Apr 7, 2022   11:00 AM — 01:00 PM

Kilian Fatras MILA, McGill University, Canada

For more information, visit: https://cerc-datascience.polymtl.ca/coffee/#

Kilian Fatras

Hybrid seminar on Zoom and in the GERAD seminar room.

Optimal transport distances have found many applications in machine learning for their capacity to compare non-parametric probability distributions. Yet their algorithmic complexity generally prevents their direct use on large scale datasets. Among the possible strategies to alleviate this issue, practitioners can rely on computing estimates of these distances over minibatches of data. While computationally appealing, we highlight in this talk some limits of this strategy, arguing it can lead to undesirable smoothing effects. As an alternative, we suggest that the same minibatch strategy coupled with unbalanced optimal transport can yield more robust behaviours. We discuss the associated theoretical properties, such as unbiased estimators, existence of gradients and concentration bounds. Our experimental study shows that in challenging problems associated to domain adaptation, the use of unbalanced optimal transport leads to significantly better results, competing with or surpassing recent baselines.

Gabriele Dragotto organizer
Federico Bobbio organizer


Room 4488
André-Aisenstadt Building
Université de Montréal Campus
2920, chemin de la Tour
Montréal QC H3T 1J4

Associated organizations