An embodied agent with a structure based on global workspace theory, trained on realistic audiovisual inputs to navigate 3D environments, performs better and more robustly at smaller working memory sizes compared to a standard recurrent architecture. Task complexity and regularization are essential for feature learning and the development of meaningful attentional patterns within the workspace.
The authors examine three contemporary theories of conscious function—Global Workspace Theory, Information Generation Theory, and Attention Schema Theory—and find that each relates conscious function to some aspect of domain-general intelligence in humans. They then observe that state-of-the-art deep learning methods have begun incorporating key aspects of these theories, though they remain far from demonstrating general intelligence. Using mental time travel in humans as a motivating example, the authors propose combining insights from all three theories into a single unified model. Such artificial agents would possess greater general intelligence and align more closely with current understanding of consciousness's functional role, making this a promising near-term AI research goal.