Substantial evidence indicates that the brain uses principles of non-linear dynamics in neural processes, providing inspiration for computing with nanoelectronic devices. However, training neural networks composed of dynamical nanodevices requires finely controlling and tuning their coupled oscillations. We just showed that the outstanding tunability of spintronic nano-oscillators can solve this challenge. We trained a hardware network of four spin-torque nano-oscillators to recognize spoken vowels by tuning their frequencies according to an automatic real-time learning rule. Our results show that non-trivial pattern classification tasks can be achieved with small hardware neural networks by endowing them with non-linear dynamical features : here, oscillations and synchronization.
We performed real-time learning with an experimental neural network composed of four coupled spin-torque nano-oscillators. The inputs to the neural network are two frequencies fA and fB that encode for spoken vowels (colored dots). The network classifies by synchronizing oscillators to the inputs. The seven vowels are classified in different synchronization configurations which are represented in different colors. Learning is performed by tuning the frequency of oscillators through their bias current. When training is finished, vowels are classified in different synchronization regions, and the recognition rate reaches 89%.
What are spin-torque nano-oscillators and how we trained them to recognize vowels (3:30 min)
Real-time learning (0:20 min)