Why neural models succeed even when they're secretly wrong

Researchers tested whether successful predictions by simplified neural models—ones that assume inputs act independently—actually reveal how neurons compute. Using maximum-entropy models and hippocampal recordings, they found that these simple models can absorb hidden complexity into their fitted parameters through a phenomenon called feature leakage. A neuron with genuinely nonlinear computations can masquerade as a linear one if you measure it under the right conditions. The distinction matters: prediction accuracy alone doesn't tell you the mechanism.