Backpropagation destroys V1 brain alignment in one epoch, tracking RSA alignment to fMRI across training for BP, FA, predictive coding, and STDP [R]

Third in a series of papers tracking learning rules vs. human fMRI (THINGS dataset, V1–IT, N=3 subjects). Previous finding: untrained CNNs match backprop at V1. This paper asks: when does training break that, and does the learning rule matter? Setup: RSA alignment measured at 8 checkpoints (epochs 0, 1, 2, 5, 10, 20, 30, 40), 5 seeds per rule, same architecture throughout. Main findings: BP drops 90% of V1 alignment after one epoch (r: 0.102 → 0.011, p = 0.031, consistent across all 5 seeds). FA drops 49%.