The transformer as a candidate canonical cortical microcircuit

The transformer circuit uniquely matches fine-grained unit-level details, population functional maps, neuronal activity patterns, and systems-level wiring efficiency across cortical signatures.

• Why it matters: A canonical cortical microcircuit is hypothesized but hard to validate; this work tests whether a specific computational motif aligns with brain organization.
• What they did: Candidate microcircuit architectures were embedded within a shared spatial macrocircuit and compared against an extensive set of cortical signatures using large-scale data and mapping techniques.
• The result: A learned multiplicative component enabling relational information without extensive long-range wiring is linked to the transformer's uniquely strong alignment, suggesting convergence between natural and artificial computation.

AI summary of the abstract · may contain errors · not medical advice · read the paper.