Random Fourier Features for PAC-Bayesian Domain Adaptation

During this internship, supervised by Guillaume Metzler, Emilie Morvant and Marie-Ange Lèbre, we worked on the PAC-Bayesian theory and its specification for Domain Adaptation. Centering on kernel approximation methods and data representation, we took the standpoint of Random Fourier Features. Using Random Fourier Features, we found a new way to look at the pseudo-kernel by develloping a novel PAC-Bayesian bound on the quality of the representation in the target domain. from this bound we derived a learning algorithm. This proposition showed promising results when benchmarked against other PAC-Bayesian techniques for Domain Adaptation on Toy datasets.

The report can be found here and the git-hub project repository is available here.