The mammalian neocortex is organized into cortical columns — vertical units spanning six layers, each with distinct cell types, connectivity patterns, and computational roles. Layer 4 receives sensory input. Layers 2/3 perform recurrent integration and send output to other cortical areas. Layer 5 projects to subcortical targets. Layer 6 provides feedback to thalamus. The architecture is remarkably conserved across sensory modalities and across species.
Koenig and Negrello map transformer operations onto this laminar organization. Attention mechanisms — which select and route information based on content — correspond to the contextual modulation observed in superficial layers. Feedforward blocks — which transform representations through nonlinear projections — correspond to the interlaminar processing between layers. The residual stream that carries information across transformer blocks maps onto the apical dendrites that integrate inputs across cortical layers.
The mapping is not metaphorical. It generates testable predictions: specific cortical layers should show attention-like contextual selection, others should show feedforward-like transformation, and the oscillatory coordination between layers should reflect the sequential processing order of a transformer block. The authors propose these as falsifiable hypotheses, not explanatory stories.
The interesting structural claim is convergence. Transformers were not designed to mimic cortical columns. They were designed to process sequences efficiently. Cortical columns were not designed at all — they evolved under selection pressures for flexible sensory processing. If the computational organization genuinely matches, it's because the same computational problem — contextual integration of high-dimensional inputs — admits a solution that two radically different optimization processes found independently. The architecture is not borrowed. It's convergent.