Organisation
The French Alternative Energies and Atomic Energy Commission (CEA) is a key player in research, development and innovation in four main areas :
• defence and security,
• nuclear energy (fission and fusion),
• technological research for industry,
• fundamental research in the physical sciences and life sciences.
Drawing on its widely acknowledged expertise, and thanks to its 16000 technicians, engineers, researchers and staff, the CEA actively participates in collaborative projects with a large number of academic and industrial partners.
The CEA is established in ten centers spread throughout France
Reference
SL-DRT-27-0397
The reliability of artificial intelligence models capable of analyzing videos (Video Large Language Models) is a major challenge, particularly in application domains where misinterpreting visual information can lead to inappropriate decisions or generate misinformation.
Currently, these models tend to confuse what is actually visible in a video with what is merely plausible or expected. As a result, they may claim that an action has been completed even when the video does not provide visual evidence that it actually occurred.
The main objective of this thesis is to equip these models with the ability to distinguish between three situations: (1) an event has occurred, (2) an event did not occur (supported by sufficient visual evidence), or (3) the available information is insufficient to reach a conclusion.
Three complementary research directions will be explored:
· The first aims to develop methods that teach models to identify the visual cues that are truly decisive in confirming or refuting an action.
· The second focuses on developing reasoning mechanisms capable of identifying the observations that should necessarily be visible if an event had occurred. This approach will enable the generation of 'non-occurrence certificates', i.e., structured justifications based on visual evidence supporting the conclusion that an event did not take place.
· The third investigates whether using these certificates as additional supervision can improve model performance while reducing the amount of training data required.
The proposed approaches will be evaluated in terms of accuracy, robustness to ambiguous situations, and data efficiency. Ultimately, this research aims to develop more reliable video models capable of justifying their conclusions rather than simply predicting what seems likely.
Sciences et Technologies de l’Information et de la Communication (STIC)
Paris-Saclay
KARA Sandra sandra.kara@cea.fr
CEA
DRT/DIASI//LASTI
CEA SACLAY - NANO INNOV
BAT. 861
Point courier 173
91191 GIF SUR YVETTE
PHAM Quoc-Cuong Quoc-Cuong.PHAM@cea.fr
CEA
DRT/DIASI
CEA SACLAY - NANO INNOV
BAT. 861
Point courier 173
91191 GIF SUR YVETTE
01 69 08 27 16