Domain adaptation from theory to practice – MATTER

Domain Adaptation (DA) is a fundamental problem in statistics, Machine Learning (ML), and data science where one wants to estimate a predictive model from labeled training data in the presence of a shift or change in the properties of the testing data. This problem is very common in practical applications and challenging due to the lack of labels in the shifted test data. Despite an active research community, DA methods remain rarely used in practice. The objective of project MATTER is to tackle theoretical and practical bottlenecks that prevent the wider use of DA in ML applications.

MATTER focuses on the root of the DA problem: the estimation of the shift between domains, which will be achieved using optimal transport, manifold estimation, and physical modeling. This will lead to novel and interpretable shift classification and estimation methods for DA. These results will be used to implement robust validation procedures that are still lacking in the DA community and pave the way for the first Automatic DA framework. MATTER will also address the general DA problem where multiple shifts are present between multiple datasets and propose interpretable and scalable (distributed) methods. MATTER will finally investigate the problem of Heterogeneous DA that can occur between heterogeneous datasets across devices and between structured data such as graphs.

The proposed methods will be validated with a new open benchmark framework on several types of data (computer vision, biomedical, audio). They will also be evaluated on a flagship biomedical application, sleep stage classification, where adaptation to the specificities of the subjects is necessary. A major output of project MATTER will be an open-source toolbox containing implementations of DA methods, the benchmark, and the first-ever AutoDA software. By addressing both theoretical and more pragmatic aspects of the field, project MATTER is a unique opportunity to unlock the potential of DA.