NeurIPS 2021 – 10 Papers You Shouldn't Miss

Authors' TL;DR We use self-supervised play to train artificial agents to communicate by drawing and then show that with the appropriate inductive bias a human can successfully play the same games with the pretrained drawing agent.