Abstract
The principle of compositionality, which enables natural language torepresent complex concepts via a structured combination of simpler ones, allowsus to convey an open-ended set of messages using a limited vocabulary. Ifcompositionality is indeed a natural property of language, we may expect it toappear in communication protocols that are created by neural agents in languagegames. In this paper, we propose an effective neural iterated learning (NIL)algorithm that, when applied to interacting neural agents, facilitates theemergence of a more structured type of language. Indeed, these languagesprovide learning speed advantages to neural agents during training, which canbe incrementally amplified via NIL. We provide a probabilistic model of NIL andan explanation of why the advantage of compositional language exist. Ourexperiments confirm our analysis, and also demonstrate that the emergedlanguages largely improve the generalizing power of the neural agentcommunication.