Global Workspace Theory — Attention Competition
Five cognitive modules compete for a shared broadcast channel via softmax selection (Baars 1988). The winner’s signal is broadcast to all modules as top-down attention.
Competing Modules
- Sensory — input strength
- Motor — output magnitude
- Predictive — world model confidence (1 – PE)
- Error — PE saliency (PE × 5)
- Memory — episodic recall strength
Emotions modulate salience: fear boosts error module, positive valence boosts motor. Winner signal injected into SNN hidden neurons at 10% strength.
References
- Baars (1988). A Cognitive Theory of Consciousness. Cambridge University Press
- Dehaene & Naccache (2001). Towards a cognitive neuroscience of consciousness. Cognition
API Reference
GlobalWorkspaceBridge(config: GWTBridgeConfig)
step(module_inputs: dict) → dict
One competition cycle. Returns broadcast_signal, winning_module, salience_scores, attention_map.
override_attention(module_name, duration=10)
Force attention to a module (voluntary control).
GWTModule(name, n_neurons, device)
compute_salience(input_signal) → float
0.7 × strength + 0.3 × novelty (change rate).
GWTBridgeConfig
broadcast_strength: 0.3 competition: 'softmax' broadcast_duration: 10 attention_decay: 0.95