Bio Mirror / Cortical Architecture
The Cortical Column
A repeated six-layer motif with a stereotyped input, processing, and output division.
The structure
The neocortex is a sheet a few millimeters thick with six layers distinguished by cell type and connectivity. Layer 4 receives thalamic input. Layers 2 and 3 handle local and cortico-cortical processing and project to other cortical areas. Layers 5 and 6 provide output — layer 5 to subcortical targets, layer 6 back to thalamus. The motif repeats across the entire sheet with modest regional variation, and columns of cells spanning the layers share tuning properties.
The function it dictates
The column is the cortex's repeated computational unit. Its layered structure implements a consistent division: receive, transform locally, send onward, and send a signal back down the hierarchy. That descending pathway is substantial — feedback connections outnumber feedforward ones — which is a strong hint that cortex is doing something more like prediction against incoming signal than pure bottom-up feature extraction.
The mirror in AI
A transformer block is also a repeated, uniform unit stacked many times, with a fixed internal division of labor: route between positions, then transform in place, then pass on. The uniformity argument is the same in both systems — one motif, repeated, with capability coming from depth and data rather than from bespoke design per layer. The residual stream is the closest artificial analogue to a shared channel that every unit reads from and writes to.
The mirror in robotics
A layered control architecture where each level operates on a different timescale and abstraction — reflexes at the bottom, trajectory planning above, task planning above that — with information flowing in both directions. The bidirectional flow is the important part: high levels send goals down, low levels send state up.
Cortical feedback is massive and continuous; transformer blocks are strictly feedforward within a forward pass, with no descending pathway at all. Cortex operates continuously in time with oscillatory dynamics coordinating activity across regions; a transformer computes in discrete synchronized steps. And the column's uniformity is anatomical fact, whereas the claim that it implements a single canonical algorithm remains a hypothesis that has resisted proof for fifty years.
Open questions
- Is there a canonical cortical algorithm, or is uniform anatomy running region-specific computations?
- What would a transformer with genuine top-down feedback compute that the current one cannot?
Further reading
- Mountcastle, "The columnar organization of the neocortex" (1997)
- Douglas & Martin, "Neuronal circuits of the neocortex" (2004)