← Reference library

REFERENCE LIBRARY

Cortical Layers & MicrocircuitCortical Layers

The cortex is two to three millimetres thick and, unfolded, about the size of a napkin — yet it is neatly divided into six layers. The layers are not decoration: which layer a connection leaves from and which one it lands in is tightly prescribed, and that landing rule is what makes "up the hierarchy" and "back down it" distinguishable at all.

Six layers, each with its own ports

Counting inward from the surface, the convention is six layers. What matters is not the names but that each layer talks to a different set of partners:

L1 (molecular layer)

Almost no cell bodies — an entire layer of landing strip: feedback from distant areas and modulatory thalamic input arrive here, caught by the apical tufts reaching up from cells below.

L2/3

The feedforward workhorse between cortical areas, sending axons in bulk to higher regions. Also the layer with the densest local horizontal wiring.

L4

The receiving dock of sensory cortex, where thalamic input mostly lands before being handed up to L2/3. In regions such as prefrontal cortex, L4 is thin or barely identifiable.

L5

The downward exit: projections to striatum, thalamus, brainstem and spinal cord. Its large pyramidal cells run an apical dendrite all the way to L1 — the clearest case of a cell with two ends doing two jobs.

L6

Mostly back to the thalamus — the descending pathway that makes roughly ninety percent of thalamic input come from cortex (→ Thalamus).

The textbook chain "thalamus → L4 → L2/3 → L5/6" is a fine starting point, but not the only route: the thalamus also projects directly to L5 and L1, and the traffic between layers is far richer than a single chain.

Feedforward lands in L4, feedback avoids it

The most useful thing about this layering: the layer a connection lands in tells you whether it is going up or coming down. Felleman and Van Essen used exactly this criterion in 1991 to arrange dozens of macaque visual areas into a hierarchy — feedforward leaves L2/3 and terminates in L4; feedback leaves the deep layers, avoids L4, and lands in L1 and L5/6.

Why it matters: a single cell can receive both kinds of input at different addresses — evidence from below arriving near the soma, interpretation from above arriving at the apical tuft. Predictive coding needs predictions and errors to travel separately; biologically plausible versions of backpropagation need the signal and the teaching signal not to interfere. Both need this anatomical rule to stand on.

The two ends of a pyramidal cell

The main excitatory cell of cortex is shaped with intent: basal dendrites are short and dense around the cell body, taking local and feedforward input; the apical dendrite is a long shaft to the surface that opens into a tuft in L1. The two ends do not simply add — when the apical end is driven strongly enough, the cell switches from single spikes to bursts, so downstream cells can tell whether the top-down stream just spoke (Topic 36 covered what dendrites compute on their own).

Inhibition is divided up along the same lines: one class of interneuron wraps the cell body and cuts output outright, while another wraps the distal apical dendrite and suppresses specifically "don't let that top-down stream be heard" (→ Inhibitory interneurons). The apical channel can therefore be switched on and off by itself.

Columns: useful, but not dogma

A small cluster of cells running vertically through the six layers tends to share preferences, which led Mountcastle to propose the cortical column and to suggest the whole cortex runs one general-purpose circuit on different inputs. The conjecture has been enormously influential and does explain a great deal. To be clear, though: not every cortical area shows clean columnar structure, and whether a column is a real computational unit remains contested — it is not settled fact.