Abstract
Artificial agents that need to exchange information to accomplish some shared goal develop emergent communication protocols. These protocols comprise opaque discrete sequences that are not human interpretable. In this sense, they share some common properties with animal communication. This talk will review some of our recent work on understanding emergent communication of artificial agents. First, I’ll present a new metric for assessing the compositionally of an emergent protocol, by forming a mapping between emergent symbols and natural language concepts. I’ll then describe a new communication channel, which integrates a discrete codebook, together with a multi-stage training process for encouraging compositional generalization. Then, I’ll describe a theoretical result on how different environments constrain the semantic consistency of optimal emergent communication protocols. We’ll conclude with an approach for unsupervised translation of emergent messages to English, which shows promising results of deciphering opaque protocols without any supervision in the form of English translations.