Carl Craver and I have written a paper arguing that functional analyses are just elliptical mechanistic explanations, contrary to the received view that functional analysis is distinct and autonomous from mechanistic explanation. Corollary: contrary to the received view, psychological explanation is not distinct and autonomous from neuroscientific explanation–rather, psychological explanation describes aspects of the same mechanisms described by neuroscientific explanation.
We are hoping to send the paper out for formal refereeing in a week or so. Any comments would be very much appreciated.
Abstract. We sketch a framework for building a unified science of cognition. This unification is achieved by showing how functional analyses of cognitive capacities can be integrated with the multilevel mechanistic explanations of neural systems. The core idea is that functional analyses are sketches of mechanisms, in which some structural aspects of a mechanistic explanation are omitted. Once the missing aspects are filled in, a functional analysis turns into a full-blown mechanistic explanation. By this process, functional analyses are seamlessly integrated with multilevel mechanistic explanations.
I emailed a list of – well, reactions I guess – to a quick scan of the paper. I think they come down to this: I am in agreement with the idea (as far as this *is* the idea in the paper!) that “functional” analysis must be grounded in a “mechanical” explanation, otherwise they are at best incomplete. I actually defend functional analysis, slightly, from the position in the paper, functional analysis can be useful methodologically and pragmatically, it is just a major error to believe it gives fundamental explanations.
I have various other comments to the paper, positive and negative. My own position is that this thinking must be advanced more aggressively, as seen in the paper’s discussion of whether this is reductionism or not. I believe the answer is neither yes nor no. How can this be? Well, is a computer program a “reduction” of some problem to code? I’m sure many would say yes, but I say no, and therein (perhaps) lies some discussion.
The paper is excellent.
I’m not entirely convinced that task analysis is just about creating phenomenally adequate models. I think they might be explanatory to some extent. I’ve just read some classical papers recently again, and I think task analysis is genuinely explanatory of the behavioral capacity even if it is unconstrained by the lower level (or constrained in a very loose way).
The main reason for that is that it is predictive and generalizable (even if you start with a single subject, like Newell and Simon with their cryptarithmetic tasks), and it seems to support counterfactuals etc. If it is general, predictive for human behavior etc., then it’s hard to dismiss as just a phenomenally adequate model which is too incomplete to be counted as explanatory. I mean, all explanations are incomplete as they have to omit some details to be practically useful (I think of explanations/predictions as lossy compression of information).
Take Newell and Simon on cryptarithmetic tasks: they can predict how people will deal with the task with a very high accuracy (around 90%). It would be a really a striking accident that they achieve such a high accuracy for such a complex phenomenon. Note that personal variations can be also easily simulated with their model. Similar is true of the Rumelhart & McClelland model of past tense learning: it is quite accurate in terms of showing the stages of learning (the rest is left out).
Overall, cognitive simulation seems to be explanatory only of the task (to some extent) and it could be thought to be a highly idealized account of the task. If the task is sufficiently trivial, the explanation could seem superfluous; yet if it is hard to understand (like, say, communication in insects), then we would think that it gets some knowledge about cognitive capacities, even if incomplete. Obviously, Newell and Simon, as well as Rumelhart and McClelland could be thought to build simply mechanism sketches (and that is a possible reading), but their accounts are also not accidentally predictive of the features of the task that they model. For some uses, it is already quite a feat. Moreover, if the high-level (I mean the level of abstraction here) specification of the task yields predictions, it may leave everything unconstrained, and for pragmatic reasons the scientists may ignore the lower levels. Note that though early Newell & Simon information-processing architecture is thought to be outdated right now but their account remains explanatory of the structure of the task, and it’s not just a phenomenally adequate and accidentally predictive model.
Yet, the only thing that the pure task analysis is explanatory about are the subtasks (task stages), their ordering, and the task result.
Anyway, my point is not really contradicting what you say. It’s just that some task analyses might be non-trivial and explanatory even if they are very loosely constrained.
And one more tiny thing: computational explanation of a behavior of a system might be thought to be more mechanistic just because you may want to explain the actual performance of the system (p. 43). Of course, you can specify the algorithm complexity with the big O notation, and that would abstract away from the actual implementation details, but for physical systems, we might want to get the values of the constants used in the notation as dummy placeholders so as to get the timing of operations right. This would be impossible for a purely functional computational explanation.