Computers that function more like the human brain rather than conventional digital systems will rely on neuronal-like networks rather than series of binary 1s and 0s. They will be able to more easily handle the vast data sets currently being generated around the world thanks to being massively parallel. To make this new generation of machines, researchers need to develop simple, energy-efficient electronic devices that mimic the brain’s building blocks – neurons and synapses.

The new device, developed by Vincent Garcia and colleagues at CNRS-Thales and the universities of Bordeaux, Paris-Sud and Evry, is made from a memory resistor or memristor (a resistor that “remembers” how much current has flowed through it). Unlike other modern-day electronics memories like those made from CMOS devices, memristors are able to remember their state (that is the information stored in them) even if power is lost. They also use much less energy.

Modelling short-term plasticity

Synapses are the biological junctions between neurons and they transform a voltage spike (action potential) arriving from a pre-synaptic neuron into a discharge of chemical neurotransmitters that are then detected by a post-synaptic neuron. These are then transformed into new spikes, leading to a succession of pulses that either become larger or smaller.

This fundamental property of synaptic behaviour is known as short-term plasticity, which is related to a neural network's ability to learn. It is this plasticity that Garcia and colleagues have succeeded in mimicking. They were able to do this thanks to the intrinsic electronic nature of their memristor synapses based on ferroelectric tunnel junctions, in which resistance is connected to ferroelectricity.

“Learning rule”

“In ferroelectric tunnel junctions, switching the ferroelectric polarization with an electric field induces large variations in the tunnel resistance,” explains Garcia. “This is called tunnel electroresistance. In the OFF state, the polarization thus points downwards and in the ON state it points upwards. What is more, polarization usually occurs via the nucleation and propagation of domains, so by applying moderate and short voltage pulses, we can stabilize various configurations in which domains with up and down polarization co-exist. This means that we have not only a binary non-volatile memory but also a memristor.”

The researchers looked into the polarization dynamics of these tunnel junctions using scanning probe microscopy to image the domains and tunnel transport. “With the help of first-principle molecular dynamics calculations, we were also able to develop a simple model to explain the polarization switching,” Garcia tells “We then applied voltage waveforms that are similar to neuronal impulses and observed how the tunnel-junction resistance changes depending on the time difference between these voltage waveforms. This ‘learning rule’, called spike-timing dependent plasticity (STDP), is powerful and should allow for unsupervised learning in artificial neural networks.”

Learning to recognize patterns in a predictable way

And that is not all, the researchers say they were also able to reproduce the STDP of the ferroelectric memristors with their model based on ferroelectric domain dynamics and show that arrays containing hundreds of thousands of ferroelectric nanosynapses can autonomously learn to recognize patterns in a predictable way.

The team, reporting its work in Nature Communications DOI: 10.1038/NCOMMS14736, is now busy building a crossbar array of ferroelectric memristors connected to CMOS-based neurons to show that its systems can indeed learn unsupervised and recognize patterns. “For example, in the H2020 ULPEC project (that has just started), led by Sylvain Saïghi at the University of Bordeaux, we will collaborate with Chronocam, who have developed an event-based camera to process the information received by this camera using a neural network containing the crossbar array. The end-user partner Bosch will use the camera (which should detect movements much faster than conventional devices) in car sensors to assist drivers in the case of unexpected objects or persons crossing the road.”