Abstract by Connor Christopherson
Translation Between Emergent Communication Protocols
When multiple agents controlled by DNNs are given mutual goals, they can learn to cooperate. When both agents write/read from shared memory, these agents can develop a language of their own which is not interpretable by humans (referred to as emergent communication protocol). Unfortunately, if a new group of agents is similarly trained, the new communication protocol generated by this group would not be compatible with the protocol of the first group, even if both groups have learned the same tasks.
We propose the use of Disco GANs, which have previously been used to create joint latent spaces for images, to translate between different communication protocols. If such translation is possible, it is conceivable that we could extend this towards translation between human text and learned emergent communication protocols. This would have applications in military (intercompatibility between drone swarms), industry (improved natural language translation), and research (understanding increasingly complex machine learning algorithms).