Description du sujet de thèse

Quantum neural networks are attracting growing interest due to their ability to project data into high-dimensional Hilbert spaces, where they can become linearly separable. In addition, these networks open the possibility of learning directly from quantum data, thanks to their natural compatibility with other quantum systems capable of generating such data.

The most common approach relies on variational quantum circuits based on qubits. However, training such networks presents several challenges, including the estimation of output gradients with respect to internal parameters, as well as the vanishing gradient problem, which arises from the dilution of information in large Hilbert spaces.

In our team, we explore an alternative approach based on coupled bosonic modes. By using coherent states and continuous operations such as displacement and squeezing, this framework maintains a structured region of Hilbert space, which may reduce information dilution and facilitate learning.

A recently proposed training method for qubit-based networks — equilibrium propagation — enables the estimation of gradients directly from measurement results, without requiring simulation or differentiation of a quantum system model.

The objective of this PhD project is to adapt this method to parametrically coupled bosonic networks, and to experimentally demonstrate the learning of coupling parameters using this approach.

Contexte de travail

The PhD project is part of the ERC project QDYNNET – Quantum Dynamical Neural Networks, led by Danijela Marković. The selected candidate will join the neuromorphic computing team at the Albert Fert Laboratory, CNRS, Thales, Université Paris-Saclay. They will work in close collaboration with two other PhD students recruited for the project. The project is supervised by Danijela Marković and Julie Grollier (CNRS).