1: \begin{abstract}

2:   We study the use of amortized optimization to predict optimal transport (OT) maps

3:   from the input measures, which we call \emph{Meta OT}.

4:   This helps repeatedly solve similar OT problems between different

5:   measures by leveraging the knowledge and information present

6:   from past problems to rapidly predict and solve new problems.

7:   Otherwise, standard methods ignore the knowledge of the

8:   past solutions and suboptimally re-solve each problem from scratch.

9:   Meta OT models surpass the standard

10:   convergence rates of log-Sinkhorn solvers in the discrete setting

11:   and convex potentials in the continuous setting.

12:   We improve the computational time of standard OT solvers

13:   by multiple orders of magnitude in discrete and continuous

14:   transport settings between images, spherical data, and color palettes.

15:   Our source code is available at \url{http://github.com/facebookresearch/meta-ot}.

16: \end{abstract}

17: