Julie Grollier
spin torque

memristor nanodevices for Artificial Neural Networks is a famous recent example

A recent example of such dynamics from device to architecture is the memristor, an analog, voltage-tunable nano-resistance with a memory effect. These resistive switching devices belong to the top 3 technologies foreseen to be able to replace CMOS in next generation binary memories. For this reason, their performances are now benefiting from a strong academic and industrial research effort. But memristors are not just interesting as binary memories.

Thanks to their non-volatility and tunability, memristors can emulate the synaptic behavior at the nano-scale. In particular, when they are driven by spiking neurons, they can implement a learning rule, called Spike Timing Dependent Plasticity. This last rule, which has been observed in biology, also allows non-supervised learning in artificial neural networks. This means that the network learns by itself to perform some tasks, like extracting particular features from a video.

This discovery has strongly contributed to revive research in large scale hardware artificial neural networks. Indeed this field is suffering from the difficulty to down-size CMOS synapses, currently limited to hundreds of Ám2 . As the network performances depend on its size and interconnectivity, replacing CMOS synapses by nanoscale memristor synapses would yield tremendous gains in terms of silicon area and energy consumption.

The drive to integrate memristor devices in large scale neural network is at the heart of many current research projects ( US DARPA SyNAPSE , EU ERC NanoBrain , EU Human Brain Flagship project, French ANR P2N MHANN, and many others).