Functions,levels,and mechanisms: Explanation in cognitive science and its problems

Vertical expansion I: A reductive model of the relationship between psychology and neuroscience

Computational functionalism is the philosophical framework of classical cognitive science. At its appearance it was welcome as an attractive metaphysical position which put forward a conception of the mental as a conceptual domain that was epistemologically autonomous though ontologically dependent on neurophysiological facts. But since the early 1980s this supposed strength became the target of a twofold criticism. At an explanatory level, functionalism was blamed for being “anti-biological”: that is, for considering neural facts quite irrelevant to the study of mental processes. From a strictly ontological point of view, it was accused of fostering a view of the relationship between mind and brain that was inadequate to support a scientific and specifically materialistic view of mental phenomena.

As to the explanatory criticism, it can be definitely said that antibiologism was not intrinsic to computational functionalism. This is quite clear if one takes into account the importance that David Marr assigned to the neural constraints on mental computations. ¹ But as a matter of fact during the 1960s and 1970s the undeniable gap between the problems and tools of neuroscience and those of AI and cognitive psychology led some cognitive scientists to extrapolate from functionalism the idea that the knowledge of the brain was of scarce or even no use to understanding cognitive processes. In doing so, they gave life to that anti-biological involution of functionalism that Dennett (1998) has ironically termed “High Church Computationalism” (p. 48). During the 1980s, however, the realization was dawning that the gap between neuroscience and classical cognitive science had been remarkably narrowed (see Bechtel, 1988, pp. 86–87; P.S. Churchland, 1996, p. 381), for neuroscience had expanded upwards, from molecular and cellular levels to the systems level; and if now it could work at upper levels of organization, taking into consideration the information processing of large neuronal populations, its important role in cognitive science was no longer deniable. And so it has been: in the last 30 years, cognitive science expanded vertically into the brain, putting neuroscience at its forefront (see Bechtel et al., 1998, §3.2; Bickle, Mandik, & Landreth, 2010, §1).

The relationship between functionalist metaphysics and the development of a strongly brain-oriented cognitive science is more complex and less direct. This is because cognitive science tends to stay aloof concerning metaphysical questions (more on this below); but also because two different ontological claims are associated with functionalism. The first claim, which is the genuine formulation of functionalism, is that psychological states are defined or individuated by their causal roles. The other claim, which is not inherent to functionalism as such, is that psychological states are not identical with neural states; rather, the former are realized by the latter. This distinction is the rationale for Kim’s (1998) reductive functionalism, in which psychological states are individuated by their causal roles and identical with neural states. Kim sees the psycho-neural identity as a non-renounceable requirement for materialism, and the error that he intends to correct lies not in the functionalist conception of psychological states, but lurks in the doctrine of psycho-neural supervenience, which a functionalist is not forced to endorse. ² As a matter of fact the doctrine of supervenience has often kept company with functionalism, mostly owing to the influence of Fodor’s CRTM, but there is no necessary connection between functionalism and supervenience.

Kim’s neo-reductive functionalism (or, better, functionalist neo-reductionism) is one of the outcomes of the metaphysical criticism of computational functionalism. However, Paul M. Churchland’s eliminative-reductive version has been much more influential on cognitive science, also in virtue of his intensive exploitation of one of the key players in the vertical expansion of cognitive science, namely connectionist/parallel distributed processing cognitive modeling.

In his recent “Functionalism at Forty” (2007), Churchland makes clear that he endorses two assumptions of classical computational functionalism: (a) “that cognitive creatures are indeed engaged in computing complex functions of some sort or other” and (b) “that these computational activities … can be realized in a diversity of physical substrates” (p. 36). In contrast, he attributes blame to the classical functionalist program for failing to distinguish the level of cerebral matter from the level of cerebral architecture. A functionalism that aspires after biological plausibility needs to view our knowledge of the functional structure of the brain as a source of constraints on computational modeling. From this point of view, the strengths of artificial neural networks (ANNs)—capacities of learning and self-organization, flexibility, robustness in the presence of perturbations, capacity for dealing with such low-level tasks as the processing of sensory inputs and motor outputs—depend on just those structural features of computation (high parallelism as opposed to von Neumann’s sequential processing) which are inspired by how the brain works.

According to Churchland, this gives rise to a deep difference between classical and connectionist computational functionalism. Assuming as a model of mentation forms of thinking that lend themselves to being codified in formal models such as deductive logic, CRTM endorses a “linguistic-rationalist tradition” in the study of human cognition, which sticks to folk psychology and intentionalist philosophy of mind in taking agents to represent the world through sentence-like structures and to perform computations that mimic logical inferences. By contrast, connectionist computational functionalism is driven by the functional organization of the brain, which “represents the world by means of very high-dimensional activation vectors, i.e., by a pattern of activation levels across a very large population of neurons” and “performs computations on those representations by effecting various complex vector-to-vector transformations from one neural population to another” (Churchland & Churchland, 1998, p. 41).

The availability of a brain-like computational modeling that breaks with the “propositional kinematics” and “logical dynamics” of folk psychology leads to a reversal of the Fodorean approach to the autonomy of psychology issue. Fodor (1974) had famously claimed that computational psychology is autonomous from neuroscience insofar as its states and processes cannot be reduced to neurobiological states and processes. Churchland accepts this claim, but only to draw an antithetical implication from it. For he thinks that the irreducibility of symbolic computational psychology—owing to its being a formalization of propositional-attitude psychology—is a strong case for its elimination. However, the elimination of folk-psychological mental states will not cause the end of psychology, provided that the latter posits only theoretical constructs that, like the representations/activation vectors and the computations/vector-to-vector transformations, can be projected onto neurological states. After the eliminative stage, the new connectionist computational psychology and neuroscience will co-evolve until they are unified by an approximate microreduction.

Vertical expansion II: Explanatory pluralism and mechanistic analysis

From Churchland’s view, therefore, the approximate microreduction of psychology to neuroscience is the pay-off of the substitution of subsymbolic distributed representations for language-of-thought style representations. But how plausible is his eliminative-reductive model of the co-evolution of psychology and neuroscience?

At least from the mid-1980s, objections to the reductive approaches to interlevel connections have been raised by the advocates of “explanatory pluralism,” a position in the philosophy of science holding that “theories at different levels of description, like psychology and neuroscience, can co-evolve, and mutually influence each other, without the higher-level theory being replaced by, or reduced to, the lower-level one” (Looren de Jong, 2001, p. 731). ³ The focus is on the growth of explanatory resources; this allows the pluralist to stay aloof both from the reductionist obsession for ontological parsimony and unification of science, and from the claim for strong autonomy of the special sciences theorists.

Against the reductionist claim that when lower-level explanations are completed, the higher-level explanations stop being causally explanatory, ⁴ explanatory pluralists deny the existence of a fundamental explanatory level, and argue that higher-level entities continue to play a causal and explanatory role even when lower-level explanations are complete. In this perspective, then, the most serious shortcoming of the reductionist conception of the relation between lower and higher levels is its unidirectional nature: since it assigns to lower levels a priority, when the higher-level and lower-level theories fail to map onto one another neatly, the blame lies exclusively on the upper-level one (see McCauley, 1996, p. 25). By contrast, the pluralistic perspective is bidirectional: the higher-level theory should be subjected to revision in light of the findings of the lower-level theory, and vice versa. Accordingly, explanatory pluralism leads us from the reductive approaches to interlevel connections to the integrative ones (see Bechtel & Hamilton, 2007, §§ 5–6). Here our focus will be on Carl Craver’s (2007) model of the integration of levels of mechanisms, which is a brilliant synthesis between Cummins’ (1983) theory of functional analysis and the blossoming literature on mechanisms.

Mechanistically explaining some phenomenon consists in revealing its internal causal structure, namely in specifying the mechanisms that produce it. According to Machamer, Darden, and Craver’s (2000) often-cited definition, mechanisms are collections of entities and activities organized in the production of regular changes from start or setup conditions to finish or termination conditions. Moreover, mechanisms have a spatial and temporal organization that explains how they carry out their activities. ⁵ Finally, mechanisms have a hierarchical organization into levels, in which lower-level entities and activities are components of higher-level entities and activities; and hence a mechanistic explanation is intrinsically multilevel. More precisely, a multilevel mechanistic explanation of some item’s activity describes it according three perspectives. The “isolated” perspective is a description of an item’s activity (the φ-ing of some X) in isolation from its context (level 0); the “contextual” perspective describes X (and its φ-ing) in terms of its contribution to a higher-level (level +1), composite mechanism; and the “constitutive” perspective describes X’s φ-ing in terms of its lower (level −1) mechanism: that is, “detailing the organized entities and activities that constitute X’s φ-ing” (Craver, 2001, p. 67).

To illustrate the multilevel nature of mechanistic explanation, Craver (2007, pp. 165–170) examines the development of the explanations of Long-Term Potentiation (LTP) and spatial memory. ⁶ He distinguishes at least four levels. At the top of the hierarchy (the behavioral-organismic level) are memory and learning, which are investigated by behavioral tests. Below that level is the hippocampus and the computational processes it is supposed to perform to generate spatial maps. At a still lower level are the hippocampal synapses inducing LTP. And finally, at the lowest level, are the activities of the molecules of the hippocampal synapses underlying LTP (e.g., the N-methyl D-aspartate receptor activating and inactivating). These are “mechanistic levels” or “levels of mechanisms”: the N-methyl D-aspartate receptor is a component of the LTP mechanism, LTP is a component of the mechanism generating spatial maps, and the formation of spatial maps is a part of the spatial navigation mechanism. Integrating these four mechanistic levels requires both a “looking up” integration, which will show that an item (LTP) is a part of a upper-level mechanism (a computational-hippocampal mechanism); and a “looking down” integration, which will describe the lower-level mechanisms underlying the higher-level phenomenon (the molecular mechanisms of LTP; see Craver, 2005, p. 390; 2007, pp. 256–258).

Thus Craver’s model of mechanistic explanation offers the chance to go beyond the controversy between reductionists and autonomists, and at the same time to comply with some requirements of both parties. In the hierarchy of levels the constitutive perspective accomplishes the reductionist “looking down,” but without contrasting with the contextual perspective, which coincides with the autonomist “looking up” (see Bechtel, 2009; Polger, 2004, pp. 196–202). In fact, it is only the integration of the contextual, isolated, and constitutive perspectives that can lead to “an ideally complete mechanistic explanation.” ⁷ Accordingly, cognitive psychology that investigates spatial memory as a cognitive phenomenon cannot be reduced to neurobiology; each discipline preserves its own autonomy. But they integrate each other through the contribution that they make to the discovery of the mechanisms that produce the phenomena under investigation.

Horizontal expansion I: Dynamicism and the radical embodied cognition thesis

Craver’s model of mechanistic explanation provides a view of the interlevel connections that fits the epistemic practices of cognitive neuroscientists better than P. M. Churchland’s eliminative-reductive model. As we will now see, however, the mechanistic approach has been attacked by some defenders of the dynamical approach to cognitive modeling, namely one of the major driving forces behind the horizontal expansion of cognitive science toward the natural and social world.

A feature of classical cognitive science is individualism (or internalism), the claim that psychological processes can be studied by taking into consideration only the “intra-cranial,” intrinsic properties of those processes, and abstracting from all the environmental, extrinsic variables. Moreover, since psychological states and processes are defined by classical cognitive science as high-level functional description of neurological states and processes, it can be said that the mind ontologically depends on the brain alone—it is the physical brain that performs or realizes mental processes.

The so-called “horizontal” expansion consists in rejecting both individualism and the metaphysical claim that the mind depends on the brain alone. On the one hand it is argued that, at least in some cases, the mind cannot be studied by bracketing the (physical and social) environment in which it works. On the other hand it is noted that minds are not disembodied entities: mental processes are in the first place control systems of a body that moves, acts, and in so doing feeds back into brain and mind. In this framework, research programs very different from each other have advocated an externalist account of explanation, according to which an agent’s cognition cannot be understood without taking into consideration its always being embodied and world-embedded.

But explanatory externalism comes in various forms. For example, the so-called “sensorimotor paradigm” is a thread of externalism that highlights the indissoluble nexus between cognition and agency. ⁸ Here the emphasis is on the embodied nature of cognition, and it is to be distinguished from that much more radical form of externalism that removes the distinction between agent and environment. This was put forward by some proponents of the dynamical approach to cognition, which we will now examine.

The application of tools of dynamical systems theory to psychological phenomena has been presented as the advent of “a third contender” in the debate on the foundations of cognitive science (see Eliasmith, 1996). In this connection, a standard reference is van Gelder and Port (1995), which was the first major presentation of the dynamical approach to cognition. According to the authors, “to see that there is a dynamical approach is to see a new way of conceptually reorganizing cognitive science as it is currently practiced” (p. 4). Such a reorganization takes a stand against not only classical computationalism but also the connectionist one—and this despite the fact that the connectionists were the first to apply dynamical systems theory to the study of cognition. However, van Gelder and Port argue, the limit of connectionism lies in the use of dynamical systems tools within a paradigm that is still computationalist and representationalist, though in a brain-like variant. The dynamicist wants to go beyond.

First, as we have mentioned above, the dynamicist dissolves the boundary between the cognitive system and the system’s environment. Coupling between the equations describing a cognizing system and those describing the environment gives rise to complex “total system” behaviors. In this perspective, “the cognitive system is not just the encapsulated brain; rather, since the nervous system, body, and environment are all constantly changing and simultaneously influencing each other, the true cognitive system is a single unified system embracing all three” (van Gelder, 1995, p. 373).

Second, the dynamicist gets rid of the mechanistic and computationalist explanation. The dynamicist expansion into the environment implies an explanatory model very different from the mechanistic one underlying the vertical expansion. In the 1950s the appeal to the mechanistic explanatory strategy by the early cognitivists was the logical conclusion of the battle waged against behaviorism and mathematical psychology, which conceived psychological explanation as discovery of laws or mathematical regularities in behavior (see Bechtel et al., 1998, p. 96). The dynamical approach, however, relaunches the covering law conception of explanation. ⁹ The dynamical analysis identifies the critical variables characterizing the state of a system and attempts to construct laws (a set of differential equations) to account for the system’s trajectory through state space. The system can no longer be decomposed into subsystems (modules) that involve computations on representations. Consequently, the dynamical explanation is seen as incompatible with the explanatory style of the computationalist mechanism (see, e.g., Chemero & Silberstein, 2008, pp. 11–13).

Dynamicism, then, puts forward “the radical embodied cognition thesis”: to understand the complex interplay of brain, body, and environment we do not need either the concepts of internal representation and computation or the mechanistic decomposition of a cognitive system into a multiplicity of inner neuronal or functional subsystems; all we need are the analytic tools and methods of dynamical systems theory (see Clark, 1997, p. 148; 2001, pp. 128–130). We think, however, that in this form the dynamicist project is not a “third contender” in the controversy on the foundations of cognitive science but, rather, the denial of the possibility of such a science—to the extent, of course, that we are right in claiming that (some form of) computational functionalism is at the core of the very idea of a cognitive science.

Horizontal expansion II: Two reformist projects

A reformist perspective challenged the dynamicist obituary for cognitive science. It uses the objections to the individualism of classical cognitive science as guidelines to reconstruct the conceptual bases of cognitive science.

Andy Clark is a leading advocate of reformism. He believes that the computational and representational framework can be reconstructed making due allowances for the embodied and world-embedded character of natural cognition but without collapsing into the radical embodied cognition thesis. Accordingly, Clark pursues the transformation of that framework into just one component in a three-tiered explanatory strategy: (a) a dynamicist account of the gross behavior of the agent–environment system; (b) a mechanistic analysis, describing how the components of the agent–environment system interact to produce the collective properties described in (a); and (c) a representational and computational account of the components identified in (b) (Clark, 1997, p. 126). Clark calls this tripartite explanatory strategy “minimal representationalism,” and places it within a wider theoretical framework: “active externalism” (Clark, 1997, 2003, 2008; see also Clark & Chalmers, 1998).

Unlike semantic externalism, where the mental contents of an agent are showed to partly depend on aspects of the environment which are clearly external to the agent, active externalism sees the environment as playing “an active role in constituting and driving the agent’s cognitive processes” (Lau & Deutsch, 2010, § 9). In the wake of Gibson’s ecological optics, this environment is viewed as a complex of affordances that are the source of a particular variety of inner states, namely the “action-oriented” representations, which, unlike the symbols in the language of thought, are personal (in that they are related to the agent’s needs and the skills that he or she has), local (in that they relate to the circumstances currently surrounding the agent), and computationally cheap (compared with Marr’s rich inner models of the visual scene).

However, action-oriented representations are only a representational “genus.” Clark rightly notices that the concept of inner representation was introduced in cognitive science to account for cases in which a cognitive system must coordinate its behaviors with environmental features that are not always reliably present to the system. In such cases the cognitive system is able to decouple from the external environment and act in an offline fashion by creating some kind of inner item that stands in for the absent phenomena. These inner stand-ins are what cognitive scientists have termed “inner representations” (see Haugeland, 1998, p. 172). Such cases of environmentally decoupled cognition are really a tough nut to crack for the anti-representationalists, who are concerned exclusively with cases of “adaptive hookup”: that is, cases in which “the inner states of a system [e.g., a sunflower, or a light-seeking robot] are supposed to coordinate … its behaviors with specific environmental contingencies” (Clark, 1997, p. 147). Such cases of adaptive hookup, however, cannot ground a general anti-representationalist argument: they are not sufficiently “representation hungry” (Clark & Toribio, 1994).

In light of these considerations, Clark replaces the classical notion of mental representation with a continuum of representational genera. At one end of the spectrum there are the inner states that border the simple causal correlation and environmental control. At the other end of the spectrum we find the type of inner stand-in that allows us to deal with the representation-hungry problems. Then between these two poles are the action-oriented representations.

According to Clark, therefore, depending on the coupling or decoupling between agent and environment, one must appeal to the dynamical non-representational explanation or the representational one, respectively. It can be objected, however, that this implies a division of labor between the two styles of explanation, and not their complementarity; as a result, they cannot be the tiers (a) and (c) of a single explanatory strategy, as Clark would want.

Moreover, it is not clear how Clark’s model of explanation can motivate the integration between the tiers (a) and (b): if the interactions between the components of the global system can be described in mechanistic terms, is there still need to conceive the system in dynamicist terms? Some clarifications about this question come from William Bechtel and his collaborators’ work (see, e.g., Abrahamsen & Bechtel, 2006; Bechtel, 1998, 2001, 2008, 2009, 2011; Bechtel & Abrahamsen, 2010; Bechtel & Richardson, 1992, 2010; Kaplan & Bechtel, 2011).

Bechtel and Richardson (2010, ch. 7) note that in the early stage of the process of developing mechanistic models scientists often assume that the processes that they are considering are performed serially. But when it is not possible for scientists to develop a linear model that is adequate to the phenomenon, they start to introduce feedback loops and other non-linearities in their attempts to develop adequate models. The outcome is what the authors define as functionally integrated systems.

Again, as in the case of representation, a continuum emerges. At one end of the spectrum we have fully decomposable (or highly modular) systems, which are composed of subsystems that are completely independent except for the mutual exchange of outputs (this is the case with Fodor’s encapsulated modules). If the interactions among the subsystems are weak but not negligible, the system is nearly decomposable. As the complexities of interaction among parts increase, the explanatory burden shifts from the parts (or, more precisely, the interactions within subsystems) to their organization (i.e., the interactions between subsystems). Thus we reach the other end of the spectrum, where we find holistic systems whose components are functionally equivalent and hence interchangeable. In between the nearly decomposable systems and the holistic ones, there are the integrated systems. In these systems, unlike the holistic systems, it is possible to isolate different parts that make distinctive contributions but also give rise to a complex set of interactions that are nonlinear, and hence much stronger than those of a nearly decomposable system. ¹⁰

Now, both Bechtel (2001) and Clark (1997) suggest that psychobiological cognition is likely to take up the intermediate space between nearly decomposability and holism, namely that of integrated systems. This allows them to denounce as spurious the opposition between an ultra-modularist conception of the parts of biological mechanisms as totally isolated parts and a radically holistic view that rejects the very possibility of decomposing the mind-brain.

Interestingly, some radical holists are neuropsychologists who use dynamical systems tools to revive the Gestaltist principle that higher-level cognition rests on the dynamical organization of the cortex as a whole (see Bechtel, 1997, § 3; Zawidzki & Bechtel, 1996, § 4). For example, Guy van Orden criticized the double-dissociation and neuroimaging studies that explain cognitive activities in terms of single causes, and promoted the approaches that rely on the notion of continuous reciprocal causation (see van Orden, Pennington, & Stone, 2001). This is an integrally holistic concept since it “implies that each and every component of a system contributes to every behavior of the whole system … . Each component affects every other component, to the extent that their independent contributions cannot be sorted out in the behavior of the whole” (van Orden & Paap, 1997, p. S93).

A piece of evidence that is supposed to confirm this holistic view is the presence in the brain of a vast number of feedforward, feedback, and collateral connections. But Bechtel has repeatedly warned that the hypothesis of the neurobiological reality of holism is scarcely plausible. He pointed out that important contrary evidence comes from the studies by David van Essen and his collaborators, which have almost completely mapped the areas of the Macaque monkey’s visual system over the last two decades (see, e.g., van Essen, 2004). The researchers have identified over 30 different areas in the macaque visual cortex and more than 300 connections between these areas; and the tool-kit of dynamical analysis can be very useful to model this vast number of feedforward, feedback, and collateral connections. However, although these regions are highly interconnected, we can still determine what each area contributes to visual information processing. That is, it is not a holistic system, but an integrated one. Indeed, Bechtel takes this work to be an exemplar of mechanistic analysis of how the brain performs a cognitive function. And in an integrated system, mechanistic analysis “provides the foundation for dynamical analysis” (Bechtel, 2001, p. 483) since the latter has explanatory force only insofar as it describes “the operations of the underlying mechanism” (Kaplan & Bechtel, 2011, p. 443), only to the extent that it reveals “aspects of the causal structure of a mechanism” (Kaplan & Craver, 2011, p. 602).

This attempt to reconcile dynamical modelling and mechanistic analysis is very attractive since it provides a way of avoiding a fracture between explanatory styles that would bring cognitive science to collapse. However, a bothersome issue arises that we shall discuss, together with a pair of further problems, in the next section.

Computationalism and dynamicism: A pacific coexistence?

The harmonization of a mechanistic model including computational explanations (henceforth M-C models) with the dynamical one faces the following problem.

As we saw in the previous section, the most plausible picture of how mechanistic and dynamical explanations coexist is provided by Bechtel’s notion of an “integrated system.” Integrated systems have parts (subsystems) that are individuated according to a mechanistic principle; at the same time, however, since the inter-relations among parts are non-linear (e.g., they cannot be reduced to simple input/output connections), their global organization requires a dynamical description: that is, the whole system turns out to be a dynamical system. Note that in this picture the burden of explanation is carried by the mechanistic component, since the mechanistic decomposition of the system in parts is a non-negotiable condition. In other words, dynamical explanations make sense only against a mechanistic background—their role consists only in, so to speak, “filling the (explanatory) gaps.”

We are quite sympathetic with this point of view, since, in the case of cognition, computational models seem to possess a higher explanatory force (see, e.g., Kaplan & Craver, 2011), and computational models are mechanistic. Yet, even in Bechtel’s picture, how the integration actually works remains to a large extent obscure. First, computational explanations require (at least preferably) modular subsystems, whereas, according to Bechtel’s model, the richness of interactions makes it difficult to regard subsystems as modules (and, if all parts were modules, of course we could dispense with dynamical explanations completely). Second, it is by no means obvious how to link the output of modules to the relevant dynamical variables of the whole system. The notion of integration is required to put together, in some way, computational descriptions and dynamical descriptions; by contrast, in the current view, the two kinds of explanations are merely alternative.

In short, the appeal to dynamical models is typically invoked for integrated (sub)systems in Bechtel’s sense, which are very weakly modular, since each of their parts is influenced by the activity in some other parts of the system. Is this degree of modularity sufficient for the standard required by the M-C model?

The answer is hardly positive. Carruthers (2006), for instance, argues that an M-C explanation requires constraints on the concept of part (or module) far more demanding than what is required for the notion of integrated system, namely (some form of) informational encapsulation and massive modularity. Although Carruthers’s argument is highly controversial, it seems difficult to deny that the typical decompositional method of the M-C model is more reliable the more the relevant mechanisms tend to be encapsulated (Bechtel, 2003). ¹²

There is a further difficulty in the integration between dynamicism and the M-C model coming from a tension between vertical and horizontal expansion. While the former involves a view of mind as a collection of functions produced by the brain—even if the pluralistic picture is not reductionist in spirit—the latter downsizes the role of the brain, both ontologically and epistemologically. On the ontological side, the externalist philosophy underlying the horizontal expansion denies that the mind depends ontologically on the central nervous system alone; correspondingly, on the epistemological side, explanations of mental phenomena cannot be found in the cerebral bases alone. In the dynamicist (and nonmechanistic) view, in particular, the role of the brain blurs in a conception of reality in which entities are undifferentiated variables and processes (a Machian view, on some aspects).

Furthermore, the assumption that embodiment and embeddedness are, taken together, a unique, coherent doctrine is quite dubious, astonishingly taken for granted by its supporters. Suffice to say that, if one defends the embodied nature of the mental, one should be committed to the ontological dependence of mind on body (psycho-physical supervenience), against the metaphysical externalism more or less explicitly endorsed by supporters of embeddedness. Even more surprising, embodiment is often described by saying that certain cognitive tasks, such as imagery or language understanding, involve the activation of motor areas. So, in this view, motor areas are not a part of the brain!

Is it possible to find a coherent synthesis of the two kinds of expansion? Most likely yes, but provided that both sides weaken their more demanding claims. More specifically, the synthesis requires a reasonable compromise in which the epistemological demands of embodiment and embeddedness are vindicated, but at least one untenable metaphysical thesis is given up. It is the idea that mind is literately constituted by extra-bodily items. Let us explain.

Arguing for the externalist character of mental processes explanation is correct to the extent that it is legitimate and probably necessary to use, in the description of a competence theory (a computational theory in Marr’s sense), a teleological-intentional vocabulary, which makes reference to aspects of the external environment. In order to say what, for instance, the goal of vision is, or what is computed by each of its subsystems, it is quite sensible to mention environmental properties. However, this does not imply that a computational state supervenes on external factors, over and above the internal factors, as the metaphysical externalism claims. In fact, a computational state can carry information about an external event or property, but this does not make that state an external “object.” A representation is one thing; what it represents is another. In counterfactual terms, an external difference implies a mental change in an agent only if the difference is detected by (some parts of) the agent (see Egan, 1995; Patterson, 1996).

However, there are good reasons to believe that, as we have seen above, some mental processes are characterized by the direct, “real-time” involvement of external factors. In these cases it seems correct to say that these factors are metaphysically constitutive of the relevant processes and states (see Clark, 2008; Wilson, 2004). In this sense the classic idea according to which cognitive processes supervene on internal cerebral states can no longer be taken for granted. Briefly stated, one can neither regard psycho-neural supervenience as a dogma nor jump to externalism across the board. ¹³

It seems to us that granting this partly internalist point does not undermine the central point of embeddedness: that is, the influence of environment on representations. After all, the supporter of horizontal expansion need not endorse more or less queer metaphysical ideas about the nature of the mental, since acknowledging the role of the environment does not require the extension of mind beside the body. Handling metaphysics cautiously is also the attitude underlying our strategy for addressing the second family of problems we are going to discuss in the following section.

Theoretical entities and scientific explanation

Whatever the attitude as regards the possibility of integrating computationalism and dynamicism, there is a question to clarify concerning the relations between the explanatory levels of the “stack.” It is a question that would arise even within a pure M-C explanation.

A good way to introduce the issue is the following. One among the criticisms directed toward the CRTM is that it is committed to some sort of dualism. The CRTM—so the charge goes—postulates the existence of mental entities, the expressions of Language of Thought (LoT), thereby reintroducing a dichotomy between mind and brain (see, e.g., Putnam, 1999, part II). Even if this charge is, as we shall see below, unmotivated, the M-C model faces the following, general challenge: since each explanatory level postulates theoretical entities of a certain kind, what is the nature of these entities, and what are the relations between them and the entities belonging to the higher and lower levels?

In the specific case of the CRTM, the charge of dualism can be rejected since the algorithmic level is a description level, not a further ontological level. The existence of LoT is an empirical conjecture waiting for (dis)confirmation, but if, ex hypothesis, there actually are LoT expressions, then they are necessarily realized by neurophysiological patterns. In other words, a LoT formula is identified with a class of functionally equivalent neurophysiological patterns. The problem is the tenability of the claim that there are patterns of neural activation whose evolution can be interpreted as a transformation of (semantically transparent) informational structures according to compositional rules is warranted (see below). The question is, in other words, whether bottom-up constraints allow confirmation of a hypothesis, the existence of LoT, which rests only on explanatory requirements specified at a high level.

Today we are extremely far from being able to confirm (or disconfirm) empirically the existence of LoT, and one reason for this difficulty is the big distance separating the description level of CRTM from the neurophysiological level. It is this distance which feeds the impression that CRTM is a kind of dualist mentalism. Nevertheless, computational theories can escape the charge of being involved in a metaphysical reification of their theoretical construals.

The non-ontological character of levels allows us to understate the problem of relations between levels in general. The conceptual apparatus involved in the M-C model associates, at each descriptive/explanatory level, functions, mechanisms (in the computational case: algorithms), and representations (= informational structures). A function, however, is not a piece of a system; it is rather something a piece of system does, and we can say that a system does something at different levels—or, more properly, there are different parts of a system which do something at different levels. Therefore functions are not “real things,” and do not raise any metaphysical embarrassment. ¹⁴

The relations between mechanisms and representations at level X–1 and mechanisms and representations at level X are either realization or constitution relations, or perhaps (in a few cases) identities. In other words, although an algorithm or a representation is individuated in abstract terms, they do not raise, at their level, ontological problems. Their “reality” is not something added to the reality of their realizers. It is easier to see this in the case of ANNs, where an activation pattern can simply be identified with a physical state, and the distinction between algorithms and representations is blurred. But the same is in principle true in the case of classical algorithms and representations, which are characterized in a more abstract way. Marr’s primal sketch, for instance, is a mathematical description of activation patterns of neurons in the V1 brain area; and the zero-crossing algorithm, which plays an important role in the construction of the primal sketch, is a mathematical description of the behavior of such cells. One can certainly prefer connectionism to CRTM, but not because the former is ontologically less embarrassing. The problem is different, and concerns the status of folk psychology: that is, the plausibility of certain top-down constraints (more on this later).

Therefore, we can conclude that the M-C model, properly interpreted (with “de-ontologized” levels as explained above), does not raise metaphysical puzzles and matches the usual epistemological standards imposed on scientific explanations. However, there still is a bothersome issue calling for a clarification: what is the relation between the pre-theoretical mind—the level of commonsense explananda—and the top level of the explanatory stack? Can we say that the model of interlevel relations based on the notions of constitution and realization accounts properly for the mental capacities and intelligent behavior of a human agent, as they are described in the naïve representation of ourselves, as beings who possess beliefs and desires, make plans, have perceptual emotional experiences, and so on?

There are two questions calling for a proper answer:

Now, one could argue that the inclusion in the stack, at the top level, of a computational theory in Marr’s sense is exactly the answer to question 1, insofar as the appropriate explananda are the capacities corresponding to the “boxes” (e.g., edge detection, surface processing, motion processing, etc.) postulated by the computational theory. This is, for example, Dennett’s (1987) position when he opposes his “intentional systems theory” to folk psychology.

Dennett’s position is a classical way of regarding the relation between the scientific image and the commonsense view of the world: even if science originates from commonsense and moves from the questions posed by commonsense, the construction of scientific theories requires amending it (more or less conspicuously). Is this view also applicable to the case of psychology?

The question of the relation between the commonsense mind and cognitive sciences has been deeply discussed by Bermúdez (2005), who gives it a distinctive label: “the interface problem.” More precisely, the interface problem is the problem of clarifying how typical subpersonal explanations in cognitive sciences, whatever their specific form, are related to commonsense psychology. The problem has not been taken into serious consideration within cognitive science. By contrast, according to Bermúdez, the way in which Dennett assesses this relation (a similar approach has been proposed by Baker, Davidson, and Putnam) is flawed, because it cuts the link between commonsense explanation and scientific explanation, making the ordinary mind “autonomous” from science.

The distinctive difficulty posed by the interface problem can be expressed through the following dilemma: either one takes seriously the way commonsense thinks of the mind, as CRTM does, but in this case the unification with lower-level explanations turns out to be hard, or one does not, but in this way we are unable to account for some shared and sensible intuitions, such as, first of all, the idea that mental states are causes of behavior.

Although Bermúdez is well aware of being unable to offer a definite solution, he sketches out multiple strategies to deal with the problem: (a) showing that the relevance of folk psychology to human behavior is strongly overstated, (b) leaving room for a plurality of theories and explanatory levels, and (c) endorsing the massive modularity thesis, which Bermúdez associates with a blurring of the distinction between perception and cognition. ¹⁵

Whatever one’s perspective on this “cocktail” of strategies, two points should be highlighted. First, Bermúdez’s proposal fits more or less well the current tendencies in cognitive science. In this sense his proposal could be regarded as a rational reconstruction of what cognitive scientists do today, as Fodor’s work was 30–35 years ago. Second, according to Bermúdez, there are no metaphysical aspects involved in the interface problem at all. When one asks questions about how interlevel relations should be characterized, the issue at stake is not the nature of mental states, but, rather, what kind of relation there is between explanatory practices, such as personal commonsense psychology, on the one hand, and subpersonal cognitive psychology, on the other. Thus, according to Bermúdez, having a good metaphysics of mind is neither a necessary nor a sufficient condition for solving the interface problem. However, even if, on the one hand, this metaphysical “deflationism” seems to be fruitful when applied to the characterization of the M-C model, on the other hand, when the problem concerns the tip of the stack—that is, it consists in the relation between the top level and the commonsense mind—it is far from clear that giving up metaphysical questions is actually an available option. It could not be since the serious difficulty of the interface problem comes arguably from the necessity of somehow relating personal predicates to subpersonal predicates, and the kind of relation between these two levels cannot easily be brought back to one among the above-mentioned candidates (supervenience, identity, etc.). Perhaps doing cognitive science without doing (or assuming) at least a bit of metaphysics is not possible. If cognitive science leaves unresolved the interface problem, it ends up accepting a dualism between natural sciences and human sciences, between scientific reality and “the world of life,” and this is in danger of rendering scientific explanations of mental phenomena irrelevant: they would be regarded as unable to explain human nature, despite their purported aim.

Bermúdez’s account seems to take for granted the view of explanation based on a stack of explanatory layers. However, it could be argued that a genuine explanatory pluralism, and, perhaps, even Bermúdez’s idea of multiple strategies, requires a more flexible picture. The notion of stack indeed involves the idea of a hierarchy of levels, suggesting that there is a “top” (e.g., computational psychology) and a “bottom” (e.g., neuroanatomy). However, this view cannot vindicate the sophisticated nature of relations between explanatory models. Moreover, the richness and variety of the phenomenological reality we want to account for does not fit well with a rigid hierarchical model; cognitive conscious states are different from sensory states, and these also differ on several aspects. Therefore we cannot assume without discussion that every kind of mental state/process will be accounted for by the same collection of explanatory layers. In this sense, borrowing from Nancy Cartwright (1999), we can suggest that the most appropriate metaphor of explanation in cognitive science is the patchwork rather than the stack. This would also fit better the proposals aiming to integrate different explanatory models (such as dynamicism and computationalism) by relaxing the constraints on the overall explanatory architecture.