This paper investigates feature transfer in classifier-trained networks, analyzing the impact of similarity between training and unseen data on unseen data clustering performance and feature extraction.

We show how to learn representations of temporal distances that exploit quasimetric architectures in offline GCRL.