A3 : Artificial Learning and Applications
Artificial Learning is a scientific discipline concerned with the design and development of algorithms that improve their efficiency through experience (observations, labelled or not). It borrows techniques and theoretical knowledge from the fields of artificial intelligence, logic, statistics, to form a very complex field. Artificial learning is now a mature field of computer science, with solid theoretical models and results, and a wide range of applications, both in industry and in many research disciplines. It is also a field that has become very popular with the general public, thanks to the success of deep learning in gaming algorithms (Chess, Go..) and the emergence of Data Science for the intelligent processing and analysis of massive datasets.
The A3 team is structured in two research poles carrying out their activities in a complementary way:
Manager : Aomar Osmani
The common thread of research in the MAARS group is the use of structures, in addition to flat data, to build learning models. We are also interested in the contribution of a priori knowledge, and how to model it in order to have robust learning models requiring less data. Learning is not only an algorithmic issue, it is also the rigorous study of all the parameters involved in the learning process, the evolution of the models, and beyond that the respect of regulations and social rules when these models are used in real situations (health, transport, connected city, etc.).
Our group is working on structuring the meta-knowledge of the biases brought about by the entire knowledge production chain and trying to provide answers to the interpretation and meaning of learned models and their impact on society.
Our research focuses on four areas :
– Relational learning and graphs :
Learning from models expressed as logical, explicit and explainable programs and from data structured in graphs. This axis deals in particular with the themes of relational learning and uncertain examples, probabilistic relational learning in a POMDP framework, graph deabstraction and abstract closed pattern search.
– Scalable unsupervised learning :
Mixture models and simultaneous learning of data structure and clustering (deep unsupervised learning). This axis also focuses on multi-variate temporal data, large-scale learning and visualisation.
– Meta-learning or learning as a modelling problem :
Modeling the context, various biases, data status, interaction between concepts to be learned, evolution of learned theories and metamodeling of hyperparameters. This axis is also concerned with evaluation and decomposition models, data relativity and the complete learning pipeline.
Manager : Younès Bennani
The central theme of the ADA research cluster is learning from data and learners, with a focus on the unsupervised paradigm. The main axes of contributions focus on multi-model clustering (multi-view and collaborative), transfer learning, learning from structured data, and unsupervised deep learning of representations. The ADA cluster has a strong historical position in unsupervised learning and much of its research is in this field. The contributions of the ADA cluster cover both fundamental research and more applied research, most often supported by academic and industrial collaborative projects. Theoretical and algorithmic contributions are often developed in parallel with applications in various fields such as: data quality and anonymisation, health, digital marketing and recommendation, diagnosis of complex systems, social network analysis, etc…
The ADA cluster’s research is organised into three main themes :
– Unsupervised multi-model learning :
The ADA cluster is interested in the development of collaborative and multi-view learning methods for distributed and heterogeneous data through emerging formalisms such as quantum learning or optimal transport theory.
– Transfer learning :
The members of the ADA cluster seek to develop approaches that detect and avoid negative knowledge transfer, particularly in the context of optimal transport theory and non-negative matrix factorisation.
– Learning from structured data :
Collaborative and hierarchical approaches have been proposed to tackle different problems such as: Partitioning of large graphs into communities; Link prediction; Analysis of multiplexed graphs.