Paper

We study the variation in ImageNet and mistake types across 2,200 vision models.