Optoelectronic and all-optical hyperspin machines for large-scale computing

Introduction

Efficient hardware for combinatorial optimization and machine learning impacts science, engineering, and society. With new computational models, photonics tackle problems intractable with conventional computing systems. However, existing devices only scale up to thousands of spins and operate at the second timescale.

Project Overview

I demonstrate photonic machines for ultrafast parallel processing of millions of spins with microsecond timescale. The strategy is minimizing a class of functions, the Ising Hamiltonian, by a new computational system that uses a high-dimensional feature space and speeds up optimization by orders of magnitude through ultrafast nonlinear optical processes.

Device Development

I build digital, optoelectronics, and all-optical classical and quantum devices and benchmark them with real-world, large-scale problems. By utilizing spatial modulation technology and a cheap, simple, and scalable design, light propagation is recurrently trained towards the ground state of a programmable Ising Hamiltonian.

Objectives

Starting from my proof-of-concept, I aim to achieve the following:

1. Energetically efficient computing of large-scale Hamiltonians.
2. Inclusion of self-optimizing all-optical nonlinear ultrafast phase-locking processes.
3. Demonstration of record combinatorial optimization by allowing the spins to evolve in a high-dimensional space to guarantee high success probability.

HYPERSPIM Framework

HYPERSPIM leverages the interplay of classical and quantum dynamics through the onset of entanglement and squeezing. The unprecedented scale and versatility allow the first quantum optimization tests for real-world complex computational tasks.

Conclusion

HYPERSPIM achieves the fastest and biggest optical computing device operating in classical and quantum regimes in an interdisciplinary route towards new photonic artificial intelligence, large-scale all-optical computing, and fundamental science.