Does Modularity Survive the Evidence of Neural Reuse? - Modules Reconsidered: Whither Modularity?

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

Does Modularity Survive the Evidence of Neural Reuse?
Modules Reconsidered: Whither Modularity?

One of the primary contentions of this chapter is that the cortical column we have just examined is probably the only robust example of modularity that could survive evidence of reuse, and this just because reuse seems almost destined to predict something very much like it: small, stable, reusable nodes appearing within various distributed networks spanning various cognitive domains. The question before us now is whether despite appearances, neural redeployment really is compatible with the degree of functional specificity that modularity demands.

One thing appears reasonably certain. If the cortical column (or Andersonian “working”) were to survive reuse as the dedicated and functionally specific cognitive unit that it would need to be, not only would reuse then be compatible with the modularity of mind, it seems fair to say the Fodorian module itself would be likely to survive in some form—at least to the extent that cortical columns retain both stimulus specificity and informational autonomy, properties that they are likely to retain if brain regions are as task-selective and functionally constrained as the evidence in § 2.4.3 suggests they are. To be sure, the neo-Fodorian module would be a shadow of its former self, barely recognizable in size and certainly no longer suited to its original role as a marker of gross cognitive function.1 But the resulting picture of the mind would still be modular.

But does the cortical column—or its Andersonian concomitant—emerge unscathed in this way? To get a sharper sense of the options available to us on this question, I shall set the overall account of reuse in the context of Pascual-Leone and Hamilton’s (2001) original metamodal hypothesis, which is an important forerunner of contemporary theories of reuse, including Anderson’s. This account trades in brain modules, which it terms “operators,” and so allows me to convey very crisply the obvious sense in which modularity is compatible with reuse. I shall then walk through the principal objections to this view. Given what I take modules to be, my criterion of demarcation must be the degree to which dissociability no longer remains tenable even in principle. If functional specificity is no more than a will-o’-the-wisp, modularity itself can be little more than that. Therefore, I shall propose a simple device by which we can usefully conceptualize the problem facing the modularist. At its core, modularity turns on evidence of specialization. What we require, then, is a scale of specificity for brain regions that makes their indicia of specificity explicit. To the best of my knowledge, such indicia have not been presented in quite the same way before.2 I conclude this section with an assessment of the long-run prospects of modularity.

The metamodal hypothesis is intended to account for the observation that “our perceptual experience of the world is richly multimodal”—that “[w]e are able to extract information derived from one sensory modality and use it in another,” and “know a shape by touch and identify it correctly by sight” (Pascual-Leone & Hamilton 2001, p. 427). The hypothesis accommodates the possibility of crossmodal recruitment, and more specifically, the supramodal dynamics we encountered in § 2.4.3. Of relevance here is the fact that it is an adaptation of Robert Jacobs’s (1999) “mixtures of experts” (ME) model. The ME model builds on two important ideas. First is the idea of functional specificity and spatial localization (i.e., the anatomical modularity assumption). Different brain regions possess different structural properties, and these differences make for differences in functional capability to the extent that some regions will be better suited to performing particular functions than others. Second is the idea of competition between modules. Brain regions become specialized for processing particular inputs through open competition, but the competition is rigged, as it were, by the functional proficiencies that characterize the different regions: “each region tends to win the competition for those functions for which its structure makes it particularly well suited” (Jacobs 1999, p. 32). Two predictions follow. One is that the differences between neural regions appear quite early in development, and might even be innate (an issue to which I return in Chapter 6). The other is that “neural modules should enforce the outcome of a competition through a set of adaptable inhibitory interactions that allow modules to suppress the outputs of other modules” (Jacobs 1999, p. 34). Accordingly, Pascual-Leone and Hamilton propose that, instead of “unimodal sensory systems that are eventually integrated in multimodal association cortical regions,” the whole cortex

might actually represent a metamodal structure organized as operators that execute a given function or computation regardless of sensory input modality. Such operators might have a predilection for a given sensory input based on its relative suitability for the assigned computation. Such predilection might lead to operator-specific selective reinforcement of certain sensory inputs, eventually generating the impression of a brain structured in parallel segregated systems processing different sensory signals. In this view, the “visual cortex” is only “visual” because we have sight, and because the assigned computation of the striate cortex is best accomplished using retinal, visual information. Similarly, the “auditory cortex” is only auditory in hearing individuals and only because the computation performed by the temporal, perisylvian cortex is best implemented on cochlear, auditory signals. However, in the face of visual deprivation, the “striate cortex operator” will unmask its tactile and auditory inputs to implement its assigned computation using the available sensory information. (2001, pp. 427—428)

The crucial message for us here is that in this picture, despite neural operators being functionally and computationally constrained, their range of application is not. Neural operators are intrinsically versatile from the point of view of which inputs they can process, limited only by the amenability of the inputs to undergo a definite sort of manipulation. Barrett and Kurzban (2006, pp. 634—635) call something like this formal domain specificity—a construal of domain specificity wherein “domain” is not understood semantically, as referring to the set of objects or stimuli comprising a traditional task category, but rather syntactically, as the formal processing competence of a system. Formal domain specificity (essentially domain generality) comports with a view of the brain in which its several regions have manifold latent afferent input channels—preexisting connections supplying the critical cortical infrastructure that makes reuse possible. And while the picture presented in the preceding quotations would suggest a certain stability or equilibrium is achieved after a suppression mechanism ensures that the best module wins (so that individual modules get tuned to particular inputs and not others), as the examples presented in Chapter 2 dramatically attest, we do not have to wait for “visual deprivation” for this hidden complexity to be “unmasked,” since it is a normal feature of healthy adult brains to exploit these channels all the time (e.g., when “seeing” the face of a loved one at the sound of their voice, or tools at the sound of a hammer, etc.). Hence, supramodal organization simply entails neural reuse. Moreover, the model demonstrates how readily modularity can be combined with redeployment, inasmuch as the latter naturally presupposes the former.

But now we must finally confront the objections to this “minimodule” view that reuse seems to entail. We can distinguish two broad lines of attack: one weak, the other far more serious and potentially fatal. It is well to address the weaker one first. Here the charge is that minimodules are “compatible with an anemic version of localization that claims simply that individual brain areas do something, and the same thing, however low-level, simple, or cognitively uninteresting, whenever they are activated” (Anderson 2007c, p. 164). Such entities can hardly be controversial, since very few people nowadays regard the brain as a “disorganized mash” (Prinz 2006). Notice that this is the same objection I raised earlier against the system view of modularity: “when the notion of modularity is denatured, it turns into a platitude” (Prinz 2006). And for all their differences, neither systems nor minimodules pretend to solve the evolutionary debugging problem. So we seem to have another case of truism dressed up as theory.

In the case of minimodules, however, I think the objection can be overplayed. It is true that minimodules are incapable of offering a simple way through the debugging problem, and this might be thought to commend the sort of modules defended by evolutionary psychologists instead. But what would be the use of a theory that resolved puzzles by ignoring reality? A theory must aim to be both tractable and realistic (Coase 1937). As convenient as it would be for us to suppose that modules are independently modifiable gross cognitive components, the evidence of neural reuse suggests that this is not how the brain is organized. So either the problem itself is real and the evidence of reuse must be explained away, or the assumption that no non-(massively)-modular brain could possibly evolve must be set aside. Surely the latter approach would be the more sensible. The debugging issue itself is to a large extent a symptom of looking at things a certain way. If we accept that, on some level, evolution has to involve the emergence of functionally exiguous neural parts, and view the engineering problem as being how preexisting parts might be combined in novel ways, concerns over debugging become far less pressing. Minimodules, in any case, are not trivial. The mind could have been (and indeed has been) modeled in very different ways (think connectionism/Parallel Distributed Processing, holism, etc.), and a minimodule hypothesis is quite demonstrably falsifiable (unlike, say, the system view). Minimodules also support robust predictions (like forward inference) and theory-building. The truth is that minimodules are as modular as they need to be—modular enough to solve the very real wiring problems posed by scaling circuits, and modular enough to rule competing accounts like functional holism and strict localization out of the question. The trivialization charge is a nonstarter.

What, then, of the more serious line of attack? Although there are developments of the argument in several directions, its general thrust is to make a lot of the fact that the brain implements a network. Anderson (2010, p. 249) frames the issue in these terms: “Instead of the decompose-and-localize approach to cognitive science that is advocated and exemplified by most modular accounts of the brain, neural reuse encourages ’network thinking.’ ” To recapitulate briefly, all networks share a number of important properties, properties that make the study of any structure that exhibits a network design far more tractable than it might otherwise be. Preeminent among these of course are nodes and edges, but a defining mark of the network approach is its focus on the global structure of interactions between nodes, rather than the individual nodes themselves. Thus, if the brain is a network, modularity goes awry, for even if we were to concentrate all our energies upon modules qua nodes (i.e., minimodules), still we would be missing the point—the key to networks lies not in their nodes, but in the structure of their interactions.

Such is the clear-cut statement of the challenge. Put in this form, however, it just overstates the case. First, the fact that a team of soccer players exhibits higher-order dynamics in no way obviates the importance of individual players to the game; indeed, their unique talents and skills are what drive the interactions that feature at the level of abstract topology. Second, and this is well worth remembering through all the hype, one should be no less judicious in one’s use of a network analogy than in one’s use of any other:

Although the terms “network” and “connectivity” are widely used when talking about regional covariation in the human brain, it is important to keep in mind that no human data at present allow us to make inferences about brain regions forming networks in the true sense of the word. In particular, under a technical definition, two brain regions form a network if they are anatomically connected, typically via monosynaptic projections. In living humans, we rarely, if ever, can say anything conclusive about anatomical connections among brain regions. . . . [C]ollections of regions are more appropriately characterized as functional systems. (Fedorenko & Thompson-Schill 2014, p. 121)

Still, even if we were to moderate the argument in allowing for such complications, the network challenge would remain. We may for convenience describe three distinct iterations of the challenge, each more persuasive than the last, which in one way or another play upon the importance of the network context for understanding local function (inasmuch as context determines meaning). The thought here is that because minimodules appear across multiple and functionally diverse neural communities, they lack the precise degree of specialization required of modules—in view of just how tiny minimodules are, the more partnerships a given minimodule enters into, the more abstract its contribution becomes, and the dumber, simpler, and more generic it will ultimately be (Klein 2012). Price and Friston (2005, p. 268) use the example of a forefinger. Its many roles could include piano playing, typing, scratching, pinching, and feeding; yet if we had to designate its overall role, we would have to settle on something explanatorily inert: “the forefinger can only do one thing—’bend’ and ’straighten.’ Its role in other tasks is . . . entirely dependent on what the other fingers and thumbs are doing and what environmental context they are in.” In short, “naming the specific function that a region performs (or even supposing it has some single specific function) involves a kind of abduction that is inherently underconstrained and uncertain” (Anderson 2014, p. 53).

Another way of framing the issue is in terms of plasticity. The more functionally versatile and unstable a brain region, the more plastic it will be (other things being equal). At the limit, ontogenetic plasticity might be so great that even sudden, swift connection changes to the neural configuration of a given brain region between alternating task demands would be possible, and functional stasis merely illusory. Up until now, neuroscientists have simply presumed that a network approach can naturally complement a modular approach—naturally, because from the minimodule perspective, modules are akin to the nodes of a coactivation graph; but the plasticity of neural regions might so undermine their functional specificity that even neuroscientists will have to give up the pretense that such node-like entities can be modules in the full-blown sense they almost always take for granted, as when they describe these regions as “functionally specific brain regions” or “regions that are selectively engaged by a particular mental process” (Fedorenko & Thompson-Schill 2014, p. 121). In the event that neuroscientists might still like to refer to these entities as modules—much in the way they conventionally use the term to describe the communities of nodes in graphs—it would be a case of terminological convenience trumping theoretical rectitude.

The weakest iteration of the challenge adds little to what has already been said, but it might note how the preponderance of afferent input pathways sustaining the brain’s supramodal organization must ever so slightly color an individual module’s operations as to rob it of a deep and lasting functional essence. The more functionally promiscuous a region—joining now with the visual system, now with the language system (say)—the more we can expect the neural context to impinge on the region’s functional capabilities. Brain regions are by and large fairly homogeneous anyway (Buxhoeveden & Casanova 2002, p. 941). Standard histological preparations and cytoarchitectonic methods often fail to reveal anatomical differences between neighboring yet functionally distinct cortical columns. Thus an important strategy by which the brain generates difference from sameness is through the formation of different interconnection patterns among neurons and regions; indeed, often among the very same neurons and regions. Input channels therefore cannot be conceived as merely useful appendages to the lines of script run by distinct neural operators, as they are themselves partly constitutive of the operations performed by them. Functional promiscuity means we cannot retain a prespecified notion of function for brain regions considered in isolation of the neural contexts in which they appear.

Now it must be said that, when put like this, the argument again runs the danger of just overstating its case. For what it seems to lead to is a variety of holism. Insofar as that is where this line of thinking is taking us, it should be resisted, for the weight of evidence does not support holism, classical connectionism/PDP, or anything like it, really. With that proviso in place, the argument is a good one—functionally distinct operators with functionally distinct input criteria can be observed in the brain, but a moderate pitch to incorporate the effects of context would not go astray. Let us call these “weak context effects.” Weak context effects are those that do not compromise a brain region’s ability to perform a well-defined, functionally specific (albeit domain-general) operation. This is consistent with how Anderson (2010, p. 295) originally defined a working: “Abstractly, it is whatever single, relatively simple thing a local neural circuit does for or offers to all of the functional complexes of which the circuit is a part.”

Evidence for stronger context effects is not hard to find. Let us call them “strong context effects.” These will constitute the basis for the second and third iteration of the network challenge; but before advancing any further on this front, I should make one point clear at the outset: there is something about strong context effects—implying as they do a much higher degree of plasticity for local circuits than we have encountered so far (details to follow)—which sits uneasily with aspects of the evidence of massive redeployment presented in Chapter 3. The problem is that strong context effects are incompatible with evidence suggesting that the units of redeployment are themselves relatively fixed in nature (even after allowing for synaptogenesis, etc.). To the extent that strong context effects obtain, then, the theory of reuse requires amendment. Anderson’s massive redeployment hypothesis, it will be remembered, predicts that recently evolved functions should be supported by more widely scattered regions of the brain than older ones, since it should on the whole prove easier to utilize existing circuits than to devise special-purpose circuitry afresh, “and there is little reason to suppose that the useful elements will happen to reside in neighbouring brain regions” (Anderson 2010, p. 246). Not only is the evidence that Anderson cites consistent with this prediction, it could also be taken to imply something more specific about the nature of local circuits, an implication that Anderson originally had no hesitation in drawing:

If neural circuits could be easily put to almost any use (that is, if small neural regions were locally poly-functional, as some advocates of connectionist models suggest), then given the increased metabolic costs of maintaining long-distance connections, we would expect the circuits implementing functions to remain relatively localized. That this is not the observed pattern suggests that some functionally relevant aspect of local circuits is relatively fixed. (Anderson 2010, p. 247)

But while this is one way of interpreting the evidence, a distributed network organization might be favored by evolution for rather different reasons. As Bullmore and Sporns (2012, pp. 336—337) point out, one reason why a general principle of parsimonious cost control might be compromised in favor of far-flung neural circuits has to do with the resilience that such organization may be presumed to confer. Robustness to adverse perturbations—“[t]he degree to which the topological properties of a network are resilient to ’lesions’ such as the removal of nodes or edges”—could well have more to do with the distributed structure of recently acquired capacities than with the functional fixity of local circuits. At the very least, the inference that local circuits are not especially plastic again “involves a kind of abduction that is inherently underconstrained and uncertain.” Anderson himself appears to have moved on from his earlier commitment to fixed local workings, but not on account of resilience per se. He has lately been convinced by the evidence of strong context effects in their own right, and as a result no longer speaks of fixed local “workings,” preferring instead the less rigid connotations of the term “bias” in describing the functional proclivities of local brain regions. For Anderson a cortical bias is “a set of dispositional tendencies that capture the set of inputs to which the circuit will respond and govern the form of the resulting output” (2014, p. 15)—an idea that reconciles a brain region’s versatility and its overall functional durability without at the same time insinuating “that each circuit does exactly one specific thing” (2014, p. 16).

So what exactly, then, are strong context effects? I think we may usefully divide them into two broad categories. The first category—motivating the second iteration of the network challenge—would appear to suggest that small brain regions can assume radically different network states, and thereby alter their basic electrophysiological configurations, depending on the requirements of the cognitive system being used. This sort of operational, on-the-fly ontogenetic plasticity of neurocognitive resources undermines the purported functional fixity of brain regions, and hence the claim that brain regions can be modular (in the true sense of being functionally specialized). The second category—motivating the third and final iteration of the network challenge—goes even further than this by throwing into question the very legitimacy of functional decomposition as a basic strategy within the cognitive sciences. Here the thought is that “it is not as if we can identify the one fixed function of an element of the system and then worry about the effect of interactions later. Rather, the interactions are often precisely what fix local function” (Anderson 2014, p. 208). This may not at first appear to be saying much more than what was said in the first instance. In fact, its ramifications are deeply unsettling for the “decompose-and-localize” approach to cognitive science, as I shall explain more fully in a moment. Let us take these two putative categories of context effects in turn.3

Evidence of swift, sudden connection changes in networks begins at the single neuron level. C. elegans has acquired fame as the nematode for which the first neural network wiring diagram was published. It contains about 300 neurons and up to 7000 synaptic connections, simple yet complex enough to serve as a model of function-structure dynamics within higher organisms. C. elegans neurons perform “more than one type of circuit function, including both motor and sensory functions,” and sometimes perform multiple functions within the same modality (Altun & Hall 2011). Beyond the straightforward implication here that neural reuse may be evolutionarily conserved, there are intimations of still more intriguing possibilities. “Neuromodulation” refers to a family of context effects in which it is possible for the same neuron to radically change function—and perform in just the opposite role—in response to changes in the electrophysiological, chemical, and genetic environment. One example is C. elegans’s olfactory neuron, AWCON, which can apparently signal both attraction and repulsion to the very same odor, depending on its neuromodulatory configuration. Another is the nocioreceptive ASH neuron, which can direct both sociality and avoidance. But neuromodulation is not restricted to C. elegans. Similar effects have been documented in both the pond snail and the honeybee, and there are enough instances within vertebrates to suggest that neuromodulation might be a basic evolutionary strategy for coping with scarce neural resources (Anderson 2014, p. 34). Of course, before such results could support more ambitious inferences regarding human cognition, we would need evidence of large-scale modulation in more complex organisms. In the simplest organisms, small modifications of even single synapses can have significant behavioral ramifications. In larger and more complex organisms, this is unlikely to be the case. In fact, evidence of such large-scale effects does exist, even within the human literature. Most suggestive of all is the evidence Cole et al. (2013) report for “flexible hubs” in the brain that “flexibly and rapidly shift their brain-wide functional connectivity patterns” in response to changing task demands. If brain regions really do move into different functional configurations, as distinct from being redeployed in the same state for different purposes, it would imply that brain regions can be neither functionally specialized (in the sense of contributing a stable and predictable operation across their various higher-order applications) nor dissociable, both because their disruption would directly impede the operation of an equivalent system—the selfsame region considered from the standpoint of any of its alternative network states—and because it could well prove impossible to identify a segregatable unit of neural tissue that retained a constant form from state to state.

One upshot of this concerns theory-building. Any theory we construct that tokens a brain region subject to strong context effects will not be able to offer a fully general explanation of what that region offers to all of its networks. Even those who do not think contextualism would undermine our ability to construct powerful theories supporting strong predictions concede that we would nonetheless be in the realm of partial generalizations (Burnston 2016). Furthermore, part of the appeal of a theory that posits functionally specific brain regions is that it supports robust inferences: one should in principle be able to infer which brain region has been engaged simply from knowing what function is being performed. Strong context effects undermine the robustness of such inferences.

None of this entails that the brain is equipotential or has an inherently open texture (as we shall see in Chapter 6). On the contrary, overlaps between the neural implementations of cognitive tasks are frequently found to involve functional and semantic inheritance (Mather et al. 2013, pp. 109—110; see also my § 3.3), implying that brain regions have a stable set of causal features that regulates their participation in various networks. This is consistent with the finding that recently acquired skills in the human lineage, such as reading and writing, have highly uniform neural substrates across both individuals and cultures (Dehaene 2005). But when the specific point in dispute is whether the mind/brain has a modular architecture, such facts alone cannot be decisive, for then the issue is not whether brain regions have specific developmental biases, input preferences, or an underlying structural and functional integrity, but precisely the degree to which brain regions are specialized. A bias is not a specialization.

To recapitulate, so far we have considered how natural the alliance between modularity and reuse can be, and proceeded to examine various objections to this view. The objections come in two forms: weak and strong. The weak objection alleges that minimodules are trivial entities, but we saw how this claim is in fact unwarranted. The stronger objection plays on the network structure of brain organization to reveal the illusoriness of functional specialization for individual brain regions, being comparable to network nodes whose functional importance is subordinate to internodal network interactions. This stronger network challenge in turn assumes three distinct forms, one emphasizing weak context effects (which we dismissed as instructive but not fatal to modularity), and two emphasizing strong context effects. The first category of strong context effects correlates with increasing ontogenetic plasticity. The objection from these context effects really does have bite, and probably compromises the modularity of any brain region that is vulnerable to their impact. I turn now to the second category of strong context effects.

The second category raises very serious doubts over the legitimacy of componential analysis, and so by implication practically all mainstream work in the cognitive sciences. The decompose-and-localize approach to cognition assumes that the mind can be understood by analogy to a machine with working parts. Central to this approach is the belief that function can be explained in terms of “bottom-up additive contributions” rather than “top-down constraints that limit, change, or determine functional properties” (Anderson 2014, p. 308). Recent discoveries suggest this confidence may be misplaced, although quite to what extent remains unclear at this stage. The starkest illustration of these effects is offered by starburst amacrine cells (SAC) in the mammalian starburst amacrine cells (SAC) retina. These are axonless neurons with dendrites arranged radially around their cell body. “What is especially interesting about these cells is that the dendrites are individually differentially sensitive to motion; they release neurotransmitter only in response to motion that is congruent with the direction the dendrite points away from the cell body” (Anderson 2014, pp. 92—93). It is tempting to think of each dendrite as a component because each appears to contribute uniquely and dissociably to effects at a higher network level.

In fact, the directional selectivity of each dendrite is due in large part to the particular blend of connections these cells have to bipolar cells and other SACs such that responses in the congruent dendrites are reinforced while responses in noncongruent dendrites are inhibited. In other words the directional selectivity of the dendrite in a given situation is due not so much to intrinsic properties of that dendrite but to global properties of the network. Global function is not built from componential local function, but rather the reverse! (Anderson 2014, p. 93)

While one could well think that the entire local network is itself a component, it should not come as a surprise if the very same dynamics “reproduce themselves at the higher level,” with the functional selectivity of whatever putative higher-level component being determined again by global network properties rather than intrinsic local features. If these dynamics apply more generally to neural networks, the assumption behind componential analysis would be substantially undermined, for then no longer would components be “temporally and functionally stable subassemblies sitting on the tabletop waiting for final construction” (Anderson 2014, pp. 93—94). Instead, the “functional organization of the whole” would be logically prior to the functionally parasitic part. Put another way, interactions between parts would be more important than the activity of parts (Anderson 2014, p. 40).

Olaf Sporns has mooted similar ideas. The traditional way of thinking about circuits is in terms of “highly specific point-to-point interaction[s] among circuit elements with each link transmitting very specific information, much like an electronic or logic circuit in a computer” (Sporns 2015, p. 92). In this account, the activity of the whole circuit is “fully determined by the sum total of these specific interactions,” with the corollary that “circuit function is fully decomposable . . . into neat sequences of causes and effects.” This is a Laplacian model of classical mechanics, “with circuit elements exerting purely local effects.” The modern approach from complexity theory and network science, however, emphasizes “that global outcomes are irreducible to simple localized causes, and that the functioning of the network as a whole transcends the functioning of each of its individual elements.” As an example of an emergent network phenomenon, Sporns takes neural synchronization, “the coordinated firing ’in sync’ of large numbers of nerve cells” (Sporns 2015, p. 93). While this phenomenon clearly depends on elemental interactions and synaptic connections, “it is not attributable to any specific causal chain of interactions in a circuit model.” Rather, it is “the global outcome of many local events orchestrated by the network as a whole.”

We can represent these varying degrees of modular specialization along a continuum running from A to E, each with the indicia represented in Table 5.1. Brain regions at or to the left of C, which marks the onset of weak context effects, will be sufficiently specialized to count as modular. Brain regions to the right of C, characterized by strong context effects, will not. Plasticity increases as one moves from A through D.

So will modularity survive evidence of neural reuse, neuromodulation, and the very strongest effects of network context? It is worth remembering that at this stage the case for the very strongest of context effects is still speculative. Russell Poldrack, for his part—whose laboratory work in this space has been pioneering (e.g., Poldrack et al. 2009)—is convinced that cognitive systems will bottom out in low-level, domain-general and functionally specific computational operations bearing a one-to-one relationship to specific cortical sites. On the other hand, if neuromodulatory and context effects are indeed as pervasive and game-changing as some people seem to think (e.g., Bach-y-Rita 2004), perhaps only a few scattered islands of modularity are all we can reasonably hope for (Prinz 2006). It is true that employing current techniques it is not actually possible to assign brain regions to definite locations on a continuum, so we cannot know for sure that only a few brain regions will cluster towards the specialist end (e.g., A through C, as mentioned). But evidence for the existence, power, and ubiquity of context effects can only proliferate at this point, one would think. (Incidentally—taking up a point I raised in § 4.3—if the “modules” reaching into central cognition turn out to have type D characteristics, central cognition will be pro tanto nonmodular after all. See § 7.2.2 for further discussion.)

Table 5.1 A Scale of Specificity Along with Indicia of Specificity for Brain Regions







Theoretical domain specificity

Strict domain specificity

Formal domain specificity




Minimal afferents

Few afferents

Many afferents

Many afferents


Participation in a single task & composite within a single task category

Participation in various tasks & composites within a single task category

Participation in various tasks & composites within various task categories

Participation in various tasks & composites within various task categories







No context effects

Negligible context effects

Weak context effects

Dynamic local network states

Local function fixed by global properties

Functional specialization

Functional specialization

Functional specialization

Functional differentiation

Functional differentiation

Possible examples

Probably none—a theoretical postulate only

Neural element common & exclusive to reading, writing & speaking; e.g., the neural basis of subjacency/wh movement

Extrastriate body area; Broca’s area

Flexible “hubs” in the brain reported by Cole et al. (2013)

Starburst amacrine cells; synchronization

Reprinted by permission from Springer Nature, Biological Theory, “Neural reuse and the modularity of mind: Where to next for modularity?” by J. Zerilli. Copyright 2019.

It is worth mentioning that the picture here is consistent with the emerging consensus around the neocortical column we met in § 4.3 (Rockland 2010; see da Costa & Martin 2010 for a historical review). Rockland takes five defining features of the column and argues that these are too rigid to do justice to the complexity of cortical organization. For instance, it is supposed that columns are solid structures, but this is not quite true, since they have a heterogeneous substructure that “correlates with reports of locally heterogeneous response properties,” very much as reuse predicts (Rockland 2010, p. 3). Their anatomy is therefore messy rather than solid. Columns also form part of widely distributed networks at several levels (again, much as reuse predicts), and for that matter are not even obligatory to cortex: for instance “comparative anatomy provides many examples of cortex apparently without anatomical columns or dramatically modified columns,” such as in whales and dolphins, whose insular cortices have cellular modules concentrated in layer II, and the giraffe, whose occipital cortex has modules concentrated in the same layer (Rockland 2010, p. 7).

I think it is fair to say, then, that while most of the cortex undoubtedly consists of module-like elements, probably only a few of these will in the end prove to be modular in the robust sense we require. The full implications of network thinking for componential analysis in particular have not sunk in, even though they promise to overturn our conception of local function almost completely. It goes without saying, of course, that to the extent that modules do exist, it will be as functionally exiguous and promiscuous brain regions. The days of classical modularity are well and truly over.