A simple introduction on Sinkhorn distances | by Jianfeng Wang | Medium
The Sinkhorn distance [1] was proposed in 2013, and the work was accepted in NIPS conference. The goal is to improve the optimization speed to solve the optimal transportation problem. Thus, there are several concepts. 1) transportation problem, 2) optimal transportation problem, 3) optimize it, 4) optimization speed improvement. The concept may be kind of new, but I will try to explain it step by step. Let’s imagine we have some mines, which are able to produce the iron ores. The iron ore will be used by some factories for further processing. We don’t need to know how they process the iron ore, or why they want to process that. Instead, we mainly focus on the amount. That is, each mine can produce different amounts of iron ore and each factory may also require different amount of iron ore. Let’s assume, we have 2 mines: A and B. We have 3 factories to use the iron ore: X, Y and Z. A can produce 3 iron ores while B can produce 7 iron ores. The unit could be tonnage, or kiloton. We don’
A simple introduction on Sinkhorn distances Jianfeng Wang 8 min read · Dec 19, 2020 -- 2 Listen Share The Sinkhorn distance [1] was proposed in 2013, and the work was accepted in NIPS conference. The goal is to improve the optimization speed to solve the optimal transportation problem. Thus, there are several concepts. 1) transportation problem, 2) optimal transportation problem, 3) optimize it, 4) optimization speed improvement. The concept may be kind of new, but I will try to explain it step by step. Transportation problem Let’s imagine we have some mines, which are able to produce the iron
Explore this link on the map →related reading
- Quick start guide - POT Python Optimal Transport 0.9.5 documentationpythonot.github.io
- Your Transformer is Secretly an EOT Solver | Elements of a Vector Spaceelonlit.com
- Wasserstein metric - Wikipediaen.wikipedia.org
- Sinkhorn's theorem - Wikipediaen.wikipedia.org
- Six (and a half) intuitions for KL divergence — LessWronglesswrong.com
- Kullback–Leibler divergence - Wikipediaen.wikipedia.org
- Visual Information Theory -- colah's blogcolah.github.io
- bv_cvxbook.pdfweb.stanford.edu
- [2405.15441] Statistical and Computational Guarantees of Kernel Max-Sliced Wasserstein Distancesarxiv.org
- [2505.06589v1] Optimal Transport for Machine Learnersarxiv.org
- bv_cvxbook.pdfstanford.edu
- [1802.08665] Learning Latent Permutations with Gumbel-Sinkhorn Networksarxiv-vanity.com