GUIDDE · Guaranteed Uncertainty for InformeD DEcisions

Линиите свързват координатора с партньорите.

Накратко на български

Цел на проекта

As machine learning (ML) is prevalent today to support decision-making, the European approach is to create an environment of trust for a fair AI through a unified regulatory framework, to ensure trust from both ‘‘people and companies’’. In line with this ambition, recent research efforts from the statistics and ML communities have been devoted to equipping ML models with provably valid tools for predictive uncertainty quantification (UQ), via methodological developments. We can now turn any point prediction into a guaranteed prediction set, or post-process estimated probabilities to get calibrated ones.Yet, a crucial question remains: what does this offer to the downstream pipeline? This project targets this gap by: A) characterizing the trade-offs between the two UQ holy grails, namely fairness (via conditional validity) and tight evaluation of the underlying model’s error (via sharpness); thus helping any decision-maker to accurately interpret UQ performance, and B) qualifying under which conditions calibration allows decision-makers to take better decisions. Based on mathematical statistics tools, the theoretical outputs of A) and B) will be confronted with a real-world scientific application: exoplanet detection, where UQ is a critical step directly influencing detection capability. They will guide the design choices of a tailored UQ method, given the specificities of the application, in the form of open-access code.The outgoing supervisor, a UQ expert, and the returning one, skilled in applying ML to real-world scientific applications, will help the PF grow into an independent researcher.This interdisciplinary project will also benefit the industry through a concrete exploitation plan, targeting France's main electricity producer and supplier, and a public-private consortium working with 16 hospitals to improve the care of severe trauma patients. The communication plan will directly contribute to raising public awareness of the possibilities for robust ML.

Оригинален текст от CORDIS (на английски).

Участници

• INSTITUT NATIONAL DE RECHERCHE EN INFORMATIQUE ET AUTOMATIQUE · Le Chesnay CedexКоординаторФранция
• THE REGENTS OF THE UNIVERSITY OF CALIFORNIA · OaklandСъединени щати

Връзки

• Виж в CORDIS
• DOI: 10.3030/101264481