Structural Inductive Biases in Emergent Communication
This addresses the problem of improving communication and generalization in AI agents, though it is incremental as it builds on existing referential games with a new model type.
The paper tackled the problem of enabling artificial agents to develop compositional language for systematic generalization by investigating representation learning in graph referential games. The result showed that agents using graph neural networks developed more compositional language compared to bag-of-words and sequence models, allowing them to generalize to new combinations of familiar features.
In order to communicate, humans flatten a complex representation of ideas and their attributes into a single word or a sentence. We investigate the impact of representation learning in artificial agents by developing graph referential games. We empirically show that agents parametrized by graph neural networks develop a more compositional language compared to bag-of-words and sequence models, which allows them to systematically generalize to new combinations of familiar features.