Spintronics-based Neuromorphic Computing for Deep Learning

This project aims to propose new models for neuromorphic computing based on promising spin electronics (spintronics) technologies (i.e. magnetic skyrmions) for deep learning applications. In particular, it aims to model new architectures of artificial synapse devices that emulate biological synapses. Neuromorphic, or bio-inspired, computing is inspired by biological structures such as brains or nervous systems. Neuromorphic computing could have numerous advantages over classical computing, namely pattern recognition, the ability to imitate learning, increased fault tolerance and, crucially, lower power requirements that can lead to more energy-efficient devices. Applications ranging from image search to voice recognition in smartphone virtual assistants are built with a so-called deep learning approach in neural networks, which has witnessed an explosion of interest by the major technology companies (Google, Apple, etc.). While existing neural network implementations are software-based, bio-inspired computing would need to be built-in on the hardware level in order to fully exploit its numerous potential advantages. CMOS, which is the mainstream technology, is an excellent way to build such systems, e.g. [1].

At the same time, there is an effort to invent CMOS-compatible nanodevices that can emulate biological synapses at the nanoscale. Spintronics nanodevices offer a number of advantages such as combining, at the nanoscale, computation with on-site magnetic memory, which is non-volatile [2]. This could allow for on-site storing of the parameters of a system, such as the synaptic weights of a neural network. Magnetic skyrmions have attracted extensive interest due to their unique physical characteristics and are considered promising candidates for information carriers in logic and memory technological applications. Magnetic skyrmions are nanoscale topologically stable vortex-like spin windings that are extremely robust due to their topology. Their small size (nanoscale), robustness against material defects and low electrical currents needed to move them can lead to next generation ultra-dense, robust and low-power spintronic devices [3]. They have recently been demonstrated experimentally at room temperature for the first time in technologically relevant multilayers [4] which opens the way for utilising them in novel applications, such as recently proposed skyrmion-based artificial synapse devices for neuromorphic systems [5]. Specifically, a multi-bit storage device using skyrmions as information carriers (bits) has recently been proposed, where the state of the device is modulated by an electric current shifting the skyrmions in and out of the device. Such a behaviour, in which the state of the system (“weight”) can be dynamically adapted to the environment, is analogous to a biological synapse and their synaptic plasticity. This project will use computational tools to explore architectures that emulate basic synaptic behaviour as well as systematically investigate the optimal parameters/strategies for efficient, low-power and parallelisable spintronic-based artificial synapses. The project will utilise GPU-based micromagnetics simulations to propose strategies for skyrmion-based bio-inspired devices that can efficiently operate as artificial synapses.

Tools to be used: Micromagnetics (e.g. OOOMF and the GPU-accelerated MuMax3 framework).

Skills desired: Engineering or Computer Science or Physics degree with strong mathematical/analytical background.

Familiarity with script programming/Matlab/Mathematica/Octave is an asset.

Keywords: spintronics, neuromorphic computing, magnetic skyrmions, post-CMOS, nano electronics, neural networks