Skip to main content
European Commission logo print header

On-chip memristive artificial nano-synapses and neural networks

Final Report Summary - NANOBRAIN (On-chip memristive artificial nano-synapses and neural networks)

There is today a revival of brain-inspired hardware architectures. Indeed, because they are massively parallel systems, neuromorphic architectures have remarkable qualities: they are fast, they don’t need a lot of power to operate, and they can work even if a large number of their components is defective. These assets are now crucial to complement current processors that are too power-hungry, over-heating and beginning to suffer from the increasing number of defective transistors. In addition, the application scope of artificial neural networks is extremely relevant in the present context of data explosion: they are excellent at data classification related tasks, such as recognition and mining.

The performances of these massively parallel artificial neural networks on chip depend on their size and interconnection degree. It is therefore very important to have small elements to emulate the plastic synaptic interconnections. Memristor devices, that can mimic synapses at the nano-scale, can be the key to the future implementation of hardware neural networks as accelerators of next generation processors.

Memristors are tiny tunable analog non-volatile resistances. The more intense the current through a memristor, and the longer it is applied, the larger the resistance variations. This specific behavior explains the name “memristor” (memory resistor) and allows for example implementing the fact that the more an excitatory synapse is used, the better it transmits. There are many types of memristors based on different physical effects. Most technologies rely on deep changes of the device materials and structure under an applied voltage to induce large resistance variations, through thermal, phase change or ionic motion phenomena. Building purely electronic memristors in contrast to these existing resistive switching technologies can upgrade their performances in terms of endurance and speed.
In the ERC NanoBrain project we have fabricated two novel types of memristors, where the resistance variations are due to purely electronic effects: the spin torque memristor, based on the magnetic tunnel junction and the ferroelectric memristor, based on the ferroelectric tunnel junction. In both cases we have obtained the multi-state, quasi-analog resistance variation of memristors through a fine control of the nanoscale domain configuration during magnetization or polarization reversal. Because the physical origins of resistance variations of our memristor devices are well understood, we can model, optimize and predict the behavior of our devices. Together with the absence of forming step, this is a decisive advantage over a number of existing memristor technologies.

Finally, we have obtained the very first results of pattern recognition in a hardware neural network based on our memristors.