A Paper by Araya’s Reinforcement Learning Team has been Published in the Journal “Frontiers in Computational Neuroscience”
On June 14, 2024, a paper titled “Design and evaluation of a global workspace agent embodied in a realistic multimodal environment” (DOI: https://doi.org/10.3389/fncom.2024.1352685) first-authored by Dr. Rousslan Dossa, Senior Researcher at the Reinforcement Learning Team (Team Lead: Dr. Kai Arulkumaran) of the Research and Development Department of Araya Inc., was published in the journal “Frontiers in Computational Neuroscience”.
Background
To understand the essence of consciousness, deeper investigations of its underlying mechanisms are required. This involves formulating hypotheses about how consciousness works and conducting experiments to test them. While some theories of consciousness can be implemented in artificial agents, they are typically tested in simplified settings.
Summary
Researchers at Araya Inc., Tokyo, Japan, in collaboration with Microsoft Research, New York, United States, have developed an artificial agent that fulfills criteria for the Global Workspace Theory (*1) and tested the agent on a navigation task within a photorealistic 3D simulator, where the agent has to follow sounds to the goal location. The results showed that, at smaller working memory sizes, the Global Workspace agent performed better than standard agents without the Global Workspace module, while these performance differences disappeared at larger memory sizes.
The agent navigates to the target in the room as shown in the map (right) based on its memory of a brief audio signal (left) and visual observations (middle).
This research underscores the importance of testing consciousness theories in more realistic settings. The successful implementation of Global Workspace Theory in a complex environment marks a significant step forward in understanding consciousness through artificial agents.
The researchers have open-sourced their code to encourage further exploration and development in this field. The code can be accessed at GitHub.
Future Prospects
This study provides evidence of the potential benefits of agents implementing the Global Workspace Theory. Additionally, it has opened new avenues for how theories of consciousness can be studied in artificial agents.
Glossary
*1) Global Workspace Theory: one of the leading theoretical frameworks for understanding how consciousness operates in the human brain.
Related Links
Reinforcement Learning Team – Team Page
Design and evaluation of a global workspace agent embodied in a realistic multimodal environment (Frontiers in Computational Neuroscience)
GitHub – multimodal-global-workspace-agent
