Skip to content

Global Neuronal Workspace (GNW)

Introduced in A neuronal model of a global workspace in effortful cognitive tasks (Dehaene, Kerszberg & Changeux, 1998).

The Global Neuronal Workspace (GNW) is a model of the brain's large-scale architecture. It explains how information processed in specialized brain areas becomes globally available for effortful cognitive tasks, and it is one of the leading theories of conscious access.

Architecture

The GNW distinguishes two computational spaces:

  • Specialized processors: a network of modular, mostly local circuits for perception, motor control, long-term memory, evaluation and attention. They process their inputs automatically and in parallel, without requiring the workspace.
  • Global workspace: a distributed set of neurons with long-range excitatory connections, mostly pyramidal neurons in prefrontal, parietal and cingulate cortex. They interconnect the specialized processors and allow information to be shared between them.

The resulting topology resembles a hub-and-periphery network: sensory inputs are processed by peripheral modules and integrated in a densely connected central core.

Dynamics

When a stimulus is strong or attended enough, workspace neurons are mobilized in a sudden, self-sustained and nonlinear activation called ignition. The information is then broadcast to many processors at once and can be maintained, compared and used for routine-breaking, effortful tasks. At the same time, the workspace amplifies the selected processors top-down through excitatory and inhibitory feedback loops, while competing representations are suppressed. Routine tasks can be solved by the specialized processors alone, whereas novel or conflicting tasks require the workspace.

Relevance for CHAIS

The GNW topology is one of the two neural architecture directions of the CHAIS project, next to oscillations. Modality-specific inputs (audio, gaze, video) could be processed in peripheral modules and fused in a central workspace, with network density, interlinking and hierarchical modularity as parameters to vary. Related work, including the spiking BOSSA model, is collected in Existing Research.