PhD student · ENSTA, Institut Polytechnique de Paris
I am a PhD student in the U2IS lab at ENSTA, Institut Polytechnique de Paris, advised by Gianni Franchi and Goran Freshe. My research is on uncertainty quantification for deep networks, trying to make uncertainty estimates accurate, cheap enough to use at scale and reliable.
Most of my work is closley related to computer vision and i am quite interested in robust detection of AI generated media.
I also maintain open-source tools like Torch-Uncertainty among others with the goal of making most uncertainty quantification methods reproducible and easy to use.
Interests Uncertainty quantification · Computer vision
NeurIPS 2026
FakeParts are deepfakes with subtle, localized spatial or temporal manipulations of otherwise authentic videos. We present FakePartsBench over 81K videos (including 44K FakeParts) with pixel- and frame-level annotations and show they reduce human detection accuracy by up to 26% with similar degradation for sota detectors.
ECCV 2026
Detecting covariate shift without supervision remains a challenging problem. We address this through Positive–Unlabeled learning and introduce S-PUNA, a geometry-aware framework that progressively identifies shifted data by leveraging the local manifold structure of visual features.
ICLR 2026
Deep Ensembles achieve strong uncertainty quantification but are costly to scale to large models. Hydra Ensembles prunes attention heads to create diverse members and merges them via multi head attention with grouped fully connected layers yielding near single model inference speed while matching or surpassing Deep Ensembles without retraining from scratch.
NeurIPS 2025 Spotlight
Torch-Uncertainty is a framework built on top of PyTorch and Lightning that streamlines training and evaluation with UQ techniques and metrics, and benchmarks a diverse set of methods across classification, segmentation, and regression.
* Equal contribution
Teaching assistant, ENSTA / Data AI Master
Teaching assistant, MVA Master