Continuous Symbol Emergence Using Variational Inference with Multi-Agent

Abstract

In recent years, research on emergent communication has gained significant attention to elucidate the phenomenon of symbol (language) emergence through a constructive approach. Especially, the collective predictive coding hypothesis has been proposed to model the phenomenon of symbol emergence through decentralized Bayesian inference of latent variables shared among multiple agents. As a method for inferring symbols, the Metropolis-Hastings Naming Game (MHNG) has been proposed, where multiple agents observe the same target and represent their observations using a shared name. Numerous extended methods for MHNG have been developed, but these methods suffer from low learning efficiency due to their use of the Metropolis-Hastings algorithm. To overcome this problem, the Variational Bayesian Naming Game (VBNG), which uses variational Bayesian methods, was proposed to enhance learning efficiency. However, VBNG is currently applied only to the emergence of discrete symbols, which limits its expressive power. In this paper, we propose a novel method that integrates Gaussian processes into VBNG to enable the efficient estimation of continuous symbols. In this experiment, the proposed method enables decentralized Bayesian inference through a naming game task with multiple agents.