[Paper Review] Deep Learning and the Global Workspace Theory

Rufin VanRullen, Ryota Kanai|arXiv (Cornell University)|Dec 4, 2020

Neural dynamics and brain function73 references4 citations

TL;DR

This paper proposes a deep learning implementation of the Global Workspace Theory (GWT) by creating a unified, amodal Global Latent Workspace (GLW) through unsupervised neural translation between multiple specialized neural networks. By enabling cross-modal information sharing via shared latent spaces, the framework supports higher-level cognition, attention, and consciousness-like integration, offering a brain-inspired AI architecture with testable neuroscientific predictions.

ABSTRACT

Recent advances in deep learning have allowed Artificial Intelligence (AI) to reach near human-level performance in many sensory, perceptual, linguistic or cognitive tasks. There is a growing need, however, for novel, brain-inspired cognitive architectures. The Global Workspace theory refers to a large-scale system integrating and distributing information among networks of specialized modules to create higher-level forms of cognition and awareness. We argue that the time is ripe to consider explicit implementations of this theory using deep learning techniques. We propose a roadmap based on unsupervised neural translation between multiple latent spaces (neural networks trained for distinct tasks, on distinct sensory inputs and/or modalities) to create a unique, amodal global latent workspace (GLW). Potential functional advantages of GLW are reviewed, along with neuroscientific implications.

Motivation & Objective

• To develop a brain-inspired cognitive architecture for artificial intelligence based on the Global Workspace Theory (GWT).
• To enable cross-modal integration of information from specialized neural networks through a shared, amodal latent space.
• To provide a practical, implementable roadmap for constructing a Global Latent Workspace (GLW) using existing deep learning components.
• To bridge artificial intelligence and neuroscience by aligning AI mechanisms with neuroscientific principles of consciousness and attention.
• To generate testable predictions for neuroscientific research on information broadcasting and neural ignition.

Proposed method

• Utilizes multiple specialized deep neural networks, each trained for distinct sensory or cognitive tasks (e.g., vision, NLP, motor control).
• Employs unsupervised neural translation between the latent spaces of these specialized modules to enable cross-modal information exchange.
• Constructs a Global Latent Workspace (GLW) as a shared, amodal representation space where information from diverse modalities is integrated.
• Leverages techniques such as cycle-consistent adversarial networks, dual learning, and domain adaptation to enable translation between non-parallel, non-overlapping latent spaces.
• Implements attention mechanisms to select and broadcast salient information to the GLW, mimicking top-down and bottom-up attention in the brain.
• Relies on long-range recurrent connections and ignition dynamics to simulate the all-or-none broadcast of information, as seen in the Global Neuronal Workspace (GNW) model.

Experimental results

Research questions

• RQ1 How can deep learning architectures implement the core mechanisms of the Global Workspace Theory (GWT) in artificial systems?
• RQ2 What role does an amodal, shared latent space play in enabling cross-modal integration and higher-level cognition?
• RQ3 How can unsupervised neural translation between distinct latent spaces support the broadcast of information across specialized modules?
• RQ4 What neuroscientific signatures (e.g., neural ignition) can be replicated or predicted by such a deep learning-based GLW framework?
• RQ5 How might this architecture support attention, executive function, and conscious-like awareness in artificial systems?

Key findings

• The proposed Global Latent Workspace (GLW) enables integration of information across multiple specialized neural networks through unsupervised translation between their latent spaces.
• The framework supports higher-level cognitive functions such as problem-solving and coordinated decision-making by enabling shared access to information across modalities.
• Neural ignition dynamics—characterized by all-or-none activation of broad brain-like networks—can be simulated in the GLW via recurrent connections and attention-based broadcasting.
• The model provides a testable framework for neuroscience, predicting that global information broadcasting correlates with sustained, widespread neural activation across prefrontal, parieto-temporal, and cingulate cortices.
• Existing deep learning components (e.g., cycle-GANs, dual learning, domain adaptation) are sufficient to implement the core mechanisms of the GLW, suggesting feasibility with current technology.
• The approach offers a path toward artificial systems with consciousness-like properties by emulating the brain’s mechanism of global information availability through a shared workspace.