Curriculum Learning as Transport: Understanding Curricula with Wasserstein Geodesics
The paper reframes curriculum learning as movement between training distributions, then tests which parts of that movement matter.
Shin and Alvarez-Melis model curricula as Wasserstein paths over discrete difficulty levels, separating ordering, exposure, smoothness, and pacing. In 12 synthetic tasks across 33 difficulty axes, they find no curriculum strategy wins broadly. Easy-to-hard ordering can improve hard-level performance even when total exposure is matched, so exposure alone does not explain the effect. The paper also reports that endpoint smoothness and pacing change where a curriculum helps along the difficulty spectrum. ArXiv · AI/CL/LG's note
Shin and Alvarez-Melis model curricula as Wasserstein paths over discrete difficulty levels, separating ordering, exposure, smoothness, and pacing. In 12 synthetic tasks across 33 difficulty axes, they find no curriculum strategy wins broadly. Easy-to-hard ordering can improve hard-level performance even when total exposure is matched, so exposure alone does not explain the effect. The paper also reports that endpoint smoothness and pacing change where a curriculum helps along the difficulty spectrum. ArXiv · AI/CL/LG's note
score 4