Abstract
Models of complex heterogeneous systems like the brain are inescapably incomplete, and thus always falsified with enough data. As neural data grow in volume and complexity, absolute measures of adequacy are being replaced by model selection methods that rank the relative accuracy of competing theories. Selection still depends on incomplete mathematical instantiations, but the implicit expectation is that ranking is robust to their details. Here we highlight a contrary finding of “brittleness,” where data matching one theory conceptually are ranked closer to an instance of another. In particular, selection between recent models of decision making is conceptually misleading when data are simulated with minor distributional mismatch, with mixed secondary signals, or with non-stationary parameters; and decision-related responses in macaque cortex show features suggesting that these effects may impact empirical results. We conclude with recommendations to mitigate such brittleness when using model selection to study neural signals.
