Abstract

Psychological Inquiry, 22: 137–138, 2011 Copyright Taylor & Francis Group, LLC ISSN: 1047-840X print / 1532-7965 online DOI: 10.1080/1047840X.2011.551323 Edward Vul Department of Psychology, University of California, San Diego, La Jolla, California All models are wrong, but some are useful. — George These trade-offs apply to predicting human cog- Box (1979) nition. If we want to predict whether a given drug will increase dopamine in Parkinson’s patients, psy- chological and cognitive neuroscience descriptions are What is the best level at which to describe human practically useless: Our question is about biochemical cognition? We could describe it using mathematical interactions, so biochemical descriptions of the brain formalisms (like Bayesian statistics) at Marr’s (1982) provide the most useful basis for psychopharmacol- computational level by specifying the sources of in- ogy. In contrast, if we want to tell a neurosurgeon formation in the world and our own inductive biases where to cut to avoid damaging the patient’s capac- that we draw on to make inferences and choose ac- ity for speech, biochemical and psycholinguistic de- tions in the world. We could describe human cognition scriptions are useless; however, theories and data from at the algorithmic level using the language of com- cognitive