Collective predictive coding hypothesis

Toward new development of symbol emergence in robotics and language evolution

Abstract

This paper introduces a new hypothesis about symbol emergence systems, called the collective predictive coding hypothesis. The hypothesis is based on a language game in which multiple agents coordinate to develop a symbol system (language) through language games while observing sensorimotor information. In particular, an approach that models symbol emergence (or emergent communication) as social representation learning. The computational model based on such an approach leads us to the hypothesis of collective predictive coding. Its generality and extendibility are also discussed. Finally, the implications of the hypothesis for symbol emergence in robotics and language evolution are discussed.