Generative models with physical structure

Physics-informed generative models for Mueller polarimetric imaging

At CMAP, École Polytechnique, I work on deep learning for Mueller polarimetric imaging, with applications to medical diagnosis. Each pixel of a Mueller image is a 4×4 matrix describing how a sample changes the polarisation of light, and the models must respect the physical constraints these matrices satisfy.

• Deep generative architectures for Mueller matrix modelling.
• Physical constraints (positivity, Mueller matrix realizability) built into the network rather than added as a penalty.
• Synthetic data generation for small clinical datasets, and evaluation by downstream segmentation.
• Uncertainty-aware learning, including geometry-aware label smoothing.
• Self-supervised pre-training of physically consistent representations (JEPA-style architectures).

Generative AI

I am particularly interested in variational autoencoders, diffusion models and representation learning. A recurring question is how a latent space can encode meaningful scientific structure, rather than only optimise a reconstruction or generation objective.

Scientific machine learning and uncertainty

My broader interests include physics-informed machine learning, Bayesian deep learning, uncertainty quantification, computer vision, explainable AI and probabilistic modelling. I favour approaches that connect learning theory with measurable physical quantities and with the limits of the available data.

Internship report

Master 2 MS2A, Sorbonne Université · CMAP, École polytechnique · 15 April–15 September 2026 · written in French

Physics-constrained Generative Models and Representations for Mueller Imaging in Low-Data Regimes. Supervisors: Erwan Le Pennec (CMAP) and Angelo Pierangelo (LPICM).

The report studies representation learning and generative models for Mueller polarimetric imaging when annotations are scarce. It introduces the Stokes-Aware Mueller Generator (SAMG), which learns a compact latent representation and rebuilds the Mueller matrix by a regularised inversion of parametrised Stokes responses. It then studies how to generate latent fields (autoregressive models, Gaussian processes, diffusion, Flow Matching) and how to use them for classification and segmentation under low annotation budgets. The experiments are methodological and are not a clinical evaluation.

© 2026 Ayoub Oulad Ali. Tous droits réservés.

VAEAC: Gaussian prior and Shapley explanations

Academic project · Sorbonne Université, MS2A · February 2026

An implementation of a variational autoencoder with arbitrary conditioning (VAEAC), used to generate samples conditional on any subset of features, so that Shapley values respect the dependence between variables. The report analyses how the Gaussian prior affects those explanations.

Code and further reading

Research code is on GitHub (@ayouboa30), in particular MuellerPT and Stokes-Aware-Mueller-Generator. A short, public-safe summary of the research is available as a research summary.