Learning to program "recycles" preexisting F-P pop codes of logical algorithms (www.jneurosci.org)

Researchers used fMRI to track how novice programmers' brains represent core programming constructs and found that learning to code “recycles” preexisting logical representations rather than creating them from scratch. Twenty-two college students were scanned before and after a semester-long Python course while reading Python functions, pseudocode, and control working-memory stimuli. A left-lateralized fronto-parietal reasoning network was engaged when reading real code after instruction — and crucially, the same network was already activated by pseudocode before instruction. Multivariate pattern analyses and representational similarity analysis showed that population codes in this network reliably distinguished “for” loops from “if” conditionals both before and after the course, with shared representational structure across time. For the AI/ML community, this supports the neural recycling hypothesis: cultural skills like programming leverage preexisting cognitive maps for logical algorithms. That contrasts with the de novo emergence of representations often seen in artificial neural networks and suggests human learning repurposes abstract reasoning circuitry. Implications include new perspectives on how to design educational curricula and human-aligned AI: models of transfer learning and symbolic reasoning might benefit from architectures or training regimes that mirror this reuse of structured, compositional representations.