Optimal transport is a notoriously computationally difficult tool that deals with probability
distributions. The main approaches (mostly relying on convolutions) still require significant
time which makes optimal transport intractable for large scale machine learning problems,
where it has most expected impact.
We propose to tackle this speed issue in the opposite way: trying to develop machine
learning tools to speed up optimal transport computations -- that, in turn, will permit to
develop time-efficient machine learning applications.