Dynamical Neural Network Based on Spin Transfer Nano-Oscillators

Spintronic technology promises to significantly increase the efficiency and scalability of neural networks by employing optimized task-oriented device components that exhibit intrinsic nonlinearity, temporal nonlocality, scalability, and electrical tunability. In particular, the functional response of spin-transfer torque oscillators can be designed to naturally emulate the building blocks of neural networks, such as short-term memory, hierarchy, and nonlinearity. We propose spin-transfer nano-oscillators as a dynamic neuron that can be used in a neural network coupled with a fully connected layer to perform classification tasks. In this concept, successive nodes of the neural network correspond to successive time steps, so that the nonlinearity and memory of the system can be naturally exploited. The tunability of the device allows to project initial configurations in well-defined regions of the phase space where classification is easily performed. Furthermore, training is performed using optimal control theory. We emphasize that the devices benefit from more realistic models compared to simpler analytical models and is robust against device-to-device variations. We tested the performance of the network on two types of datasets and obtained 99% accuracy. Although these systems are computationally expensive, their hardware implementation is simple and inexpensive.