Quantum optical neural networks are Variational Quantum Circuits composed of alternating layers of Gaussian and non-Gaussian transformations, which can be trained to become virtually any device.
As the field of integrated optics makes leaps year after year, our results will soon be implementable in real devices. We aim at developing a variety of methods and to build a solid knowledge base to take advantage of the cutting edge technology of integrated optics.
What we are working on:
- Developing fast differentiable simulations to allow for the computation of gradients. These are necessary in order to run optimisation algorithms such as Gradient Descent and its many variants and improvements. As the technology matures, the computation of the gradient itself can happen directly in the quantum hardware. We have recently developed an algorithm to simulate and optimize photonic networks which is up to 100x faster than the previous state-of-the-art .
- Applying enhanced versions of the Gradient Descent algorithm such as higher derivative methods (e.g. Hessian) and methods that take into account the curvature and topology of the parameter space (e.g. the Natural Gradient). These are known to accelerate the rate of convergence, which leaves us with more resources for optimising larger systems.
- Devising regularisers and loss functions that enhance the convergence of the optical network toward a desired objective, for example entanglement-based regularisers and entropy-based loss functions. Exploring the use of entropy-based loss functions will open new possibilities to optimise also devices used for Quantum Key Distribution.
While this program is being developed, we will constantly evaluate the current performance of our methods. As the main benchmark of the effectiveness of the various ideas, we optimise a one-way quantum repeater. We already have experience on this precise type of task and therefore we have a baseline to compare against. As a bonus, the one-way quantum repeater is a device which still lacks a CV implementation, and our suite of optimisation methods will likely yield viable candidate architectures.
 “Fast optimization of parametrized quantum optical circuits” Filippo M. Miatto, Nicolás Quesada, arXiv:2004.11002 [quant-ph]