Prior Distribution and Model Confidence

Maksim Kazanskii, Artem Kasianov

We study how the training data distribution affects confidence and performance in image classification models. We introduce Embedding Density, a model-agnostic framework that estim…