Why brain-like learning algorithms converge differently than standard backpropagation

When RNNs are trained under biological constraints—using only local information instead of full backpropagation—they reach different solutions that are mathematically simpler (low-rank). Williams, Payeur, and Lajoie analyzed this using dynamical systems theory, showing that locality constraints create qualitatively distinct learning dynamics and convergence rates. This matters because it explains trade-offs between biological plausibility and learning performance.