THE FEEDBACK

By Aliya Rumana
Center for Philosophy of Science
University of Pittsburgh

Abstract:
David Marr famously argued that computational theory (i.e., analysis at the computational level) was required to explain both “what the device does and why”. In a series of papers, Oron Shagrir and William Bechtel argue that computational theory explains how certain mechanisms are appropriate for certain tasks by showing that identity holds between the corresponding mechanisms and the tasks at an abstract, computational level of description. Call this the “computational identity account” of “mechanism-task fit” (or “M/T fit”). Inspired by their work, I propose an alternative account that grounds M/T fit in constraint satisfaction, where the mechanism is appropriate to the task because the mechanism’s properties satisfy all the task-related constraints. I use retinal edge detection and sound localisation as two cases to demonstrate that constraint satisfaction may be a better way to ground M/T fit than identity. This account of M/T fit isn’t confined to the computational level of description, so I describe it as “task-fitting explanation” rather than computational theory. I argue that task-fitting explanation is a species of constraint-based explanation: it is interested in which features of a mechanism make possible above-chance correct task performance for the mechanism. As such, it is “modally complementary” to mechanistic explanation, which, I argue, is interested in which activities done
by a mechanism’s parts make actual competent task performance for the mechanism.

Explaining Mechanism-Task Fit in Neuroscience was recently published in the British Journal for the Philosophy of Science.

Commentary from Aliya:

Philosophy of neuroscience grew to prominence with a surge of interest in mechanistic explanation, following a landmark 2000 paper by Peter Machamer, Lindley Darden, and Carl Craver. A quarter century later, interest in mechanistic explanation is waning and many philosophers of neuroscience are interested in characterising alternative forms of explanation (e.g., functional, topological, efficient-coding, teleological, and constraint-based).

However, Carl Craver, David Kaplan, Gualtiero Piccinini, and others have raised an important challenge for these “alt-mechanistic” projects. They acknowledge that these explanations may look different from mechanistic explanations, but they are only explanatory to the extent that they accurately represent mechanisms. When we correct the mistakes and fill in the missing details, these explanations will turn into full-blown mechanistic explanations. In other words, their non-mechanistic appearances can be chalked up to their inaccuracy or incompleteness. This objection is often misunderstood (and maligned), but I think it is essential to an objectivist approach to neuroscience and exceptionally difficult to refute.

Yet we can find a blueprint for a response to this objection in David Marr. In a prescient move, Marr’s alt-mechanistic argument started with the completeness conditions of mechanistic explanation. He said that even if we had a complete description of a neural mechanism, we would still an explanation for why this mechanism made possible successful performance on a task. It was not enough to talk about causal structure between stimulus and response. Marr insisted that an explanation of what I call “mechanism-task fit” had to describe some higher-order relationship between that causal structure and the structure of the task. Note how this flips the mechanist script: it says that the completeness conditions of mechanistic explanation form a proper subset of the completeness conditions of some superordinate kind of explanation.

Unfortunately, Marr’s argument is notoriously open to interpretation. Yet I’ve always admired the interpretation offered by Oron Shagrir (in solo work and in collaboration with Bill Bechtel), who suggests that a mechanism is appropriate to its task when a coarse-grained, computational description of the mechanism is equivalent to the coarse-grained, computational description of the task. Even so, I find this reading dissatisfactory. In particular, I worry that it fails to capture the alt-mechanistic spirit of Marr’s original argument. After all, it suggests that coming up with “computational-level” explanations is just a matter of looking for some way of coarse-graining our description of the neural mechanism so that it’s equivalent to a description of the task.

In Explaining Mechanism-Task Fit in Neuroscience (BJPS), I offer another reconstruction of Marr’s view that, I hope, does more justice to its alt-mechanistic aims. On my view, a mechanism is appropriate to its task when the mechanism’s features satisfy all the constraints on successful task performance. These constraints have hierarchical structure: some are generic to all task performers, some are generic to certain types of task performers, and some are specific to the mechanism itself. Discovering these constraints is a substantive question for neuroscience, not one that reduces to finding the right way of coarse-graining a mechanism description. Thus, I believe this gives us an ontologically robust alternative to mechanistic explanation.  

To be clear, though, this “task-fitting explanation” (as I call it) fits quite well with mechanistic explanation. In fact, they are what I call “modal complements”: task-fitting explanations are interested in which features make possible competent task performance whereas mechanistic explanation is interested in which activities make actual competent task performance. Both are required for genuinely complete explanations of competent task performance.