Can brain-like learning rules discover hidden structure in complex data?

Researchers tested whether biologically plausible learning rules—using only local signals between neurons rather than brain-wide error signals—could learn hierarchical structure in high-dimensional data. Self-supervised approaches using contrastive learning succeeded where direct feedback approximations failed, matching supervised backpropagation's efficiency. This matters because it suggests the brain's actual plasticity rules may be sufficient for learning abstract representations without the unrealistic machinery backpropagation requires.