Abstract
There is renewed interest in simulating language emergence among deep neuralagents that communicate to jointly solve a task, spurred by the practical aimto develop language-enabled interactive AIs, as well as by theoreticalquestions about the evolution of human language. However, optimizing deeparchitectures connected by a discrete communication channel (such as that inwhich language emerges) is technically challenging. We introduce EGG, a toolkitthat greatly simplifies the implementation of emergent-language communicationgames. EGG's modular design provides a set of building blocks that the user cancombine to create new games, easily navigating the optimization andarchitecture space. We hope that the tool will lower the technical barrier, andencourage researchers from various backgrounds to do original work in thisexciting area.