Prof. Cuturi's research lies at the interface between machine learning, deep learning and optimal transport theory.

A fundamental property of deep learning frameworks is that they rely on automatically differentiable operators. I will show in this talk how a fundamental building block of CS, ranking, can be adapted to "fit" this computation model. To do so, we leverage the fact that sorting can be seen as a particular instance of the optimal transport (OT) problem. The first part of the talk will show how one can use a mathematical hammer (OT) to hit a fly. I will then show how this hammer can be made considerably more efficient and differentiable using regularization, to propose drop-in replacements to usual sorting algorithms that are automatically differentiable.