Last time I discussed the examples of research on memory, visual object recognition and categorisation. This kind of functional modelling work is the rule, not the exception in cognitive science - it's how we're taught to work and how the field moves along.
@PsychScientists A mechanism is a graph with at least three boxes and two arrows.— Tim van der Zee (@Research_Tim) May 24, 2016This kind of program feels like it's heading towards mechanism . Every division into new sub-capacities comes from work showing the two sub-capacities function differently and are therefore the result of different mechanisms. Every new representational model adds a new component (part or process) that handles another part of the capacity. There is one basic problem, however. None of these models make any explicit reference to any real parts or processes that have been empirically identified by other work - for example, 'working memory' still refers to a capacity, not a component. This means there is no reason to think this new capacity maps onto any particular parts and processes or if it does, to which parts and processes.
We are, in effect, doing science backwards: modelling first, running experiments later, and the result is that we are not actually on a trajectory towards mechanistic models, just better functional ones. This is a problem to the extent you want access to the many real benefits mechanistic models offer, in particular the ability to explain rather than simply describe a mechanism (see Bechtel & Abrahamsen, 2010 and the last post). This post reviews whether functional models explain or whether they can be part of a trajectory towards an explanation. The answer, unsurprisingly, will be no.
Functional Models Do Not Lead to Mechanistic ModelsAs we discussed, mechanistic 'how-actually' models are the thing you build once you have independently investigated the mechanism and identified the relevant real components. Bechtel & Abrahamsen (2010) note that cognitive science does not build models this way. Instead, the approach is to build a model that hypothesizes some components and then demonstrate that the model can produce the behavior in question. We model build, and run experiments later. The problem is that this only allows you to build 'how-possibly' models, or at best, 'how-probably' models, because "Empirical evidence can rule out specific models, but usually cannot decide between competing architectures." (pg. 332).
Bechtel & Abrahamsen demonstrate this fact with a classic argument in the cognitive literature, between symbolic (e.g. Pinker & Prince, 1988) and connectionist (e.g. Rumelhart & McClelland, 1986) modellers trying to model language. These two camps have traded models and objections for years. One group would model a linguistic phenomenon the other couldn't handle (e.g. learning the past tense of irregular verbs), the other would tweak until they could handle it and come back with something they could do that the first camp couldn't. This goes on without end, and reveals that the various experiments cannot decide which architecture is the real one because each can be made to generate the effects. (Anyone who has been doing psychology for a while should find the form of this arguing very, very familiar.)
The problem is that there are always many possible ways to generate a given behaviour, but typically only one way that the mechanism actually generates that behavior. If you make your model without reference to an already-understood actual mechanism (as you do in any flavor of functional analysis) all you end up with is one of the many functional, 'how-possibly' options instead of the one mechanistic, 'how-actually' version and you won't ever figure out which option is the real one.
Levels of AnalysisTo decompose a mechanism into the right real components, you have to be cutting at the right level. One reason why cognitive models aren't mechanistic is that we have yet to settle on the correct level to begin the analysis. We either go too high (with mental representations) or too low (with neural mechanisms), and the result is a set of non-mechanistic models that cannot be reconciled (think how hard cognitive science rejects the idea of reducing cognition to neural activity, for example). In addition, even when you do try to reconcile them, a search for the neural correlates of those mental representations is confounded by neural degeneracy and neural reuse; see this earlier post for details. This is another way to frame the impossibility of going from functional to mechanistic models; the many-to-one mappings are underconstrained.
Bechtel has another example to make this point about levels clear.
The situation might be compared to the state of fermentation research in the late nineteenth century. By describing the potential intermediates formed in the process of fermentation as themselves undergoing fermentations, physiologists looked too high. They provided little explanatory gain, since researchers were appealing to the phenomenon to be explained to describe the operations that were to provide the explanation. In contrast, by focusing on the elemental composition of sugar and alcohol and appealing to operations of adding or deleting atoms to explain organic processes such as fermentation, chemists focused too low. The chemists clearly appealed to operations on components in a mechanism to explain the phenomenon, but this approach was underconstrained. Researchers lacked principles for determining which operations were really possible.Researchers eventually identified another level of analysis - biochemistry - at which the operations relevant to fermentation function. Because of this identification, the field was, at last, able to move forward in developing mechanistic models. Right now, cognitive science has not identified such a level for itself (although to foreshadow, we have an idea).
Bechtel, 2008 p. 989.
Summary
Bechtel & Abrahamsen (2010) pose a challenge to cognitive science, to shift the research strategy to one that makes explanatory mechanistic models a possibility because of all the advantages that come with such models. They do not think it is impossible, although they recognize it will be hard:
Computational modeling in circadian rhythm research .... suggests a path for cognitive modeling that can attend to dynamics without ignoring mechanism: equations in which the variables include properties of parts and operations as well as time. It also highlights the value of computational modeling that is anchored in an empirically derived mechanistic account. This kind of dynamic mechanistic explanation will not come easily in cognitive science, in that it must target extremely complex mechanisms. However, the necessary empirical and conceptual tools are increasingly available, as is the will to use them in this way.There are two ways to meet this challenge. The first, advocated by Weiskopf (2011) and Chemero & Silberstein, is to reject the idea that reductionist, mechanistic models should be the goal and that there are perfectly acceptable explanatory alternatives. This post addresses the Weiskopf suggestion; we will review the Chemero argument in the next post. The second, advocated by us, is to accept the challenge and work to identify the correct level of analysis to ground efforts to appropriately decompose cognitive systems into real components we can then mechanistically model. We will introduce this argument in the penultimate post of the week.
pg 332
References
Bechtel, W. (2008). Mechanisms in cognitive psychology: What are the operations? Philosophy of science, 75(5), 983-994.
Bechtel, W., & Abrahamsen, A. (2010). Dynamic mechanistic explanation: Computational modeling of circadian rhythms as an exemplar for cognitive science. Studies in History and Philosophy of Science Part A, 41(3), 321-333.
Pinker, S., & Prince, A. (1988). On language and connectionism: Analysis of a parallel distributed processing model of language acquisition. Cognition, 28, 73–193.
Rumelhart, D. E., & McClelland, J. L. (1986). On learning the past tenses of English verbs. In J. L. McClelland, & D. E. Rumelhart (Eds.), Parallel distributed processing: Explorations in the microstructure of cognition, Vol. 2. Psychological and biological models (pp. 216–271). Cambridge, MA: MIT Press.
Weiskopf, D. A. (2011). Models and mechanisms in psychological explanation. Synthese, 183(3), 313-338.