Mechanising Conceptual Spaces using Variational Autoencoders

In this pilot study, we explore the Variational Autoencoder as a computational model for conceptual spaces in a social interaction context. Conceptually, the Variational Autoencoder is a natural fit for this purpose. We apply this idea in an agent-based social creativity simulation to explore and understand

the effects of social interactions on adapting conceptual spaces. We demonstrate a simple simulation setup and run experiments with a focus on establishing a baseline. While ongoing work needs to identify if adaption was appropriate, the results so far suggest that the Variational Autoencoder appears to adapt to new artefacts and has potential for modelling conceptual spaces.