Relationships between artificial intelligence and human experts or even "lambda" users (in the context of human-machine interfaces or gamification processes) offer many advantages and open up prospects for the application of AI in sensitive areas (medicine, justice, etc.). Thus, these human-IA interactions can improve the performance of AI algorithms, help optimise complex processes or make consensual decisions, while playing an important role in the development of trust mechanisms towards AI (see Research Area 2) and the extraction of new knowledge. It is then a question of providing explanations, intelligible by an expert in the application domain, about the decisions and predictions suggested by a "black box" algorithm. All these interactions can take different forms such as "Human in the loop" or "AI in the loop" or the creation of consensus mechanisms. This first line of research focuses on these different forms, which are not mutually exclusive.
The work package is coordinated by Prof. Christine Decaestecker and Dr. Alberto Franzin at ULB. On 3 November, a kick-off meeting of the WP1 activities on Human-IA interaction was organised. This meeting, in a hybrid format, was attended by about 40 researchers (of which 19 in person), including PhD students, postdocs, Principal Investigators (PIs), and researchers from the CRAs.
Twelve presentations were given on this occasion. A general introduction of this first research axis for the attention of the researchers linked to the different partners was carried out so that all could have a good understanding of the work expected in the framework of this work package.
This kick-off also allowed PhD students and postdocs to present their research work associated with WP1, notably explainable AI, the use of imperfect and/or incomplete annotations on biomedical images, active learning, and other aspects of the use of AI in relation to humans, including the collaborative development of open source software.
The programme also featured two longer invited presentations on more general topics of interest to participants. One on deep learning in histopathology in the presence of multiple expert annotations, by Adrien Foucart (research assistant in the last year of his PhD at LISA, ULB), and one on biases in machine learning, by Christine Decaestecker. The participants gave positive feedback on the quality of the presentations and the interest of the research topics presented.
So far, 32 researchers have identified WP1 as their primary research area (see breakdown in the table below). However, the kickoff meeting also saw the participation of researchers (from universities and CRAs) whose main research area is in other WPs, but who are also interested in human-IA interaction, which is a very broad area. We welcome the interest in this research area. The next meeting is scheduled for early 2022. This will be an opportunity for researchers to share their work but also to revisit the correspondence between this work and the sub-tasks of the grand challenges now that most of them have been activated. Regular meetings will then be planned to ensure good synchronisation of research and Grand Challenges and to support the work of young doctoral students funded by ARIAC. These will be held remotely, co-modally or face-to-face when health conditions allow.