Keynotes | K&Ts | GACs | Talks | Posters | Search

Poster B in Poster Session B: Tuesday, August 4, 2:00 – 3:45 pm, Kimmel Center, Shorin & Rosenthal Rooms

What Does the Global Latent Workspace Learn?

Yu-Ang Cheng¹, Sixuan Chen¹, Zhouyang Lu¹, Xizheng Yu¹, Grégoire Dhimoïla¹, Thomas Serre¹; ¹Brown University

Presenter: Sixuan Chen

Cognition requires more than parallel processing of separate sensory streams. It demands shared representations that can support integration and flexible reuse across perception, memory, and action. Global Workspace Theory offers a framework for how this is achieved: specialized processors operate in parallel while a shared workspace broadcasts selected information across brain areas. Recent deep learning architectures make this idea concrete by connecting modality-specific modules through a common latent space. Yet, it remains unclear what structure these training procedures actually produce and when the resulting workspace is genuinely multimodal rather than merely compressive. We develop a theoretical framework in a simplified linear setting, a regime where the problem admits closed-form solutions. We show that the nature of the learned workspace depends critically on what it is trained to do. If each module only reconstructs its own input, the workspace compresses each modality independently. If modules translate into each other's format, the workspace aligns with the statistical structure common to both modalities. This suggests that a workspace becomes integrative only when training pressures translation across modalities, not compression within each modality alone.

Topic Area: Methods, Tools, Theory & Neural Coding