REALM

The RealM research project aims to rethink updates for machine learning models particularly in the context of federated learning, where a super model centralizes updates for multiple instances. Drawing inspiration from software testing approaches, RealM seeks to prevent updates from causing deviant behavior. In addition, the project envisages the creation of a model management system, similar to a package manager, which would centralize the various models, managing their versions and documentation, as well as test sets to evaluate their performance over time. RealM aims to guarantee the reliability and security of models deployed in complex systems, while facilitating their evolution and continuous evaluation over time.

With the support of Wallonie-Bruxelles International and the FRS - FNRS.