Soutenance de Victor Sechaud
| When |
Nov 02, 2026
from 10:00 to 12:00 |
|---|---|
| Where | Amphi A |
| Attendees |
Victor Sechaud |
| Add event to calendar |
vCal
iCal
|
Scientific observation and the measurement of physical quantities are intrinsically based on the use of sensors, whose function is to translate a real-world phenomenon into a usable signal. However, this acquisition process is inevitably subject to distortions: the measured signal represents only a partial, distorted and noisy projection of the original information.
An inversion problem involves reversing this measurement process to recover the source signal, a mathematically complex task that is often poorly defined, given that the acquisition operator is generally not invertible and that, in theory, an infinite number of signals could account for the same observation.
Whilst deep learning has revolutionised this field by offering excellent reconstruction quality, it typically requires vast databases of ‘ground truth’ data (the perfect signals paired with the measurements). However, in many applications such as medical imaging, astronomical photography or ultra-high-speed imaging, such reference data are impossible or extremely costly to acquire.
The central objective of this research is therefore to design and validate self-supervised learning methods capable of training high-performance reconstruction models using only raw measurement data, while providing strong theoretical guarantees, particularly for non-linear inverse problems. These guarantees revolve around two fundamental issues: on the one hand, model identification, which consists of proving that the set of admissible signals can be uniquely determined from the measurement data alone; on the other hand, signal recovery, which ensures that there is a unique solution associated with each observation within this set.