Interpretable quantum machine learning enabled by integrated photonics

Classical Machine Learning algorithms are notoriously difficult to interpret, making error identification and debugging extremely challenging. In this project, we aim to explore the potential advantages that quantum photonics can bring to ML algorithms. Diverging from conventional, gate-based variational quantum algorithms, our approach draws inspiration from the quantisation of projective simulation (PS), an interpretable classical machine learning model. The proposed model shares the graph structure with classical PS[3], where decision-making is realised through a traceable and interpretable random walk.
In our implementation, this random walk becomes a quantum walk of a single photon undergoing unitary transformations, with the variational parameters becoming the phases of a homogeneous mesh of Mach-Zehnder interferometers.