Université Paris-Saclay, CEA-List
I work on model quantization and what it does to the reliability of deployed vision and vision-language models: which predictions it changes, why, and how to keep the good changes and repair the bad ones.
Advances in Neural Information Processing Systems
Good and bad prediction flips separate along the SVD rank of the quantized model's own feature space, so a post-hoc gate can keep the good flips and repair the bad ones.
International Conference on Machine Learning
A large-scale evaluation of CLIP-family VLM quantization beyond accuracy: in specific regimes, quantization improves accuracy, calibration, OOD detection and noise robustness at once.
IEEE International Conference on Vehicular Electronics and Safety
On real adverse-weather driving images, 8-bit quantization costs a road-scene detector no blur robustness and makes a recognition encoder more robust to blur it never saw, while both run faster and smaller as true INT8 models.