2.1. The Mechanism Approach

In the past decade, a number of philosophers of science have begun to develop an approach to the nature of models, explanation, and discovery based on the notion of a mechanism (Bechtel and Richardson Reference Bechtel and Richardson1993; Glennan Reference Glennan1996; Machamer, Darden, and Craver Reference Machamer, Darden and Craver2000; Craver Reference Craver2001; Craver and Darden Reference Craver, Darden, Machamer, Grush and McLaughlin2001). A mechanism is defined as “entities and activities organized such that they are productive of regular change from start or set-up to finish or termination conditions” (Machamer, Darden, and Craver Reference Machamer, Darden and Craver2000, 3). The setup conditions are entities and their properties (e.g., parts, structure, spatial relations, and orientations) and enabling conditions (e.g., available energy, pH, electrical-charge distributions). Termination conditions are idealized states or parameters that constitute a privileged endpoint (e.g., rest, equilibrium, elimination of something, production of something). The intermediate activities that take the mechanism from the setup conditions to the termination conditions are intervening entities and activities (e.g., causal chains, forks, joins, cycles) (Machamer, Darden, and Craver Reference Machamer, Darden and Craver2000).

Mechanisms are characterized by both temporal and spatial organization. With respect to temporal organization, a mechanism, typically, has stages that are ordered and that have characteristic rates and durations. With respect to spatial organization, the stages of a mechanism are typically localized with respect to particular entities and their activities, and are connected to each other in various ways. These connections often depend on the structure of the relevant entities and their spatial interrelations. Mechanisms can be described with varying degrees of completeness. A complete mechanistic description will exhibit “productive continuity” in terms of “bottoming out” lower-level mechanisms, where bottoming-out mechanisms are “the components that are accepted as relatively fundamental or taken to be unproblematic for the purposes of a given scientist, research group, or field” (Machamer, Darden, and Craver Reference Machamer, Darden and Craver2000, 13). In contrast, a mechanism schema will be “a truncated abstract description of a mechanism that can be filled with descriptions of known component parts and activities” (Reference Machamer, Darden and Craver2000, 13), whereas a mechanism sketch is “an abstraction for which bottom out entities and activities cannot (yet) be supplied or which contain gaps in the stages” (Reference Machamer, Darden and Craver2000, 18).

Mechanism descriptions vary not only with respect to degree of completeness or specificity, they can also vary with respect to their “vertical” scope. To capture the idea that a mechanism can consist of other mechanisms, Craver has developed the notion of a mechanism hierarchy. In Figure 1, we see a mechanism S, which is Ψ-ing, composed of smaller entity Xs, which are Φ-ing. Each of these Xs is itself a mechanism consisting of even smaller entity Ps, which are σ-ing (Craver Reference Craver2001).

Figure 1.

Mechanism hierarchy (adapted from Craver Reference Craver2001).

Mechanistic descriptions can be used for various explanatory purposes. Craver distinguishes three kinds of mechanistic explanations: etiological, constitutive, and contextual. Only the second—constitutive—is relevant for our purposes. The explanandum in a constitutive explanation is the activity (or “phenomenon”) exhibited by a mechanism, taken as a whole or, more precisely (if one adopts the view that explananda are always questions, which we do), how (actually, possibly, or plausibly) S exhibits the activity (process, capacity) of Φ-ing. The explanans is then, simply, a description of the internal mechanistic structure of S, in terms of its lower-level entities and activities and their interrelations, in virtue of which S does, or can, or possibly could, or plausibly could Φ. In other words, a constitutive mechanistic explanation of how S Φs makes reference to the mechanism that constitutes S and its Φ-ing (Craver Reference Craver2001).

A similar explanatory notion has become standard in the philosophy of psychology where the “phenomenon” in question is usually taken to be a capacity rather than an activity. Cummins suggests that a system's or organism's capacity to do something can be explained by a functional analysis of that capacity. By a functional analysis of some capacity C, he means, roughly, a breakdown of the complex capacity into a set of constituent capacities ( $c_{1},\,c_{2},\,\ldots,\,c_{n}$), with a specification of the sequence in which those constituent capacities must be exercised for the result to be a manifestation of the complex capacity (Cummins Reference Cummins1975). Von Eckardt points out that a functional analysis for a complex capacity C only provides us with a possible explanation of how the organism has C. The reason is that most complex capacities can be broken down into various alternative sequences of constituent capacities, the exercise of each of which is sufficient for the exercise of the complex capacity. What is required for an actual explanation is not only a functional analysis but some reason to believe that the system or organism in question actually has (or exercises) C in virtue of actually having (and exercising) the constituent capacities of the functional analysis in the order specified by the functional analysis, namely, the functional analysis must be structurally adequate (Reference Von Eckardt (Klein) and Walker1977). Von Eckardt further suggests the following as sufficient conditions for structural adequacy:

The model that describes a possible structural decomposition of O such that if it were true of O, conditions (1)–(3) would be met, is a functional component model of how O has C.

Note that all of this is easily translated into the language of the mechanism approach. (Indeed, the breakdown of O into component [physical] parts whose operation results in O's exercising C is explicitly referred to in Von Eckardt (Reference Von Eckardt (Klein) and Walker1977) as a “mechanism.”) A functional analysis can be viewed as a mechanism sketch that begins to provide a constitutive mechanistic explanation of how O exercises C. It is only a mechanism sketch because constituent capacities are hypothesized without reference to the constituent entities that (presumably) have those capacities. However, a functional analysis combined with a functional component model is equivalent to a complete constitutive mechanistic explanation. Both can be further specified with respect to modality or epistemic strength, and for both we can distinguish an explanation of a capacity itself vs. an explanation of the exercise of a capacity.

2.2. Cognitive Neuroscience

In the 1970s and 1980s, there were two multidisciplinary fields studying the mind/brain without much interaction: cognitive science and the neurosciences. As Bechtel has emphasized, both acknowledged the importance of the other in principle, but for both there were reasons for not taking research in the other all that seriously. On the one hand, cognitive science paid lip service to the importance of neural realization but there was no clear picture, even in principle, of how classical (“symbolic”) computational models could be realized in the brain and the data coming out of neuropsychology and neurology was not very useful in constraining those models. On the other hand, the major successes in the neurosciences had been at the cellular and molecular levels but there were no techniques for studying higher cognitive processes other than the classical neurological and neuropsychological ones (Bechtel Reference Bechtel, Machamer, Grush and McLaughlin2001).

All that changed in the 1980s with the emergence of the technologies of PET, MRI, and fMRI. Within a decade there were eight major centers for cognitive neuroscience funded by the McDonnell Foundation and the Pew Charitable Trust, a new Journal of Cognitive Neuroscience, and a new Cognitive Neuroscience Society. The aims of this new field were, first, to further the development of adequate psychological models of cognition by generating and bringing to bear a variety of kinds of neural data and, second, to develop adequate hypotheses about the neural localization and realization of cognitive structures and processes described psychologically. Michael Gazzaniga, the first editor of the Journal of Cognitive Neuro science, wrote in his opening editorial that the journal was for cognitive scientists who “now believe that it is maximally fruitful to propose models of cognitive processes that can be assessed in neurobiologic terms” and neuroscientists who believe they must “actually come to grips with the complexities of psychological processes involved in any particular mental capacity” (Gazzaniga Reference Gazzaniga1989). Something very similar could be said of the field itself.

Another way to understand the explanatory aims of cognitive neuroscience is provided in Posner and Rachle's nice introduction to the field, Image and Mind (Reference Posner and Raichle1994). There they distinguish five levels for linking the cognitive and neural levels of analysis—the cognitive system, the mental operation, the performance domain (or pathway activation), the neural system, and the cell—each of which is associated with its own methods of investigation. For our purposes, what is important to note is that the typical cognitive neural model includes an information processing story of the capacity under study (usually, at a fairly crude level) plus hypotheses about the functional localization of the constituent capacities and operations posited by that story. The hope is, then, that in the future, the model will be supplemented by hypotheses detailing the neural implementation of these operations.

4.1. Constitutive Explanation in Cognitive Neuroscience Is Representational

As we indicated earlier, a type of explanation, namely, functional analysis, similar to constitutive mechanistic explanation, has been part of the philosophical reconstruction of cognitive science for more than twenty-five years. It is similar because it seeks to explain certain global properties of a system or organism in terms of that system's or organism's constitution. More specifically, it seeks to explain how organism O has such-and-such properties (including capacities) by reference to properties (including subcapacities) of parts of O. It has been understood for many years, however, that constitutive explanation in cognitive science is a particular species of constitutive explanation, namely, it is a type of constitutive explanation that involves appeal to representations. In particular, the basic explananda of cognitive science are the human cognitive capacities along with a host of properties of those capacities that involve intentional states (e.g., pragmatic evaluability, coherence, and reliability). To explain the intentionality of mental states, cognitive scientists have adopted the strategy of “semantic heritability.” Roughly speaking, the strategy is to explain intentionality by reference to representations with content and significance for their “owner.” Given this strategy, the explanation of how a person exercises a certain capacity (e.g., reads a sentence) appeals to a computational story (either “symbolic” or connectionist) that details the steps or stages the person's mind/brain goes through as it goes from a visual (or auditory) representation of an inscription (or utterance) of the sentence to a representation of its meaning, that is, the representational state underlying our “understanding” of the sentence. The explanation of how a person has (rather than exercises) a certain capacity appeals to the computational and representational resources required for any exercise of that capacity. Both kinds of explanation are constitutive and both involve representations.

What does all of this have to do with cognitive neuroscience? If theoretical models in cognitive neuroscience are information-processing models mapped onto the anatomy and physiology of the brain, then if constitutive explanations in cognitive science involve representations, then so will constitutive explanations in cognitive neuroscience. And, indeed, this is precisely what we find, although sometimes the hypotheses are framed in terms of intentional capacities rather than strictly in terms of representations. For example, it has been known for some time that in monkeys there are two main streams of projections from primary visual cortex to other cortical regions (Ungerleider and Mishkin Reference Ungerleider, Mishkin, Ingle, Goodale and Mansfield1982): a ventral stream from V1 to the infero-temporal cortex and a dorsal stream from V1 to posterior parietal cortex. There has been some dispute, however, about the function of these two streams. Ungerleider and Mishkin's hypothesis was that the ventral stream played a major role in the visual identification of objects, whereas the dorsal stream concerned the localizing of objects in visual space. However, Goodale and Milner have argued that the important functional distinction is between visual perception and the visual control of action (Goodale and Milner Reference Goodale and Milner1992; Milner and Goodale Reference Milner and Goodale1993). According to this account,

4.2. Mental Representations

We have suggested that cognitive neuroscience, like cognitive science itself, invokes mental representations to explain cognition. What is it about resorting to representations that poses problems for the mechanism approach? To answer this question, we must consider what mental representations are, with a focus on how they are constituted.

Mental representations are best understood in terms of Charles Peirce's general triadic theory of representation. They consist of: (1) a representation bearer, which has some sort of (2) content, which, in turn, has (3) significance for an interpreter who, in this case, is the person in whose mind/brain the representation resides. Insofar as a mental representation has these three aspects, a constitutive account of mental representation must explain in virtue of what any given mental representation has (1), (2), and (3). A constitutive explanation of a mental representation's bearer is the most straightforward. If one views the mind/brain as not only a representational system but also a computational one, then the bearers of mental representations will be nonsemantic, purely formal, computational structures or states. Precisely what kinds of computational structures or states will depend on whether one views the mind-brain as a “classical” computational system, as a “connectionist” system, or as some other kind of computer. For example, if the mind/brain is viewed as a classical computational system, then the representation bearers of mental representations will be data structures. If it is viewed as a connectionist computer, the representation bearers will be patterns of activation (for recurrent representational states) or patterns of weights (for stored representations). All of this is, obviously, quite compatible with a mechanism approach.

In contrast, difficulties for the mechanism approach emerge when we consider how both the content and significance of mental representations might be realized. Consider first, content. Suppose we are trying to explain how people exercise the capacity to provide a use verb (e.g., ‘hit’) in response to a visually presented name of an object (e.g., ‘hammer’). A cognitive psychologist or cognitive neuroscientist will most likely hypothesize that the early stage of exercising this capacity is to form a representation of the stimulus word. Let's suppose further that the representation will be an orthographic one, that is, it will represent the word ‘hammer’ in a way that is similar to a standard printed representation, by a sequence of component representations of the letters that spell the word ‘hammer’.

A major foundational question of cognitive science (and, ipso facto, cognitive neuroscience, insofar as it incorporates the information-processing level of description) is how content is realized naturalistically. Philosophers (and a few psychologists) have proposed a variety of theories of content determination (see Von Eckardt Reference Von Eckardt (Klein)1993 for an overview) the details of which need not concern us here. The important point is that many of these theories are externalistic. That is, the naturalistic properties and relations of a representation bearer that realize its content are taken to be properties and relations that extend beyond the head, and, indeed, in the case of the evolutionary version of the teleosemantic theory of content determination, into the distant past. Furthermore, there is every reason to believe that the correct theory of content determination will have, at a minimum, an externalistic component. Clearly, such externalistic components will not be part of any neural mechanism.

What about significance? The significance of a mental representation is how it is used by its “owner.” Von Eckardt (Reference Von Eckardt (Klein)1993) argues that within cognitive science (and so, again ipso facto, within cognitive neuroscience) the significance of a representation for some subject S consists in the set of all possible determinate computational processes contingent upon entertaining (activating) that representation in S. In other words, it consists of something like downstream “narrow” functional role. Since the computational consequences of entertaining a representation are “in the head,” it might not be immediately obvious why this account of representational significance would cause problems for the mechanism approach to constitutive explanation. The reason turns on the fact that any given (fairly specific) type of representation, say, the orthographic representation of the word ‘hammer’ mentioned above, will play a role in many capacities and, hence, will have a wide variety of computational consequences. However, when we are delineating the mechanism underlying a specific capacity, say, reading, we will be focusing on only some of those consequences. So there are likely to be other consequences that, on the one hand, constitute part of the significance of the representations that are part of the mechanism of reading, but that, on the other, play no role in the mechanism itself.

A similar challenge to a purely mechanistic reconstruction of cognitive neuroscience arises in the domain of psychopathology. For example, according to the Diagnostic and Statistical Manual of Mental Disorders, 4th edition (American Psychiatric Association 1994), a delusion is “a false belief based on incorrect inference about external reality that is firmly sustained despite what almost everyone else believes and despite what constitutes incontrovertible and obvious proof or evidence to the contrary” (1994, 765). So in order to explain, at the psychological level, in virtue of what people have delusions, cognitive neuroscientists will have to resort to the content of beliefs and, hence, to representations. But, as we have seen, a mechanism approach that embraces M1–M4 will only be able to capture the formal aspects of the relevant representations, not all of their content nor all of their significance.

Psychopathology introduces another feature that raises problems for a mechanistic reconstruction of cognitive neuroscience. This feature is already introduced by our example of delusion, namely, the fact that many psychopathological phenomena are classified normatively. A variety of norms are referenced in the classification of psychopathology, some based on statistics or natural design (“abnormal”), some that are social or prescriptive (“bizarre,” “inappropriate,” “deviant”) and some that are epistemic (“unreasonable fear,” “poor insight,” “nonresponsiveness to evidence”). The point is that if we are interested in providing a constitutive explanation of some piece of psychopathology, part of the explanation must include an account as to why the cognition or behavior merits the evaluative label. But such an account will involve a comparison of the cognition or behavior with a norm and determine whether the norm is statistical, based on natural design, based on community standards, or epistemic. Any such account will extend well beyond what one would naturally assume to be the mechanism of (or the breakdown of the mechanism of) the cognition or behavior in question.

In sum, then, there are several reasons why constitutive explanation in cognitive neuroscience is more than mechanistic explanation and, hence, why the mechanism approach is unable to reconstruct all aspects of cognitive neuroscience. First, the constitutive explanations of human cognition provided by cognitive neuroscience make reference to representations and computational processes involving representations. But it is not possible to supply the required multilevel, part-whole constitutive mechanistic explanation of all aspects of any given cognitive capacity because representations are not fully constituted (realized) by lower-level neural entities or activities in the relevant part-whole mechanism hierarchy. In particular, while the neural realization of the representation bearers involved in any given capacity will be located within the relevant part-whole mechanism hierarchy, the realizations of neither the content nor the significance of those representations will be entirely so located.

Another reason that the mechanism approach cannot handle all aspects of cognitive neuroscience is that any constitutive explanation of a pathological cognition or behavior must include an explanation of the criteria by which the cognition or behavior merits an evaluative label. But such explanations require a comparison with some norm and neither the norms nor the relation to those norms will fall within the mechanism hierarchy underlying the cognition or behavior in question.