Tag
This article discusses a research paper showing that the disagreement rate between two deep networks trained with different random seeds can accurately estimate generalization error using only unlabeled data, revealing a surprising connection called Generalization Disagreement Equality.
A reflection arguing that in multi-model setups, the consensus output is less valuable than the disagreements, which reveal genuinely contested parts of a problem. The post questions whether consensus should be the goal and how to distinguish productive disagreement from noise.