Mathias Vast
I study how to improve the robustness, effectiveness and efficiency of neural retrieval models — opening them up with interpretability first, then improving their architecture.
About
I'm a 3rd-year PhD student in Information Retrieval within the MLIA team at ISIR, Sorbonne Université, working in collaboration with Sinequa. My research is supervised by Benjamin Piwowarski and Laure Soulier.
Improving the Robustness, Efficiency and Effectiveness of Neural Rankers in Information Retrieval
The goal of this thesis is to jointly improve the robustness, effectiveness, and efficiency of neural IR models, and in particular of cross-encoders. To this end, we first studied their adaptation to new domains without supervision, before turning to interpretability in order to understand how they model relevance internally. These findings, combined to the exploration of the most effective training strategies, ultimately led to design MICE, a new cross-encoder architecture that significantly reduces computational cost while preserving, and even improving, the model’s effectiveness and robustness.
Publications
an IR researcher's site should let you retrieve!
0 results retrieved — perhaps not my work?
Contact
Open to discussions on IR (as a whole) and interpretability — feel free to reach out at vast@isir.upmc.fr.