On some two-way barriers between models and mechanisms.

A number of recent as well as classic ideas suggest that there are constraints and limits on the explanatory role that computational, mathematical, and neural net models of visual and other cognitive processes can play that have not been generally appreciated. These ideas come from mathematics, automata theory, chaos theory, thermodynamics, neurophysiology, and psychology. Collectively, these ideas suggest that the neural or cognitive mechanisms underlying many kinds of formal models are untestable and unverifiable. Models may be good descriptions of perceptual and other cognitive processes, but they cannot in principle be reductive explanations nor can we use them to predict behavior at the molar level from what we know of the neural primitives. This discussion is an effort to clarify the appropriate meanings of these models, not to dissuade workers from forging ahead in the modeling endeavor, which I acknowledge is progressing and is making possible our increasingly deep appreciation of plausible and interesting cognitive processes.