Trustworthy Artificial Intelligence
Leader : Anne JEANNIN-GIRARDON
Associated members : Cécilia ZANNI-MERK, Ulvyia ABDULKARIMOVA, Samer EL ZANT
Temporary teaching staff : Romain ORHAND, Anna OUSKOVA-LEONTEVA
PhD students : Lalla Aicha KONÉ, Quentin CHRISTOFFEL, Nasser GHANNAD, Hiba KHODJI, Tam’si LEY, Julien MICHEL, Yohan SOLON
Objectives and approaches
Our goals are to design and develop more explainable, more ethical and more autonomous artificial intelligence systems.
Our approaches cover different fields of artificial intelligence in a broad sense and include:
- knowledge engineering and computer ontologies ;
- graphs, both in their theoretical and applicative aspects
- artificial reasoning and neuro-symbolic approaches
- evolutionary computation (evolutionary algorithms, genetic programming)
- collective intelligence and complex systems theory
- epistemology of artificial intelligence.
One of the specificities of our team lies in the exploitation of traditional bioinformatics methods to generate knowledge that can then guide artificial intelligence approaches. This original combination is based on the plurality and complementarity of the team's internal skills. In addition, we rely on concepts from the human and social sciences to better understand the notions of explicability, ethics and autonomy inherent to artificial intelligence. We are working on a form of machine learning capable of developing ethical laws associated with a principle of locality rather than universal scientific laws.
Fields of application and current projects
We are interested in the quality control of biological data, in particular gene sequences. In this context, we develop original approaches for the detection of gene prediction errors in multiple sequence alignments based on deep learning models. Due to the opacity of these algorithms, we are also interested in the design and development of parsimonious and quantifiable post-hoc explainability approaches.
AI and reasoning
We are interested in combining (and coupling) deep learning models (neural network-based systems) with semantic technologies (decision trees, rule-based systems, knowledge bases, ontologies, knowledge graphs, etc.) to design and develop explainable and ethical neuro-symbolic AI systems.
Category theory provides a better understanding of the ontologies that are constitutive of AI. Grothendieck's topoi allow us to think about local ontologies, in order to consider creating AIs whose objective is no longer to find universal laws written with arithmetic operators, but local laws written with logical operators, which would then allow to describe an ethic, which it is necessary to imagine for a good interaction between AIs and humans.
AI and graphs
Graph models support the properties of data traceability, model transparency and comprehensibility of the output of these models. Human interpretability is also improved over the multi-dimensional datasets typically used by machine learning. Explainable graph models can be used for feature extraction from complex data for learning, for knowledge representation and as a deep learning model by Graph Neural Networks. The main application domain is network monitoring and cyber security.
We also do fundamental research in graph theory, in order to improve our understanding of the general properties of graph structure (especially in relation to their various colorations), of the extremal values of invariants (extremal theory) and of the behavior of associated polynomials (Tutte polynomial). We also study the contribution of such an approach to identify rules and properties in genes.
In the field of education, we have released POEM v3 (Personalized Open Education for the Masses), the education flagship of UniTwin CS-DC. This version of POEM has been validated by the University of Strasbourg for course evaluation. The primary goal of POEM is to provide a social and intelligent platform for a massive and personalized educational ecosystem, in order to facilitate knowledge acquisition through personalized curricula involving the learner. Version 3 allows to implement a 4P mass education while automatically enhancing a warehouse of unstructured educational content. It is based on a MOOC system to which it adds the added value of AI algorithms to select and propose individual lessons adapted to the objectives, level and skills of the student.