Michaël Soumm

Welcome to my research page!

I am currently a Postdoctoral Researcher in Deep Learning. My research is conducted jointly at:

• IDS Lab at Télécom Paris under the supervision of Pietro Gori
• MLIA Lab at Sorbonne University under the supervision of Alasdair Newson

Current Research

My research focuses on disentangled and controllable latent representations in deep learning, with applications in computer vision and medical imaging.

Specifically, I explore how to structure latent spaces to improve interpretability and enable meaningful data manipulation.

Disentangling Latent Representations in Neuroimaging

In medical imaging, particularly neuroimaging, the separation of healthy and pathological patterns is a key challenge. I apply contrastive analysis (CA) and generative models to identify salient imaging patterns that differentiate psychiatric patients from healthy subjects, with the goal of improving diagnostic insights.

Generative & Diffusion Models for Structured Representations

State-of-the-art generative models—such as diffusion models—provide powerful tools for learning meaningful latent structures. My work aims to adapt these models to neuroimaging, improving image synthesis quality and representation disentanglement for better interpretability.

Research Interests

• Latent Space Structuring: Learning disentangled representations for deep learning applications.
• Contrastive Learning: Identifying key patterns in neuroimaging data.
• Generative & Diffusion Models: Leveraging modern generative approaches for medical imaging.

📌 Stay tuned for research updates and code releases!