Thematic Axis: Representation and qualification of uncertainty
Leader(s)
Description
Here, we study methods to obtain and exploit information about the, usually implicit, uncertainty of an AI model or system. This includes both ways to assess uncertainty emerging from the training conditions of a machine learning or deep learning model (complexity of the problem, properties of the model, bias in the data), and the representation of uncertainty within knowledge systems, so that it can be reasoned upon. Those more theoretical aspects not only have direct implications on the applicability and adoptability of AI systems, but also lead to a greater ability to exploit uncertainty information in the building of such systems, as is the case for example in active learning.
The motivation for focusing on uncertainty qualification and representation as a research axis relates to its transversal nature: uncertainty is a fundamental property of most AI systems, and addressing it rigorously has implications across most of the other axes considered here. The most direct connection is with the frugal AI axis, through active learning: uncertainty estimates allow a system to identify the most informative examples to label in such a process. he explicability and interpretability axis is also closely linked, since communicating uncertainty is itself a form of explanation. The connection with the knowledge representations and reasoning axis runs in both directions: knowledge representation formalisms (e.g. for belief revision) can provide languages for representing and manipulating uncertainty, while the study of uncertainty also raises questions about how it should be encoded in knowledge graphs and ontologies. Finally, the NLP and LLMs axis is relevant for the reason that the challenges that already exist in qualifying uncertainty in classical machine learning and deep learning models are amplified as the scale and level of complexity of LLMs