STAGE Study of the impact of domain shift assumptions on transfer learning methods

In classical machine learning, one assumes that the source data used to train an algorithm comes from the same distribution as the target data it is applied to. This assumption is not true for many applications: for instance, a human activity recognition model trained on young people may not perform well when applied to older ones [1]. This kind of distribution shift issue happen in numerous real scenario of machine learning applications and often degrade the model performance. To correct these shifts, different machine learning approaches have been developed within the field of domain adaptation or transfer learning [2].

Within the framework of the industrial chair IDAML [3] of ENS Paris-Saclay, a library of transfer learning methods, ADAPT [4], has been developed to facilitate the access of these techniques to the industrial sector.
Based on the library already developed, the project will consist in studying the dependence of the different transfer learning methods on their main hypotheses. This will allow to identify in which context each method can be efficiently applied and which type of method is to be preferred according to the nature of the problem. This work will thus help industrials to more easily identify which tool is best suited to their problems.

[1] https://www.ensiie.fr/wp- content/uploads/2020/10/PosterLudovicMinvielle-1.pdf
[2] Yishay Mansour, Mehryar Mohri, and Afshin Rostamizadeh. “Domain adaptation: Learning bounds and algorithms. In COLT, 2009.
[3] https://www.centreborelli.fr/partenariats/chaires/chaires-industrielles-2/
[4] https://github.com/adapt-python/adapt