Viet-Anh on Software Logo

What is: Distributional Generalization?

SourceDistributional Generalization: A New Kind of Generalization
Data SourceCC BY-SA -

Distributional Generalization is a type of generalization that roughly states that outputs of a classifier at train and test time are close as distributions, as opposed to close in just their average error. This behavior is not captured by classical generalization, which would only consider the average error and not the distribution of errors over the input domain.