Compositionality and Generalization in Emergent Communication Using Metropolis-Hastings Naming Game

This study investigates the emergence of compositionality and generalization within Emergent Communication (EmCom) systems, focusing on emergent language using the Metropolis-Hastings naming game (MHNG). Although the MHNG has been used in previous EmCom research, this is the first study to explore compositionality, the ability to construct complex expressions by combining simpler elements, and generalization, the ability to apply learned patterns to new situations, in emergent language within this language game. We introduce the novel Inter-VAE+VAE model, which equips each agent with dual variational autoencoders (VAEs), specifically designed for cognitive tasks that mirror human perception and language processing. This model, facilitating predictive coding, allows agents to refine their world models through collective experiences, a process rooted in the collective predictive coding hypothesis and differing from isolated learning approaches. Our model was evaluated against baseline models, including the $\beta- \text{VAE}$ and $\beta-\text{TCVAE}$, and was further compared with implementations of the Lewis signaling game using the dSprites and 3Dshapes datasets. The results from these evaluations indicate an improvement in agent communication within the MH naming game and underscore the model's potential in replicating key aspects of human language, particularly compositionality and generalization, in artificial systems.