The Adaptable Mind: What Neuroplasticity and Neural Reuse Tell Us about Language and Cognition - John Zerilli 2021

Varieties of Modularity
Modules Reconsidered: Varieties of Modularity

4.2.1 Themes and trends

The nineteenth-century phrenologists were probably the first to emphasize the specialization of brain functions. Gall and Spurzheim (1835) hypothesized that there were “about thirty-five affective and intellectual faculties” localized in distinct regions of the brain. As almost everyone knows, however, they got the details horribly wrong, for they fallaciously assumed that the activity of a cortical faculty would be reflected in its size, and that its size in turn would be reflected in the relative prominence of cranial bumps. This led them to endorse the pseudoscientific practice of gauging personality from the shape of a person’s skull. Wrong though they most assuredly were in this respect, the idea that brain function can be mapped to local structure was not itself a bad idea. It soon received empirical support in the work of the neurologists Gustav Fritsch, Eduard Hitzig, Paul Broca, and Carl Wernicke—Broca and Wernicke being of course the first to discover the so-called “language areas” of the brain (Bridgeman 2010). Indeed, by the end of the nineteenth century, the idea of mapping the brain was well on its way to becoming the equipment of every working scientist in the field. In fact “the notion of cognitive function being subdivided and routed to different regions of the brain has remained a central idea in neural science and a fundamental principle in the clinical practice of neurology” (Pascual-Leone & Hamilton 2001, p. 431).

Corresponding to a rough division between “mind” and “brain,” one may trace the course of two distinct but parallel traditions originating in the work of these nineteenth-century neurologists. The first is a structuralist tradition whose methodology, guiding assumptions, and theoretical concerns are predominantly biological (i.e., neurological and anatomical). From a certain point of view, Fodor’s archetype could be said to fall broadly within this tradition—notwithstanding the subordinate and strictly dispensable role played by structural properties in his overall account (Anderson & Finlay 2014, p. 5; Fodor 1983, pp. 71, 98—99; Coltheart 1999)—as may both the neural network graph-theoretic module (see § 4.2.2) and neuroscience “brain module” (see § 4.3), which I come to shortly.

An alternative approach investigates questions of cognitive architecture from the standpoint of a classic computationalist or functionalist. In the guise of evolutionary psychology or “massive modularity,” for example, it “retains the Fodorian focus on computation, and with it a focus on the algorithmic (or heuristic) efficiency of purported psychological solutions to adaptive problems such as food choice, mate selection, kin identification and cheater detection” (Anderson & Finlay 2014, p. 5). It does not, however, entail specific commitments about implementation beyond those required for functional independence.¹ (See Sternberg 2011, pp. 158—159, for an overview.)

These two (ideally) complementary approaches to the mind/brain are reflected again in the central assumptions underpinning much of the effort within neuropsychology, cognitive neuropsychology, and cognitive neuroscience (Bergeron 2007). Bergeron calls these the “anatomical modularity assumption” and the “functional modularity assumption.” Recall that in Chapter 1, we provided a general definition of an anatomical module. It is worthwhile restating this definition in such a way as to reveal more clearly its relationship to a “functional” module. The functional modularity assumption

is the idea that the architecture of human cognition largely consists in a configuration of cognitive modules, where a “module” is roughly defined, following Jerry Fodor (1983), as a domain specific, innately specified, and informationally encapsulated system. . . . What this means is that human cognition can be decomposed into a number of functionally independent processes, and that each of these processes operates over a distinct domain of cognitive information. Moreover, since these processes are brain processes, to hypothesize that the capacity to do A and B depends on two distinct cognitive modules—one responsible for the capacity to do A and the other responsible for the capacity to do B—is to hypothesize that the brain processes cognitive information related to A separately from the way it processes cognitive information related to B. . . .

What makes the A module/process distinct from the B module/process is their functional independence, the fact that one can be affected, in part or in totality, without the other being affected, and vice versa. (Bergeron 2007, pp. 175—176)

The anatomical modularity assumption, then,

is the idea that the cognitive modules which compose cognition (or at least most of them) each reside in some specific and relatively small portion of the brain. . . . The anatomical modularity assumption is in fact the functional modularity assumption plus a claim about the implementation of functionally modular processes in the brain. (Bergeron 2007, p. 176, my emphasis)

Stripped to their essentials, functional modularity implies functional dissociability, while anatomical modularity implies both functional dissociability and neural localization. As I argue later, functional dissociability—functional modularity, pure and simple—represents the essence of any modular account worthy of the name.

What I have so far failed to mention, though it will in fact be crucial to appreciating the implications of neural reuse, is that cognitive modules have been generally postulated to account for higher level or gross cognitive functions; i.e., for the sorts of psychological capacities that might appear in the ontologies of cognitive psychology. Even if one restricts one’s gaze to the history of the structuralist/neurological tradition, one will not be surprised to learn that, in the main, the project of mapping function to structure has proceeded with a fairly coarse taxonomy of psychological capacities in hand. The phrenologists, for their part, merely translated the categories of Thomas Reid’s faculty psychology onto a plan of the skull (acquisitiveness, friendship, sagacity, cautiousness, veneration, etc.) (Poldrack 2010, p. 753). Broca’s postulation of a “language” area associated with motor aphasia, though no doubt empirically better supported than Gall and Spurzheim’s assumptions, hardly served to sharpen the focus on what the brain itself is actually doing when facilitating speech; for “what warrants the thought that [such] characteristics [as those found in faculty psychology] will be useful to structuring the neuroscience of behaviour and divide the brain at its functional joints?” (Anderson 2014, p. xvi). Consider Russell Poldrack’s illuminating reductio ad absurdum (cited in Anderson 2014):

Imagine that fMRI [functional magnetic resonance imaging] had been invented in the late 1860s rather than the 1990s. Instead of being based on modern cognitive psychology, neuroimaging would instead be based on the faculty psychology of Thomas Reid and Dugald Steward, which provided the mental “faculties” that Gall and the phrenologists attempted to map onto the brain. Researchers would . . . almost certainly have found brain regions that were reliably engaged when a particular faculty was engaged, . . . [and] Gall and his contemporaries would have taken those neuroimaging results as evidence for the biological reality of his proposed faculties. (Poldrack 2010, p. 753)

What reasons have we for imagining that the taxonomies of modern day psychology will fare any better in carving the brain at its true functional joints? Clinical evidence of dissociations aside (about which I shall have more to say later), widespread evidence of neural reuse strongly suggests that attempts that seek to impose upon the brain a set of categories devised (largely) autonomously of the brain, and molded from a wholly different set of considerations from those guiding brain science generally, are doomed to repeat the same basic phrenological mistake. What is needed is “the development of ontologies that let the phenomena speak on something closer to their own terms” (Anderson 2014, p. xvii).

The structuralist tradition in fact does admit of some exceptions to this questionable trend in what might even be seen by some as a clear premonition of neural reuse theories. As Bergeron’s (2007) helpful discussion reminds us, the fact that Carl Wernicke postulated a sensory speech area, often wrongly dubbed the “language comprehension area,” makes it all too easy to forget that Wernicke himself was “very resistant to postulating any cerebral centers beyond what he referred to as the ’primary’ (motor and perceptual) ’psychic functions’ ” (2007, p. 184). Wernicke could well be credited with the elaboration of an entirely original approach to the structure—function relationship in the brain. In this approach, only the sensory and motor functions are allocated distinct and dedicated neural anatomy. Higher psychological functions such as those implicated in language production and comprehension are supposed to depend on the interactions of these low-level sensory-motor systems. This arguably anticipates modern theories of reuse that predict that higher cognitive functions resolve in the interactions of lower-level elements. Bergeron certainly thinks so, and even suggests that Wernicke must have been operating with an implicit understanding of the difference between a cognitive working and a cognitive use, the distinction that, as we saw in Chapter 3, Anderson made central to his original presentation of the massive redeployment hypothesis. If Bergeron’s conjecture is correct, Wernicke’s great methodological innovation—what set him apart from the phrenologists and even his predecessor Paul Broca, for example—consisted in his cautious reluctance to infer cognitive working (i.e., essential functional contribution across all task categories, considered in isolation of neural context) from cognitive use (i.e., high-level/gross cognitive function), an inference obviously susceptible to Poldrack’s reductio.

In the same vein, the father of modern neuroscience and champion of the neuron doctrine, Santiago Ramón y Cajal, “was decidedly not a supporter of either the definition of psychological ’faculties’ or their assignment to discrete, localized neural ’organs’ ” (Anderson 2014, p. xv):

In [his] view, brain function is to be understood in terms of a hierarchy of reflexes, in the most sophisticated instances of which one responds not just to external but also to internal, and not just to current but also to stored stimuli. . . . In such a brain there can be no region for circumspection or poetic talent, for although a particular sensory experience or association may be stored in a particular place . . . the behavioral characteristics of the organism are realized only by the fluid activity of the whole system in its environment. (Anderson 2014, pp. xv—xvi)

The idea that specific circuits could be cued by various stimuli across both internal and external environments is a tolerably clear presage of the metamodal hypothesis of brain organization, which we encountered briefly in Chapter 2, and underwrites the possibility of neural reuse. (I revisit the metamodal hypothesis in more detail in the next chapter, as it bears greatly on the questions facing us there.)

4.2.2 Graph theory and network neuroscience

There is another usage of the term “module” that one often comes across in the literature. It is perhaps a testament to the immense versatility of modularity that it has descriptive utility well beyond the confines of cognitive science. Modules play an important role in fields as diverse as developmental and systems biology, ecology, mathematics, computer science, robotics, and industrial design. One interesting application of the term occurs in the study of networks, and neural networks in particular. Unfortunately, there is a danger of confusion here, because the network concept is significantly looser than the classical one in cognitive science. Thus it sometimes happens that different researchers, all of whom work in the cognitive sciences broadly speaking (including brain science), refer to “modularity” but mean different things by it.

A “network” is any organization with a weblike structure. The Internet, airline routes, food webs, and electrical grids spring immediately to mind, but these are only the most obvious examples among a great variety of phenomena displaying network design, including genetic regulation and protein interaction (Bullmore & Sporns 2012; Caldarelli & Catanzaro 2012, pp. 23—25). Networks manifest a number of important universal properties (Caldarelli & Catanzaro 2012, pp. 4—5). At the most elementary level, all networks comprise a collection of nodes (or “vertices”) and the various connections (or “edges”) between them (see Fig. 4.1). In a map of airline routes, for example, a single airport would be represented by a node and the route between any two of them by an edge. Because the focus of attention is the global structure of interactions between nodes, rather than the individual nodes themselves, the basic representational vehicle can be the same in every case; namely, a graph depicting nothing more than these nodes and their all-important interconnections (Caldarelli & Catanzaro 2012, pp. 4, 12; Anderson 2014, p. 12). In graph theory, a “module” is defined as a community of “densely interconnected nodes” where “the existence of several [such] communities is characteristic of [a] modular [network]” (Bullmore & Sporns 2012, p. 342; Caldarelli & Catanzaro 2012, pp. 89—90) (Fig. 4.1). In network neuroscience specifically, network models take the form of neural coactivation graphs, where modules are identified as communities of nodes that are functionally coactive (see later in this chapter). In the context of neural networks, then, “modularity refers to the existence of multiple communities of neurons or brain regions as defined by patterns of [functional] connectivity” (Bullmore & Sporns 2012, p. 342).

Figure 4.1 Nodes, edges, modules, and hubs in a network. Nodes are sometimes also called vertices.

(Reprinted by permission from Springer Nature, Nature Reviews Neuroscience, “The economy of brain network organization” by E. Bullmore and O. Sporns. Copyright 2018. Bullmore & Sporns 2018, p. 342.)

The point is explained very simply by Caldarelli & Catanzaro in connection with the importance of fMRI:

When humans perform an action, even one as simple as blinking, a storm of electrical signals from the neurons breaks out in several areas of the brain. These regions can be identified through techniques such as functional magnetic resonance. Through this technique, scientists have discovered that different areas emit correlated signals. That is, they show a special synchronization that suggests that they may influence each other. (Caldarelli & Catanzaro 2012, p. 27)

Furthermore,

These areas can be taken as nodes and an edge is drawn between two of them if there is a sufficient level of correlation. Also at this level, the brain appears as a set of connected elements [i.e., “modules”]. Each action of a person lights up a network of connected areas in the brain. (Caldarelli & Catanzaro 2012, p. 27)

That is, the neuroimaging data resulting from a functional connectivity analysis can be represented as a graph—a neural coactivation graph—in which nodes represent individual brain regions and edges denote the likelihood of coactivation between two nodes during a particular task (Anderson 2014, p. 12).

Why should this prove instructive for cognitive architecture? It turns out that the abstract topological features of these neural coactivation graphs frequently (if only roughly) track the functional taxonomies of cognitive psychology, cognitive neuropsychology, and the computationalist/functionalist tradition more generally (Anderson 2010, p. 303, 2014, p. 42). This sense of the word “module” therefore seems as if it might have a natural affinity with the modules to which philosophers of psychology have become accustomed. But closer inspection shows this to be a tentative link at best.

Firstly, being in effect sets of reusable (i.e., domain-general/task-selective) nodes, these graph-theoretical modules are not your typical dissociable ones (although see Pessoa 2016 for discussion); nor, for that matter, are they intended to track encapsulation, domain specificity, automaticity, or the half-dozen other features typically ascribed to modules within the computationalist tradition (Stanley & De Brigard 2016). Quite simply, the usage here is sui generis. Secondly, while there no doubt is a standard and more orthodox usage of the term “module” in neuroscience (one that moreover does offer some support to the classical conception from cognitive science, as I discuss later), its meaning is in fact much closer to what is represented by the nodes of a coactivation graph than by the communities of nodes in such a graph (see, e.g., Pascual-Leone & Hamilton 2001, p. 443; Pascual-Leone et al. 2005, p. 396; Caldarelli & Catanzaro 2012, p. 27; Fedorenko & Thompson-Schill 2014, pp. 120, 121; Zador 2015, p. 44). That is, the standard sense of “module” in neuroscience trails far more closely the idea of small, individual brain regions with discrete subfunctional profiles than it does the idea of high-level/gross-functional composites. The anomaly results from the fact that network techniques were developed independently of neuroscience and with a distinctive usage and vocabulary. When network methods were eventually adopted by neuroscientists, an idiosyncratic usage was introduced into a discipline that already had a fairly settled meaning for the term “module.” In neuroscience, “module” typically refers to a cortical column (akin to a node in the coactivation graphs just discussed), and this, as we shall see further in §§ 4.3 and 5.1, is a twentieth-century refinement of the anatomical module within the structuralist tradition.

4.2.3 Separate modifiability as the touchstone of modularity

A common objection to accounts of cognitive architecture that downplay or question the modular hypothesis is that modularity has not been given due credit for the uniquely versatile concept that it is, and that the dissenters have simply fettered themselves with an impossibly narrow and needlessly structuralist conception of cognitive architecture that is unwarranted in all the circumstances (the circumstances being the Cognitive Revolution, the fact that no one seriously denies that the mind has a rich internal structure, the unquestionable boon of functional decomposition as an effective research strategy in the cognitive sciences, etc.). Jungé and Dennett (2010, p. 278) appear sympathetic to this point of view: “A software theory of massive modularity—programs evolved to serve particular adaptive functions within brains—without commitments about implementation (unlike anatomical modularity) could survive [the evidence of neural reuse] largely untouched.” At issue here is whether a nondissociable system of some variety could be regarded as modular—whether, say, a language or norm acquisition device comprising very many smaller domain-general neural regions could in some sense be a module.² Against this suggestion is the claim that functional dissociability ought to fix a definite threshold beneath which a system cannot be regarded as modular. Here I shall contend for the latter view.

Recent developments in neuroscience have no doubt added to the luster of the “system” module, as I shall call it, and even encouraged the view that such modules represent what was always the most important contribution of modular theories to our understanding of the mind (see further in this chapter). But actually the system module has been around for a long time. Its fortunes can nowhere be more illuminatingly charted than in the annals of generative grammar. Generative grammarians are notorious for prevaricating on the issue of linguistic modularity—one can easily locate passages that would suggest the modularity in question is anatomical or at the very least functional (Chomsky 1975, pp. 40—41, 1980a, pp. 39, 44, 1988, p. 159, 2002, pp. 84—86; Pinker & Jackendoff 2005, p. 207; Fitch et al. 2005, p. 182; Collins 2008, p. 155), and others where what they seem to have in mind is little more than a “domain of inquiry”—“[t]he view that internal cognitive systems can fruitfully (for purposes of successful theory-construction) be studied independently of other such systems” (McGilvray 2014, p. 235; Chomsky 2005, p. 5). Notice that the module-as-domain-of-inquiry very effectively neutralizes the sting of neuroscientific evidence, since there is really no evidence that neuroscientists can adduce against the existence of such a module (a point to which I return later). Indeed, the system module is frequently endorsed by playing down the significance of implementation and emphasizing its “methodological value as a research heuristic” (Badcock et al. 2016, p. 11; see also Scholl 1997). But let us return to the other theme of this section, the notion of dissociability.

In a straightforward sense, a system is dissociable if it is functionally specialized—if it can (in principle) be modified without directly impeding the operation of a comparable system.³ This is a matter of degree. So if a neural system n consisting of primitives {p₁, p₂, p₃ . . . p_n} contributes some specific and functionally discrete operation f such that any element of the set {p₁, p₂, p₃ . . . p_n} is dedicated to f, n will be pro tanto dissociable; the more elements in the set {p₁, p₂, p₃ . . . p_n} are dedicated to f, the more dissociable n will be.⁴ On this understanding, a speech production center will be dissociable to the extent that its impairment has no direct effect on any system “considered with the same grain of analysis” (Carruthers 2008, p. 295) (e.g., numeracy, rhythm, speech comprehension, episodic memory, IQ, etc.), even though it might ramify to compromise a higher-level functional system that draws upon the speech production center for processing (e.g., singing, signing, etc.). In the context of neural reuse, we may presume that a working’s impairment will ramify to all higher-level functional composites in which it plays an active role; and yet so long as no other working is directly put out by such an intervention, the working remains sufficiently discrete to be regarded as dissociable. (Whether brain regions as small as workings are truly dissociable in this sense is another question. I take it up in Chapter 5.)

Notice that when spelled out in this way—and all I have done is follow through with the logic of dissociability as it is commonly understood (see, e.g., Carruthers 2008, p. 258; Coltheart 2011, pp. 227—228)—the requirement could be thought to lose much of its explanatory power. For what it entails is that the smaller and more functionally promiscuous a neural system gets—remembering that neural reuse itself implies that the typical brain region will be both extremely small and highly versatile—the more difficult to quarantine the effects of regional impairment, since those effects will presumably ramify to all affected distributed systems. An evolutionary psychologist might allege that nothing theoretically significant can follow from the fact that a tiny brain region is dissociable if its impairment will disturb the operation of many higher-level cognitive systems. It is only when modules directly implement gross cognitive functions (e.g., sentence parsing, cheater detection, face recognition, and the like) that the effects of modification can be contained in a way that makes dissociability an important constraint on cognitive theory. For then evolution itself can have a clear role to play in shaping cognitive systems by selectively modifying brain regions in a way that does not reverberate detrimentally across distributed systems. This indeed was thought to be a major argument in favor of modularity—the neat solution it offered to the so-called evolutionary debugging problem (Cosmides & Tooby 1994). In contrast, any account of modularity in which modules come out as small and promiscuous is an account of modularity that no longer promises to solve the debugging problem. And (it may be alleged) any criterion of modularity that casts modules in such a diminutive role cannot be considered especially salient.

Now I am defending dissociability as a criterion of modularity. My position must therefore seem a little curious, for am I not by defending dissociability actually defending the wrong sorts of modules—given the sorts of modules that this criterion delivers if the redeployment hypothesis is correct? I can certainly see how an evolutionary psychologist would be puzzled by my position. But, as I shall explain later, I do not think the evolutionary psychologist’s reasoning here is persuasive—frankly, the sorts of modules she is after are very unlikely to be found anywhere beyond the most primitive domains, and the search for them at all reflects a misunderstanding of the brain and its evolution: the debugging problem is not a deep one. Like it or not, therefore, it looks as if we are going to have to rest content with a diminutive role for modules—which may not be such a bad thing anyway. For while dissociability may not ultimately meet the desiderata for a theory of evolutionary psychology, it ought to safeguard a respectable threshold for modularity nonetheless. It furnishes a kind of cognitive “movable type” for the mind, and mechanisms that can support robust laws, generalizations, and predictions (e.g., “forward” inferences from cognitive tasks to brain areas) (Burnston 2016). If—

for a given neural area A, there is some univocal description D, such that D explains the functional role of A’s activity whenever A functions

—it should be possible to formulate a theory tokening A providing “functional descriptions that apply over a range of instances of functioning,” and “functional explanations in particular contexts that are relevant to contexts not yet explored” (Burnston 2016, pp. 529, 531). This would be a “very powerful theor[y] in terms of generalizability and projectability” (Burnston 2016, p. 531).⁵

So what, then, of Jungé and Dennett’s suggestion? The problem, as I see it, is that it confuses modularity with faculty psychology more generally, and so reduces it to a platitude. Being neither controversial, falsifiable, nor particularly interesting, the system view fails to live up to the theory’s venerable reputation. On such an expansive definition, who would not emerge as a defender of modularity? Certainly few theorists in the cognitive neurosciences would deny the utility of functional decomposition as an effective research strategy (Prinz 2006; Piccinini & Craver 2011; Boone & Piccinini 2016).⁶ And of course a higher-level cognitive system composed of shared neural elements might well exhibit natural kind properties, such as a systematic set of procedures for dealing with typical inputs (Chomsky 1980a; 2006; Pinker 1997). But it is difficult to see how such a definition could have any substantively worthwhile theoretical upshots, certainly of a kind that could possibly justify the enormous effort spent in advancing modularity as some sort of solution to a deep and longstanding set of issues. On this weak view, what would the modularity of cognition explain about cognition beyond the simple fact that the mind, too, may be investigated using the techniques of natural science (i.e., “divide and conquer” works here, too)? If the answer is “not much,” this cannot be a good account of modularity—assuming that by “modularity” we mean a substantive doctrine. In the weak construal, modules turn out to be little more than fruitful perspectives on the mind, the mind considered from this or that particular point of view; say, the point of view of its linguistic capabilities, its pitch discrimination capabilities, its problem-solving capabilities, and so on (in principle, without limit). Such perspectives unquestionably give us useful entry points into what would otherwise be intractably complex, and allow us to figure out what it is that the mind actually does. But it is hardly surprising that a targeted coming-to-grips with a complex object should yield significant insights. The same strategy is familiar in one form or another in virtually all domains of rational inquiry, be they physical, chemical, biological, psychological, or otherwise. That “science works here, too,” I take not to be an interesting claim, if it comes to that, because it does not so much as provide a theory of cognition at all: if anything, it says more about science than it does about cognition. Furthermore, it is not entirely clear that the behaviorists would have spurned the sort of modules in view here. What they denied was the existence of sui generis principles, or the computational/architectural/modality independence of certain capacities (e.g., language), which I regard as evidence of true modules. They would not as a rule have denied that a partition of their subject matter could lead to interesting results. Recall the title of Skinner’s Verbal Behavior—I think it is fair to say that Skinner sought quite literally to explain the language faculty, albeit in terms of general associationist learning mechanisms, and is this not nearly comparable to the system sense of a module now under discussion? The behaviorists may have offered a shallow theory of human capacities, but even it did not appear to preclude modules in this sense (see, e.g., Chomsky 1979, pp. 49—51). Nor for that matter is there any logical reason why a connectionist or holist would have to rule them out, either. Contemporary Parallel Distributed Processing models of cognitive architecture in fact do have a sort of generic componential motivation behind them (O’Reilly 1998; Jilk et al. 2008).

It is worth being clear about exactly why system modularity fails the test of being “interesting.” It is easy to be misled here by the genuinely “interesting” results that have been achieved as a consequence of adopting the system view; i.e., by what has been learned about distinct domains of psychology as a result of iterative functional analysis (task analysis, decomposition, “boxology,” etc.). The system module’s notable successes, as well as its historical association with the computational theory of mind and the view of the mind as richly and intricately structured, are apt to lead to an exaggerated estimate of its true significance. Any theory of the mind pitched at the level of faculties (or analyzable parts, components, units, etc.)—as modularity most assuredly is—must tell us what it does about the mind through what it tells us about the faculties (or whatever the relevant units of analysis happen to be). If it does not speak “through the faculties,” as it were, it cannot so much as count as a faculty psychology, since the properties of the mind to which a faculty psychology brings our attention are, in the first instance, properties of the mind’s faculties. This point is at once obvious and yet so readily overlooked that it needs to be emphasized. That the mind is richly structured, that the mind is a computer, that the mind obeys laws exhibiting a clear mathematical structure, and so forth—these statements, if they are true, are true of the mind generally, meaning that it ought to be unsurprising if the divisions of the mind are correspondingly rich, intricate, computational, systematic, and so on. None of these properties attach to faculties per se. Moreover, learning that the language faculty has such-and-such features, or that vision operates in this or that fashion, need not tell us much about faculties qua faculties either, as against telling us about this or that particular faculty. Thus neither general claims about the mind associated with the Cognitive Revolution, nor specific claims about specific faculties, hard-won though these insights may have been, automatically get reckoned as among the distinctive insights marking out a truly general theory of faculties, which is after all what a faculty psychology aims to be. Contrast such claims with those of a well-developed faculty psychology (e.g., Fodor 1983). The roster of properties associated with Fodorian modularity (domain specificity; encapsulation; shallow, fast, and mandatory processing; hardwiredness; etc.) do not amount to a list of properties pertaining to the mind generally, nor to specific faculties considered independently, but to all faculties qua faculties. This is what made his theory interesting. So, as easy as it is to roll the system module in the glitter of the Cognitive Revolution, a frank assessment of this module demands that we isolate clearly what it is the theory that posits such modules says about the mind at the level of faculties—and when we do this, I maintain, we will be hard put to find anything that would not heartily be conceded by anyone who believes in the power of science (be they classical modularists, connectionists, holists, and, as I suggested, probably even behaviorists, mutatis mutandis).

Furthermore, lest it be thought that the very idea of functional decomposition can underwrite the system view—for one must admit that decomposition proceeds in a curious fashion where computers are concerned; namely, via the execution of subroutines by simpler algorithms or subagencies (“homunculi”), surely a nontrivial design feature of such systems—it need only be pointed out that homuncularity is not the same as modularity. Careful psychologists have always understood the difference, and that modularity is really a special type of homuncularity (Mahon & Cantlon 2011, pp. 149—151), just as homuncularity is a special type of decomposition (van Gelder 1995, p. 351). It is interesting to observe in this connection that David Marr, one of the chief architects of the computational theory of mind, did not see computationalism (and therefore, we may surmise, homuncular functionalism) as providing a free pass to his “principle of modular design.” Modularity seems for Marr to be an added feature that some computational systems, for largely heuristic reasons, might be thought to possess:

Any large computation should be split up and implemented as a collection of small sub-parts that are as nearly independent of one another as the overall task allows. If a process is not designed in this way, a small change in one place will have consequences in many other places. This means that the process as a whole becomes extremely difficult to debug or to improve, whether by a human designer or in the course of natural evolution, because a small change to improve one part has to be accompanied by many simultaneous compensating changes elsewhere. (Marr 1976, p. 485)

So, although homuncularity is not so generic as “mere decomposition,” it is nowhere near as important a principle as modularity, either. Accordingly (and for additional reasons I canvass later), we should withhold the more serious designation from generic system subcomponents and procedures that are non-dissociable.

So far, I have said nothing about two important features of classically modular systems: domain specificity and informational encapsulation. Can they get the system module over the line? Actually, the question itself is incoherent. Consider that once a module is allowed to consist of shared parts, it can no longer be domain-specific, except perhaps in an abstract sense (see § 7.2.1). This is because the “module” will be sensitive to potentially many domains, since its parts are presumably domain-general (see § 4.3). Put another way, domain specificity⁷ requires a functionally integrated unit that can respond to specified inputs. While the component modules of a composite consisting of shared parts would be functionally integrated, it is not obvious that the composite itself would be, although it might be said to have a sort of ad hoc integrity when in use. Notice also that a composite is unlikely to be informationally encapsulated “precisely because in sharing parts [it] will have access to the information stored and manipulated by [other high-level systems]” (Anderson 2010, p. 300). Anatomically distributed and overlapping brain networks simply must share information on some level (Pessoa 2016, p. 23). Lacking both of these properties, then, one or the other of which has been considered definitive (Coltheart 1999; Fodor 1983, p. 71), its postulation does not quite serve the purposes many would assume. One might have supposed that the system module could be more strongly motivated if at least it had the property of either domain specificity or encapsulation (in a concrete and unambiguous sense). And yet just because it is a composite, it can be neither truly domain-specific nor (in all likelihood) informationally encapsulated.

In something of a reductio, then, the system view of modularity leads only to the claim that the mind can do different things at different times. Certainly a more ambitious and theoretically interesting claim than this is available; namely, that the mind can do different things at the same time; but as far as we know this requires functional specialization; i.e., separate moving parts (real modules), since the prospect of neural time-sharing appears low (see § 7.5 on the time-sharing problem). The pervasiveness of cognitive interference effects and processing bottlenecks in stimulus-rich environments that impose overwhelming attentional demands is enough to make this clear (Anderson 2010, p. 250). In short, either these debates are trifling, or the claims at stake are more adventurous than the system view permits. Here I shall presume that the more adventurous reading is correct, and that, in any event, functional dissociability really ought to be considered the sine qua non of modularity.

Bear in mind also that, in the context of cognitive neuropsychology, modules have been defined largely by reference to what the dissociation evidence has revealed; i.e., “on the basis of the specific behavioral effects of brain lesions” (Bergeron 2007, p. 177). Bergeron calls this inferential strategy the “functional modularity inference.” Basically, “the presence of highly selective cognitive impairments (dissociations) like prosopagnosia and various linguistic processing deficits suggest [sic] the functional independence of at least some cognitive processes,” and this in turn licenses the postulation of functionally independent modules subserving those processes (Bergeron 2007, pp. 176, 177; Gazzaniga 1989, p. 147). The fact that brain lesions are often also relatively localized suggests that such modules reside in a “specific and relatively small portion of the brain” (the “anatomical modularity inference”) (Bergeron 2007, p. 176; Gazzaniga 1989, pp. 947, 950). Make no mistake, the legitimacy of these inferences is hotly contested, since noisy dissociations are compatible with a system’s being dissociable, and clean dissociations compatible with a system’s being substantially nondissociable. In the first instance, “there are a variety of reasons, well explored in the neuropsychology literature, for which lesions to brain systems can produce noisy rather than clean patterns of breakdown even when the systems required to complete the task are modular” (Barrett & Kurzban 2006, p. 642). A good example would be a focal lesion at the border of two adjacent modules—the breakdown would not be clean, yet the two systems would be modular. In the second instance, even perfect (or “double”) dissociations cannot conclusively establish that the affected systems are modular, for a lesion might only compromise a small isolable component of an otherwise highly interpenetrative circuit. Damage to this component might result in the system depending on that component being independently impaired, but it does not follow from this that the system would be functionally dedicated (although admittedly it would be dissociable at the margins). In light of this, it may come as something of a surprise to be told that these arguments

have failed to deter theorists from employing either of the inferential strategies. Indeed, the functional modularity inference continues to be one of the most common approaches among cognitive neuropsychologists for inquiring about the structure of cognition. Similarly, the recent cognitive neuroscience literature abounds more than ever with cases involving the use of the anatomical modularity inference. (Bergeron 2007, p. 177)

But what the persistence of these inferences bears witness to is the fundamental role that dissociation evidence plays in the search for modules, and that functional specificity itself continues to be the lodestar for deciding upon whether, and if so to what extent, the mind is modular within the major disciplinary fields concerned with this question. As we have seen, the same assumption underwrites evolutionary psychology and massive modularity, the central claim of which is that the mind is predominantly composed of parts selectively shaped by evolutionary pressures (Carruthers 2006). As two prominent evolutionary psychologists state their position (Barrett & Kurzban 2006, p. 630): “. . . we intend an explicitly evolutionary reading of the concepts of function and specialization: modules evolved through a process of descent with modification, due to the effects they had on an organism’s fitness.” This view predicates the existence of systems that, though perhaps spatially extended and neurally interspersed, are dissociable in principle:

Psychologists generally agree—as do we—that because cognitive architecture is instantiated in brain architecture, the two will be isomorphic at some level. . . . However, at a larger, macroscopic level, there is no reason to assume that there must be spatial units or chunks of brain tissue that neatly correspond to information-processing units. An analogy might be to the wiring in a stereo, a computer, or other electronic system: Individual wires have specific functions, but at the level of the entire machine, wires with different functions might cross and overlap. For this reason, removing, say, a three-inch square chunk from the machine would not necessarily remove just one of the machine’s functions and leave the rest intact. In brain terms, it could be, and probably is, that macroscopic regions of brain tissue include neurons from multiple information-processing systems with multiple functions. (Barrett & Kurzban 2006, p. 641)

And of course other contemporary models of cognitive architecture, such as the successful ACT-R model, also posit the existence of independently modifiable subsystems of the brain.

When it comes to clarifying just what makes modularity interesting, one final set of considerations may be suggestive. While terminological nuances can hardly be decisive in an area like this, I think it is no coincidence that massive modularity bottoms out in claims about the separate modifiability of functional components. This is because the very word “module” evokes images of movable parts that can be assembled and reassembled in a variety of distinct combinations and that may be affected independently of one another. If all modularity amounts to is the claim that the mind can do different things at different times (rather than the stronger claim that it can do different things at the same time), and this suffices to call it “modular,” it ought to be permissible to say that a knife that cuts both meat and bread is modular. And yet no one thinks of knives as modular (unless they are Swiss Army knives, which actually come with different blades). It is, I think, instructive that other anatomically nondissociable systems with shared parts, such as nervous systems, reproductive systems, endocrine systems, and the like—all of which may be singled out for their natural kind properties—are termed “systems.” One never hears of digestive modules or reproductive modules. The “modules” of developmental biology and neuroscience that do have shared and reusable elements are an anomaly of network science. Most biologists, including developmental biologists, continue to think of modules as “anatomically distinct parts that can evolve independently” (Wolpert 2011, p. 115). Limbs and vertebrae would be modular on this view (being organs), but not the larger anatomical systems they compose.

It is well worth stressing here that my argument should not be read as an instance of mere carping or terminological pedantry. There are certainly occasions when scruples over the use of words reveal carping tendencies, and nothing much beyond that, but this is not one of them. Philosophers and cognitive scientists who allege a “module” for this capacity and a “module” for that capacity must be taken to be saying something substantial; i.e., something more than merely the fact that we have the capacities in question. To dignify these capacities with the honorific title “module” is, I would suggest, an attempt to invest the capacities with special-purpose, special-structure status. If philosophers and cognitive scientists persist in referring to modules for X and Y in the face of contrary evidence (i.e., evidence suggesting that the X and Y “modules” are not special-purpose, special-structure devices), they betray a willingness to exploit the connotations of a powerful term for rhetorical purposes. For, if by alleging that there is a module for X or Y the speaker intends only to say that we can give systematic accounts of X and Y—where X and Y represent particular foci of the scientific gaze upon the mind—the speaker is only avowing a belief in the efficacy of the scientific method in the realm of cognition, which I take it no naturalist would deny. In such circumstances, it would be better to drop the term “module” altogether, and settle for a less loaded (and therefore more honest) term like “capacity,” “faculty,” or “system.”

So, while it is true that I am insisting on correct usage, this insistence is not without justification, and not without consequences should laxity prevail. In some ways the issues here are analogous to those that have arisen in the philosophy of biology over the proper use of the word “innate” (see Chapter 6). Neither those who urge elimination—because the word engenders confusion and fallacies of ambiguity amid a plethora of conflicting folk-biological intuitions—nor those arguing that a technical definition can be given should be seen as engaging in a merely feeble semantic dispute.