November 29th

14:30, room 2014 (Shannon 660) (see location):


Amaury Habrard (Universite Jean Monnet de Saint-Etienne)


Title: Domain Adaptation with Optimal Transport: Mapping Estimation and Theory


Abstract:

Domain adaptation (DA) is an important topic in machine learning that
occurs when the distribution of training data (source domain) is
different but related to the distribution of test data (target
domain). A classic solution for this problem is to try to reduce the
divergence between the two distributions while ensuring good performance
on the training source data. Recently, a new approach based on Optimal
Transport (OT) has been proposed with the appealing idea to minimize the
Wasserstein distance between source and target samples in order to move
the source close to the target. However, this approach has the drawback
to limit the transport map to the samples used during the learning phase.
In this talk, I will start by an introduction to
Domain Adaptation and Optimal Transport. Then, I will first present a
method to tackle the drawback mentioned before where the idea is to
learn a mapping of the data jointly with the coupling computed by
Optimal Transport. This allows to smooth the result of Optimal
Transport and to "generalize" it to out-of-samples instances. Second,
I will discuss some theoretical perspectives for Domain Adaptation based
on Optimal Transport with a particular focus on the multiview setting.


Contact: guillaume.charpiat at inria.fr