Taming the shrewdness of neural function: methodological challenges in computational psychiatry