Nat. Commun. 8, 14736 (2017)

In neural networks like the brain, information flows between neurons through synapses, and learning is achieved by changing the strength of synaptic connections. This strength evolves depending on the relative timing of electrical signals from neighbouring neurons. And now, Sören Boyn and colleagues have demonstrated unsupervised learning using artificial neural networks made from ferroelectric tunnel junctions.

Memristors based on ferroelectric tunnel junctions can act as solid-state equivalents of synapses because their resistance, which depends on the history of their exposure to electrical signals, can emulate synaptic strength. Boyn et al. showed that the intrinsically inhomogeneous nature of the polarization switching in such junctions can emulate the necessary conditions for synaptic strength adjustment. They then provided a corresponding model that can predict the evolution of the resistance of ferroelectric synapses for varying neural inputs.

They showed that a network comprising nine input neurons, connected to five output neurons by a crossbar array of ferroelectric memristors, can autonomously learn to recognize patterns made from 3 × 3 pixel images in a predictable way.