[3L3-GS-8] Robot and real worlds: symbol emergence / knowledge extraction

[3L3-GS-8-02] Symbol emergence using Variational autoencoder

〇Yuto Yoshida¹, Yoshinobu Hagiwara¹, Akira Taniguchi¹, Tadahiro Taniguchi¹ (1. Ritsumeikan University)

Keywords:Symbol emergence in robotics, Deep generative model, Multi-agent

In this paper, we propose a computational model that realizes symbol emergence between two agents observing real images using Variational autoencoders(VAEs).
When humans communicate information with others, they communicate using signs such as words and signals.
This study is conducted to build computational models that reproduce the emergence of symbolic communication between humans to obtain better understanding.
In this study, we used VAE to model the representation of symbol emergence between two agents that perform category formation from real images. Using the SERKET framework, the representation learning is effectively influenced by the symbol emergence at the social level.
The results of experiments demonstrated that categories are formed from real images observed by agents, and signs are shared appropriately among agents through symbolic communication.
In addition, the images recalled by the agents confirmed that the objects in the recalled images were shared among the agents.