Selective Routing

Gating and control of information flow.

Executive Summary

Across biology and AI, selective routing is implemented by mechanisms that gate and direct limited-capacity processing and broadcast to consumer systems. Attention—especially when supported by an internal model of its own state—serves as the principal control policy for routing, while oscillatory and field-mediated processes tune effective connectivity. The computational benefits include efficiency, robustness under missing inputs, and function-specific specialization that aligns processing with goals.

Papers

Evidence

Confidence

Key Insights

Unified Insights

Selective routing is organized around a limited-capacity bottleneck/workspace that gates access and enables global broadcast to many modules.

Supporting Evidence (6)

Taking AI Welfare Seriously : Identifies GW indicators: parallel modules, limited-capacity workspace with selective attention, and global broadcast.

In Search of a Biological Crux for AI Consciousness : Highlights GWT/gNW: consciousness involves being in a global workspace and broadcasting to consumer systems.

Consciousness without a cerebral cortex: A challenge for neuroscience and medicine : Proposes a mesodiencephalic bottleneck gating which information reaches awareness and long-term storage.

Concepts of Consciousness : Defines access as content reaching an Executive system that controls reasoning and action, implying gating to a central hub.

Don’t forget the boundary problem! How EM field topology can address the overlooked cousin to the binding problem for consciousness : Theorizes a well-bounded 4D EM pocket selectively incorporating competing modules’ activity, echoing workspace-like gating.

Cytoelectric coupling: Electric fields sculpt neural activity and “tune” the brain’s infrastructure : Links oscillatory coherence/LFP dynamics to attentional selection and information routing, a mechanism for gating.

Contradictory Evidence (1)

Shared functional specialization in transformer-based language models and the human brain : Demonstrates distributed, headwise specialization in transformers, suggesting effective routing without a single centralized bottleneck.

An internal model of attention (attention schema) stabilizes and improves top-down control of selective routing; without it, attention becomes less controllable and less adaptive.

Supporting Evidence (4)

The Attention Schema Theory: A Foundation for Engineering Artificial Consciousness : Argues an attention schema aids control of attention by representing its properties, state, and consequences.

The attention schema theory in a neural network agent: Controlling visuospatial attention using a descriptive model of attention : Ablating the attention schema in a DQN agent severely impaired selective control and task performance.

The attention schema theory in a neural network agent: Controlling visuospatial attention using a descriptive model of attention : Human behavioral evidence: without conscious awareness, exogenous attention remains but endogenous control breaks down.

Conscious artificial intelligence and biological naturalism : Highlights AST among theories tying consciousness to attention control and global information sharing.

Contradictory Evidence (3)

The attention schema theory in a neural network agent: Controlling visuospatial attention using a descriptive model of attention : Shows that exogenous attention and some routing occur without awareness/attention schema, though control is diminished.

Interpreting and improving natural language processing in machines with natural language processing in the brain : Replacing learned attention with uniform attention benefited shallow layers, suggesting that detailed, learned control is not universally beneficial at all processing stages.

Shared functional specialization in transformer-based language models and the human brain : Transformers display effective content-sensitive routing via learned heads without an explicit attention schema.

Oscillatory coherence, ephaptic coupling, and thalamic mechanisms gate inter-areal communication, providing biophysical control over routing and decision variables.

Supporting Evidence (5)

Cytoelectric coupling: Electric fields sculpt neural activity and “tune” the brain’s infrastructure : Reviews evidence that oscillatory coherence mediates attention and routing of task-relevant information.

In vivo ephaptic coupling allows memory network formation : Shows field-to-activity Granger causality exceeding activity-to-field, with directed FEF→SEF influence during memory—consistent with field-mediated routing.

A Roadmap for Using tFUS to Investigate the Neural Substrate of Conscious Perception : Proposes causal thalamic gating tests; reports pulvinar’s effect on decision confidence and duty-cycle-dependent changes in sensitivity/criteria.

Event-related delta, theta, alpha and gamma correlates to auditory oddball processing during Vipassana meditation : Meditation altered distractor gating: decreased early delta (2–4 Hz, ~100–500 ms) to distractors and increased late alpha desynchronization (~500–900 ms).

Consciousness without a cerebral cortex: A challenge for neuroscience and medicine : Argues cortico–mesencephalic thresholding and thalamic relays gate what reaches awareness—a subcortical control point for routing.

Contradictory Evidence (1)

Don’t forget the boundary problem! How EM field topology can address the overlooked cousin to the binding problem for consciousness : Provides a speculative EM-topological selection mechanism; while compatible with field effects, it is not yet empirically established.

Function-specific specialization and learned attention policies provide computational benefits: efficiency, robustness under missing inputs, and alignment with task-relevant structure.

Supporting Evidence (6)

Shared functional specialization in transformer-based language models and the human brain : Shows emergent headwise functional specialization—content-sensitive gating by attention heads.

Shared functional specialization in transformer-based language models and the human brain : Headwise features map to cortical ROIs and syntactic dependencies; shuffling head features abolishes structure, indicating necessity of specialized routing.

TRIBE: TRImodal Brain Encoder for whole-brain fMRI response prediction : Modality dropout trains flexible routing across modalities, avoiding over-reliance and enabling predictions when inputs are missing.

A composite and multidimensional heuristic model of consciousness : A meta-controller arbitrates between model-based and model-free routes to minimize cost, a form of routing optimization.

Interpreting and improving natural language processing in machines with natural language processing in the brain : Removing learned attention harms deep layers’ brain prediction, indicating learned routing benefits deeper processing.

Concepts of Consciousness : Executive access enables content to affect reasoning and action—functional benefit of being routed to control systems.

Contradictory Evidence (1)

Interpreting and improving natural language processing in machines with natural language processing in the brain : Uniform attention improved shallow layers, suggesting early-stage processing may benefit from broader, less selective routing.

Subcellular structures (microtubules with MAPs) are proposed as additional substrates for routing via soliton-based logic, but current evidence is preliminary and not yet linked to conscious access.

Supporting Evidence (2)

On the Potential of Microtubules for Scalable Quantum Computation : Proposes MAP-mediated soliton transfer implementing XOR-like routing across microtubule networks.

Cytoelectric coupling: Electric fields sculpt neural activity and “tune” the brain’s infrastructure : Suggests electric fields can tune cellular infrastructure, potentially consistent with subcellular routing influences.

Contradictory Evidence (1)

Consciousness without a cerebral cortex: A challenge for neuroscience and medicine : Emphasizes mesodiencephalic/circuit-level bottlenecks as the gating locus, not intracellular logic.