Hardware neural networks for energy efficient machine learning tasks

Reservoir computing is a machine learning model based on the use of a recurrent, black-box, artificial neural network called a reservoir, where the input is processed by being projected into a higher dimension space. This setup can be used to perform interesting and concrete tasks, such as speech and pattern recognition. Usually, the neurons constituting the reservoir are modeled as a software using classic CMOS units. Spintronics, which describes the spin-dependent behavior of conduction electrons in metallic materials and the influence of magnetic fields on the latter, can be used to implement a hardware and eco-energetic version of the artificial neurons (and hence, the reservoir) needed for the solving of a problem in the framework of reservoir computing. This also allows to avoid the limitations inherent to the von Neumann architecture, on which the classical software neural networks are usually implemented. To do so, the non-linear dynamics of the magnetization of a specific type of magnetic nanostructures is used. These nanostructures are called spin-torque vortex nano-oscillators (STVOs). In this work, the dynamics of a STVO is modeled using the Thiele equation approach (TEA), which prevents the use of expensive and time-consuming micromagnetic simulations. Different original models (numerical and analytical) are then used to do so, leading to the simulation of an entire neural network. These models are tested using a custom-made benchmarking pattern recognition task in order to assess their characteristics. Parametric studies are also performed to test the impact of different parameters on the efficiency of the simulated neural network.