Multimodal Continuous Symbol Emergence Using a Probabilistic Generative Model Based on Gaussian Processes

This study proposes a novel probabilistic generative model based on the Gaussian process that enables symbols to emerge from two agents. In this model, agents can create symbols and share their meaning with other agents in a bottom-up manner by interacting with each other and with the environment. The symbols are formulated as shared continuous latent variables from which two agents' observations are generated according to the Gaussian process; therefore, symbol emergence is considered equivalent to an inference of the latent variables. For the inference, we used the Metropolis–Hastings (MH) naming game, in which shared latent variables can be inferred without directly observing another variable's internal state. Hence, symbols can emerge while maintaining agent independence. In the experiment, we used a multimodal object dataset, and examined the emergence of symbols representing the objects. To confirm that the symbols and meanings were appropriately shared among the agents, we evaluated whether agents could select an object represented by a symbol yielded by another agent. The results show that the agent could select an appropriate object using the symbols that emerged.