Notes on sub-module addressing

In Notions on addressing of LMs, “input ports”, etc, I talked about the need to send messages to particular parts of (say) a Learning Module. In an attempt to ground this goal in human biology, I had ChatGPT generate a summary of the levels and information flow thought to exist in a human cortical column. After a bit of editing, here’s the result:

Cortical Layer & Sublayers Columnar Flow Role & Connections
Layer I – Molecular Layer
Function: Integration/modulation; combines apical dendrites from deeper layers
Inputs: Nonspecific thalamic nuclei, other cortical areas via horizontal fibers
Outputs: Local dendritic modulation of Layers II–III

Integration / Context
Role: Top-level integration across multiple cortical columns
Inputs ↑: From distant cortical and nonspecific thalamic sources
Outputs ↓: To pyramidal neurons in II–III
Layer II – External Granular Layer
Function: Cortico-cortical communication
Inputs: Association fibers (ipsilateral), commissural fibers (contralateral via corpus callosum)
Outputs: Association & commissural fibers to II–III in other columns

Local ↔ Distant Cortical Communication
Role: Horizontal distribution of local computations
Inputs ↑: From Layer IV and other columns
Outputs ↔: To other cortical areas
Layer III – External Pyramidal Layer
Function: Integration and relay to other cortical regions
Inputs: Other cortical areas, Layers II & IV
Outputs: Association (ipsilateral) & commissural (contralateral) fibers
Sublayers: IIIa (smaller pyramidal neurons), IIIb (larger pyramids, extrastriate cortex output)

Integration & Relay
Role: Bridge between local column and distant cortical targets
Inputs ↑: From II–IV in same/neighboring columns
Outputs ↔: To other cortical regions
Layer IV – Internal Granular Layer
Function: Main recipient of sensory input (especially in primary sensory cortex)
Inputs: Specific thalamic relay nuclei (e.g., lateral geniculate nucleus, ventral posterolateral nucleus), other cortical areas
Outputs: Primarily to Layers II–III locally
Sublayers (in V1): IVa (magnocellular pathway), IVb (motion processing), IVcα (magnocellular LGN), IVcβ (parvocellular pathway)

Primary Sensory Input Hub
Role: Entry point for feedforward sensory info
Inputs ↑: From thalamus
Outputs ↑: Feedforward to II–III
Layer V – Internal Pyramidal Layer
Function: Major output to motor and subcortical targets
Inputs: Layers II–IV
Outputs: Brainstem, spinal cord, striatum via pyramidal tracts, including corticospinal tract and corticobulbar tract
Sublayers (in primary motor cortex): Va (smaller pyramids, subcortical targets), Vb (giant Betz cellslower motor neurons)

Motor & Subcortical Output
Role: Convert integrated info into motor commands
Inputs ↑: From II–IV
Outputs ↓: To motor and integrative centers
Layer VI – Multiform / Polymorphic Layer
Function: Feedback to thalamus, modulating sensory input
Inputs: Layers II–V, thalamus
Outputs: Corticothalamic projections to specific & nonspecific nuclei
Sublayers: VIa (projects to thalamic relay cells), VIb (to local cortex + thalamic reticular nucleus)

Thalamic Feedback / Gating
Role: Adjust thalamic relay neuron gain
Inputs ↑: From cortex & thalamus
Outputs ↓: To thalamus, reticular nucleus

Columnar Flow Summary

I’d like folks here to examine this (i.e., sanity check it), but I suspect that it’s mostly accurate. So, on to the next sub-topic…

A lot of the data paths described above happen within each cortical column. These will be handled by the process (or set of processes) that implements each Learning Module, so we can ignore them for the moment. Meanwhile, here are some inter-column data paths:

Incoming

  • Layer I (Molecular Layer) ← distant cortical and nonspecific thalamic sources
  • Layer III (External Pyramidal Layer) ← Layers II–IV in neighboring columns
  • Layer IV (Internal Granular Layer) ← feedforward sensory information
  • Layer VI (Multiform / Polymorphic Layer) ← cortex and thalamus

Outgoing

Let’s assume that we can address a message to a particular cortical column (or set of columns, using the publish-subscribe pattern). How should we specify the target level(s), let alone the types of neurons we want to receive the message. Suggestions?