“Stochasticity is an essential ingredient for constructing ‘neuronal populations’ and our brain naturally uses these to represent signals and cognitive states,” he explains. “We have now shown how these populations can be constructed from phase-change neurons. This is important for building dense, scalable neuromorphic systems for memory applications and computing,” he tells nanotechweb.org.

Neuromorphic systems could also provide an alternative to how we process information in today’s computers, he adds. Namely, they could perform calculations with co-located storage and thereby overcome the performance bottleneck of Von Neumann computers (in which computational units are physically separated).

Lincoln Lauhon of Northwestern University, who was not involved in this work, says that brain-like computing on existing solid-state platforms is impractical and inefficient. “However, the new devices made by Eleftheriou and colleagues emulate key neuronal behaviours, including stochasticity within a small footprint, suggesting a path towards co-locating energy efficient information processing and storage functions at high density."

The IBM team, describing its work in Nature Nanotechnology doi:10.1038/nnano.2016.70, says that it is now busy looking into novel algorithms based on artificial spiking neurons and their populations.

There's a full special issue on research using nanostructures for synaptic electronics in Nanotechnology.