Symbol Emergence and Representation Learning by Integrating GPLVM and Neural Network

Abstract

In this paper, we propose a novel model for multi-agent symbol emergence that integrates Gaussian Process Latent Variable Model (GPLVM) and neural networks, in which shared symbols between two agents can emerge. In the proposed model, the agents create symbols bottom-up by interacting with the environment and share their meanings by interacting with others. Using GPLVM, the symbols are represented as continuous variables that are more expressive than discrete variables. For the inference of symbols in GPLVM, we utilize Metropolis-Hastings Naming Game (MHNG). MHNG is a method that enables agents to acquire shared symbols and their meanings by communicating the symbols between them without directly observing the other's internal states. Furthermore, we introduce the neural network called Neural Conversion Adapter (NCA) that converts the features extracted from observation to low dimensional latent variables that are internal states of each agent. NCA facilitates the inference of appropriate latent variables for representing symbols. Experiments show that the GPLVM-based symbol emergence model can generate shared symbols and that more explicit symbols can be learned by combining NCA with GPLVM.