[A-10] システム数理と応用

[A-10-15] Basis Optimal Transport with GPU-parallelizable Sinkhorn's Solver for Size-variable Distributions

^○Jianming Huang, Hiroyuki Kasai (Waseda Univ.)

Keywords:Optimal transport、GPU parallelization、Size-variable distribution

The Sinkhorn’s solver provides an effective way to compute optimal transport (OT) problem with GPU parallelization, which greatly accelerates the solution. However, the application of GPU parallelization for Sinkhorn’s solver has a strict condition: all pairs of distribution should have a fixed size. Many real-world datasets such as graph datasets usually have variable sizes, which are not able to benefit from GPU parallelization. To overcome this difficulty, we propose a method to translate the problem of size-variable distributions into size-fixed ones, named Basis OT (BOT).