Coherent Spintronic Networks for Neuromorphic Computing

Introduction

Neuromorphic computing uses networks of artificial neurons highly interconnected by artificial synapses to perform vast data processing tasks with unmatched efficiency, as needed, for instance, for pattern recognition or autonomous driving tasks. The synaptic connections play a paramount role in creating better hardware realizations of these networks.

Connectivity Challenges

However, it is very complex to realize large interconnectivity by electronic circuitry. COSPIN overcomes this connectivity constraint by using the eigen-excitations of the magnetic system - the spin waves - to connect state-of-the-art artificial neurons based on spintronic auto-oscillators.

Project Goals

COSPIN’s main goal is to create and experimentally validate innovative physical building blocks for a novel nano-scaled, all-spintronic network structure which incorporates all necessary properties for neuromorphic computing, including:

• High nonlinearity
• Interconnectivity
• Reprogrammability

Methodology

By design, COSPIN works at the boundary between oscillator-based computing and wave-based computing. It uses:

1. Interference
2. Frequency-multiplexing
3. Time-modulation techniques
4. Spin-wave amplification

These methods significantly increase the connectivity between neurons.

Reprogramming Techniques

Reprogramming of the network is implemented by:

• A direct physical link to magnetic memory solutions
• Reconfiguring spin-wave circuits

By using coherent wave interference and nonlinear wave interaction, COSPIN paves the way for novel coupling phenomena for complex artificial neural networks far beyond the state-of-the-art of current hardware realizations.

Experimental Approach

Using cutting-edge micromagnetic simulations enhanced by inverse design methods, the artificial networks will be designed and tested prior to their nano-fabrication.

Investigation Methods

Experimental investigations will be mainly carried out using micro-focus Brillouin light scattering. This allows for local investigation of the individual neurons and synapses, and significantly simplifies the interpretation of the network dynamics.