Open Problem: How fast can a multiclass test set be overfit?

摘要

We ask how many measurements of the accuracy on a multiclass benchmark are needed to achieve a given amount of overfitting.