Machine Learning
Cortex
Computational architecture organized around selective local processing, bounded global integration, persistent hypotheses, and learned gating.
Research question
Can useful computation be organized around selective local processing, bounded global integration, persistent hypotheses, and learned gating rather than applying equivalent global computation at every stage?
Method & architecture
Cortex decouples routine fast-path sensory reflex from deliberate global cognitive workspace integration. Inputs are filtered by local processors; only significant or anomalous signals trigger gated recruitment of bounded global hypothesis slots that compete for final output synthesis.
- Selective Local Processing: Low-entropy signals are handled by local reflex paths without engaging global state.
- Bounded Global Workspace: A finite set of hypothesis slots (H1, H2, H3) prevents combinatorial state explosion.
- Falsifiability Invariant: The architecture remains an empirical hypothesis until controlled experiments demonstrate an advantage against appropriately matched baselines.
Results & findings
Gating Hysteresis Requirement
OBSERVEDSimple threshold gating produces thrashing when input ambiguity hovers near decision boundaries.
Advantage Against Matched Baselines
UNRESOLVEDThe architecture remains a hypothesis until controlled experiments demonstrate an advantage against appropriately matched baselines.