Structural Inductive Biases in Emergent Communication
Abstract
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.
- Publication:
-
arXiv e-prints
- Pub Date:
- February 2020
- DOI:
- 10.48550/arXiv.2002.01335
- arXiv:
- arXiv:2002.01335
- Bibcode:
- 2020arXiv200201335S
- Keywords:
-
- Computer Science - Computation and Language;
- Computer Science - Artificial Intelligence;
- Computer Science - Machine Learning;
- Computer Science - Multiagent Systems;
- Statistics - Machine Learning
- E-Print:
- The first two authors contributed equally. Poster presented at CogSci 2021