R2.1. The importance of neural context

For many, these observations will simply add fuel to the skeptical fire being built under the use of neuroimaging in cognitive science (see Coltheart Reference Coltheart2006; Klein Reference Klein2010; Roskies Reference Roskies2007 for discussions of the general issue). Certainly there is reason to be cautious, but the potential value of neuroimaging is so vast that it would be foolish to forego its use. So how should we address this issue? The target article emphasizes that although neural regions are not use-selective, they may be “working selective,” and so clearly one promising empirical project is to begin to define local workings with various sorts of cross-domain modeling. What was less clear in the target article, but helpfully raised by Gomila & Calvo and Reimers, is that there is another, complementary empirical project: Although local regions are apparently not use-selective, networks of regions may well be use-selective. That is, it might be possible to recover selectivity by attending to larger scale patterns of regional co-activation and coherence. Cognitive tasks and task categories may turn out to have characteristic signatures at the network level (for discussion, see Anderson et al. Reference Anderson, Brumbaugh, Şuben, Chaovalitwongse, Pardalos and Xanthopoulos2010; Poldrack Reference Poldrack2006).

Whether and to what degree specific identifiable networks of interacting regions turn out to be use-selective is an empirical question, one that is only now starting to be investigated. But it seems clear that this is currently the most promising, and perhaps the only viable way to uncover use selectivity in the brain. Note the implication that knowing what activity in a given region means for the tasks engaging the brain will require careful attention to the neural context of that activation – to what else the brain is doing when it exhibits the activation of interest. Seeing activation in Broca's area may give one confidence that it is doing something of importance (although see Klein's commentary and Klein [2010] on the uncertainty of this inference), but knowing what else is active along with Broca's may tell us what that something is, the use to which Broca's is being put.

R2.2. Bodily, environmental, social, and cultural context

This point about the value of attending to neural context was somewhat eclipsed by my attention to the importance of local neural workings. Moreover, my near exclusive attention to neural facts appeared to Toskos Dils & Flusberg; Gomila & Calvo; Immordino-Yang, Chiao, & Fiske [Immordino-Yang et al.]; and Rozin to eclipse the value of attending to even broader contexts. I certainly endorse the general point that broader contexts – bodily, environmental, social, cultural – matter to our ascription of function. If we don't understand what an organism is doing, we can hardly hope to understand what its brain is up to; and figuring out the best way to describe or characterize a behavior certainly involves detailed attention to context. On the other hand, this is where semantics – and, specifically, the imprecision of the term function – can sometimes get in the way of science. It is an obvious point, and barely worthy of scientific attention, that a single mechanism can be construed as serving multiple functions; the alarm system detects motion, and protects the house from intrusion. This is less an observation about the fundamental nature of alarm systems than about the potential variety of our epistemic interests when we ask “What is it doing?” The cases of scientific interest are those where a single mechanism is put to genuinely different uses, the way many people (myself included) use their e-mail inbox also as a “to-do” list. Note the implied converse, that I thereby put different mechanisms – my inbox, my task list—to the same use.

So, is it the case, as Toskos Dils & Flusberg hypothesize, that the same cognitive behaviors can emerge from different, contextually variable sets of neural circuits? It is an interesting question and worth pursuing. But here is where attention to context, and its role in defining the conditions under which we will call one behavior the “same” as another, becomes crucial to the scientific enterprise. There can be no doubt that the same behaviors mechanically defined (writing on a piece of paper, say, or calculating exchange rates) can involve different neural circuits. But of course writing out a confession, or a love letter, or an essay are vastly different enterprises, and we should expect correspondingly different neural involvement. These would be cases where the neural context tracks the task context in ways that are interesting to discover, but also unsurprising.

What would be somewhat surprising is the discovery of different unique sets of circuits for the very same function, where there is no discoverable contextual or other difference to explain the apparent redundancy. Here would be a failure of the neural context to track task context because of the surfeit of neural mechanisms for the task in question. The discovery of such an example would be very interesting and illuminating, although it would not be a specific counterexample to neural reuse. Nothing about the theory suggests that there is only a single neural token instantiating any given type of neural working, much less a single, unique high-level neural network for every identified use. Some redundancy, especially in workings, is to be expected; it will sometimes be the case, because energetic constraints favored the outcome, or perhaps just as the result of chance, that different neural circuits developed the same working or came to subserve a similar use. And the discovery that such redundancy was extensive at the use/network level, and, more importantly, that differences in which networks subserved specified uses did not track context in some identifiable way, would be very puzzling and would affect far more than the theory of neural reuse, including the dynamic models that Toskos Dils & Flusberg favor.

Immordino-Yang et al. ask a similarly interesting and challenging series of questions. Can the same network of neural circuits in fact serve quite different uses, a difference that would only become apparent once cultural context was considered? Here again, it is not the least surprising (although not for that reason uninteresting) that cultural context affects which neural resources are brought to bear on cognitive tasks; for, after all, the context may well change (if only in subtle ways) the nature of the task. One would expect the neural context to track the environmental/cultural context in this way. What would be harder to assimilate is if it were often the case that the same network subserved different uses at the cultural level – genuinely different uses not arising from shift in epistemic perspective – without there being a detectable neural difference. It would be a bigger challenge because this would imply the existence of many cases where neural context does not track environmental context, and this would leave a large explanatory gap that behavioral science has not yet discovered a way to fill. Here again, this would not be a challenge specific to neural reuse; the discovery of radical context dependence in behavior would not undermine the discovery that neural resources are deployed in support of multiple uses across task categories and that differences in uses are better explained by patterns of inter-neural cooperation than by differences of activity in individual brain regions. But it certainly would suggest that this was only a part – perhaps a very small part – of the explanation of behavior. There are perhaps some hard-core neuroscientists who think that neural facts are the only relevant inputs to behavioral science, but I am not a member of that tribe, and the implications of neural reuse seem largely orthogonal to that debate.

R2.3. Context and complexity

Still, there is an interesting quasi-dilemma that is illuminated by the advocates of context. Insofar as neural context tracks broader context, then although initial attention to context would be necessary to identify the nature of the cognitive task or behavior in question, the lab-minded neuroscientist could then (safely?) focus on the brain, for the broader context would be reflected therein. This somewhat blunts the force of the argument for the necessity of contextualization. On the other hand, although the discovery of cases where neural context did not track broader context would appear to strengthen the case for the necessity of contextualization in science, the attendant increase in the complexity of the problem space could encourage the community to simply ignore such cases as beyond the current reach of scientific method. If there is no neural difference despite relevant differences in context and behavior, to what degree are subjects aware of the difference, and controlling their behavior with respect to context? How do the non-neural aspects of intention and control manifest themselves? It is incumbent on advocates of context to go beyond gestures to dynamic modeling or genetic mechanisms; they must both identify examples of this sort and describe an approach to understanding them that promises more illumination than mystification (Chemero Reference Chemero2009). I should be clear that I am not faulting the commentators for not doing so here; this is an ongoing challenge for all of us in the neuro-genetic-social-behavioral sciences.

R2.4. Niche construction

For a final word on the topic of context, it is important to keep in mind the facts that context is itself malleable, and that we are among the most important agents of that change. Iriki makes the case not just for the importance of neural niche construction, but also for the possibility that the changing neural context influences our evolutionary pathway, by inducing stable cultural changes and thereby changing the environment within which selection operates. Both Rozin and Lindblom make similar points. Culture (and especially language) can speed up preadaptation, both by increasing the degree and frequency of innovation and by buffering group members against selection pressures that might otherwise tend to weed out initially maladaptive exaptations. There is an extremely interesting literature emerging at the intersection of brains, genes, and culture (Boyd & Richerson Reference Boyd and Richerson2009; Hawks et al. Reference Hawks, Wang, Cochran, Harpending and Moyzis2009; Kitayama & Cohen Reference Kitayama and Cohen2007; Richerson et al. Reference Richerson, Boyd and Henrich2010; Suomi Reference Suomi2004), and I would be pleased if neural reuse turned out to be a more amenable perspective for coming to grips with these interdependencies than competing proposals on brain organization, such as modularity and localization (something Reimers suggests in noting the many parallels between neural reuse and molecular and epigenetic reuse).

R3.1. Workings versus plasticity

Still, neural reuse does suggest that these two processes will be mutually constraining, not to say antagonistic, and that opens some very interesting avenues for further exploration. I think the matter should be framed in the following way. The regions of the developing brain are likely to (and the massive redeployment hypothesis positively predicts that they will) have some innate functional biases, the strength and specificity of which undoubtedly vary from region to region. Where the nature of experiential input and the characteristics of the task demands being placed on the organism are consistent with these neural biases, then plasticity and reuse can act happily in concert. Neural plasticity generates local workings that reuse can arrange into various circuits that subserve the uses required for the organism's flourishing. (Apropos of which, it should be noted contra Michaux, Pesenti, Badets, Di Luca, & Andres [Michaux et al.] that nothing in the theory of neural reuse requires denying the necessity of experience to shaping local circuitry; more on this issue in section R4, para. 3.) However, where the nature of the input or the characteristics of the task are inconsistent with existing cortical biases or established workings, then these processes can easily come into conflict.

The experiments reported by Sur et al. (Reference Sur, Garraghty and Roe1988) and cited by many commentators here offer an excellent paradigm to further explore these sorts of conflicts. As is well known, Sur et al. (Reference Sur, Garraghty and Roe1988) rewired the ferret cortex by redirecting right visual field outputs to auditory rather than visual cortex. The result of this manipulation was the induction of neural circuitry in auditory cortex resembling that typically found in normally developing visual cortex – the classic “pinwheel” organization, for instance. The rewired cortex apparently instantiated workings typically associated with visual processing, such as orientation and direction selectivity, and subserved typical visual uses, such as orienting toward a visual stimulus (von Melchner et al. Reference von Melchner, Pallas and Sur2000). Plasticity is clearly a powerful force in the development of the brain. It is not, however, omnipotent; visual acuity in the right visual field was significantly lower than in the left visual field. This finding is consistent with the existence of congenital cortical biases potentially in conflict with the nature of sensory inputs, which had to be overcome to accommodate visual stimuli.

From the perspective of neural reuse, it would be interesting to have a more complete behavioral inventory of these animals. Although in this particular case behavioral evidence would have to be treated with caution, given the multiple neural ablations these experiments involved, such information could nevertheless offer some clues as to what other uses the now missing auditory workings might have served. What performance impact did the induction of “visual” circuitry into auditory areas have on other functions relying on the same neural region? Were the neural complexes underlying these other behaviors systematically altered by the rewiring? If so, how? Certainly, this is a paradigm that could be used to systematically investigate such questions for various regions of the brain.

Other opportunities for investigating the potential conflicts between plasticity and neural reuse come in the form of manipulations not of neural wiring, but of the task environment, and in particular manipulations of the order in which tasks are learned. Before local neural circuits have acquired very specific workings, and before these workings have been incorporated into multiple functional complexes subserving different uses, it may well be that the most efficient way to acquire novel capacities is inducing plasticity in local circuitry. But later in development such plasticity could prove costly, and learning may favor neural reuse as a strategy. If it is the case that different tasks induce different local workings when learned early, then it might be possible to systematically investigate the conflicts between plasticity and reuse by looking for specific order effects in learning. For instance, it might be easier to learn task A after task B than after task C, even when A is learned at the same stage of development in each case. (Naturally, choosing appropriate tasks would take some ingenuity; that it will be harder to learn calculus after learning Spanish than after learning algebra is obvious and uninteresting. That it might be easier to learn simple arithmetic after manipulating objects in gloves than after manipulating objects in mittens looks potentially more interesting.) Similarly, it may be that learning task D after A and B disrupts A, but does not do so when learned after tasks A and C, because in the former case the local workings needed to allow for neural reuse as a learning strategy have not developed, leaving plasticity as the only option. Reimers suggests some similar developmental studies, and I know that the entire community eagerly awaits the release of the first analyses from the various longitudinal fMRI studies currently underway (Paus Reference Paus2010).

R3.2. Evolution or development? Both!

In short, I think that the neural reuse model is much friendlier to the developmental perspective than it might have appeared in the target article (Dekker & Karmiloff-Smith and Moore & Moore rightly point out that development was under-discussed there) and that the two perspectives together suggest some novel avenues for further exploration. I think this account may also shed some light on the issue of how fixed I take neural workings to be (Aisenberg & Henik, Immordino-Yang et al., Toskos Dils & Flusberg) and how I take them to be fixed (Michaux et al.). While I think innate cortical biases are a likely feature of our neural organization, workings emerge over time, driven by experience and task demands. Although I think the brain eventually achieves a maturity that is marked in part by the existence of strong and stable workings, plasticity always remains a possibility, whether in response to extraordinary task demands or to physical injury. In this light, I think one can understand neural reuse as a learning strategy that greatly increases the flexibility of the brain while avoiding some of the potentially disruptive effects of local plasticity (especially plasticity that occurs late in development).

This may make it sound like I am giving up on the phylogenetic claims of the massive redeployment hypothesis. I am not. Instead, I am suggesting that the evolutionary and developmental aspects of learning – especially when considered in the neural context – are typically complementary, mutually influencing, and extremely difficult to disentangle. Developmental trajectories, even those highly sensitive to local context, may nevertheless depend on specific evolutionary inheritances. Genetic effects can be influenced by environmental factors such as resource availability (Devlin et al. Reference Devlin, D'Andrade, Uh and Biagi2004), and even social factors such as parenting styles (Suomi Reference Suomi2004), and may themselves rely on typical developmental trajectories which, although not themselves hard-coded, have been driven long enough by stable environmental factors to have become established among the dependencies of the genetic pathway.

R3.3. Workings versus polyfunctionality

This may assuage some of the concerns that my workings were too fixed to account for the dynamic nature of the brain, but several commentators question the very notion of local workings. Aisenberg & Henik; Brincker; Gomila & Calvo; Jungé & Dennett; Katz; Petrov, Jilk, & O'Reilly [Petrov et al.]; and Toskos Dils & Flusberg all argue that local regions might well be natively polyfunctional, obviating the need for any explanation based on reuse. It is true that much of my imaging data is consistent with this possibility, as they show at most that neural regions subserve multiple uses, and multi-use could certainly result from the polyfunctionality of these regions. Moreover, as Jungé & Dennett, Klein, and Ritchie & Carruthers point out, the imaging data are also consistent with there being multiple local workings in close proximity, such that the multiple uses only appear to use overlapping circuitry due to the poor spatial resolution of current functional imaging techniques. And, as I noted in the target article, these data are even consistent with there being no local functions at all. If brain functions are primarily established not by the structure of local circuitry but by the relative response patterns of neurons or neural assemblies (if, that is, functions are defined by the relational rather than the local properties of neural assemblies), then multi-use could result when these assemblies cooperate with different partners, thereby establishing different relational – and therefore different functional – properties.

But imaging data demonstrating neural overlaps are not the only data I cited, and I think the broader picture sits uneasily with these possibilities. First, there are the data suggesting that more recently evolved uses are subserved by more broadly scattered neural circuits than are older uses. If we may presume an increase in the metabolic cost of establishing and maintaining more broadly scattered functional complexes, then, if polyfunctional local circuits were an option, one would expect uses to be consistently subserved by localized functional complexes. These data seem to favor the existence of local and relatively defined cortical biases. Second, there are the data on cognitive interference and cross-domain interactions. These data would appear to weigh against the possibility of multiple local workings, and favor actually shared neural components. Third – and most telling in my view – are the cases where there appears to be functional or semantic inheritance that results from the sharing of components. This suggests that the functional contributions of the shared local neural circuits are stable and detectible across multiple uses, a possibility apparently inconsistent with both relationally defined function and polyfunctionality. I recognize, of course, that these arguments are more suggestive than definitive, and will be more or less compelling depending on one's other intellectual commitments. In the end, the only evidence that can truly establish the case is the consistent discovery of local workings that can explain the multiple uses to which the local circuit is put. I am laying an empirical bet that this will prove possible, but I must recognize that the evidence may not break my way.

To counter the worries that apparent neural overlaps might be a side effect of the relatively poor spatial resolution of fMRI, Klein suggests that experiments leveraging neural adaptation effects might be in order. Functional magnetic resonance imaging adaptation, fMRIa (Krekelberg et al. Reference Krekelberg, Boynton and van Wezel2006), exploits the fact that neurons exposed to the same stimulus attenuate their responses to that stimulus over time, resulting in a decreased BOLD signal. Klein's idea is roughly this: Choose two tasks that appear to involve some overlapping neural circuits, attenuate neural responses by repeatedly engaging in one task, and then switch tasks. If the attenuation disappears in the overlapping region, this would be evidence that “fresh” neurons from a distinct neural subpopulation were responsible for the second task; whereas if the attenuation remained, this would be evidence that the very same neurons were responsible for processing in both tasks.

Let me first of all heartily endorse any and all calls for converging evidence from multiple techniques, and for multiple alternate uses of MRI in particular, for example, Diffusion Tensor Imaging (DTI) (Hagmann et al. Reference Hagmann, Cammoun, Gigandet, Meuli, Honey, Wedeen and Sporns2008); fMRI coherence analysis (Muller et al. Reference Muller, Lohmann, Bosch and von Cramon2001; Reference Muller, Mildner, Lohmann and von Cramon2003; Sun et al. Reference Sun, Miller and D'Esposito2004); Multi-Voxel Pattern Analysis (Periera et al. 2009), and so forth. Nevertheless, I do have some concerns about this particular suggestion. First, although there is good evidence for neural adaptation as the result of repeated stimuli, that is, as the result of being asked to represent the same thing, there is less evidence for task adaptation, that is, for the idea that there is a reduction in neural response as a result of being asked to do the same thing. This matters, because in most cases of neural reuse, the hypothesis is that the region of overlap is not being used to represent the same thing in both tasks, so any inherited neural response suppression between tasks would have to result from the region being asked to do the same thing, that is, express the same working in both tasks. Second, even if one were to observe neural response suppression as the result of repeating a task, it would remain difficult to interpret the outcome of any experimental manipulation. For, consider the case where the BOLD signal in a region continued to be attenuated during the second task. This could be because they use the same brain regions, and there is some inherited response suppression (apparent evidence for reuse); or it could be because practice at the first task makes the second task easier, or changes the strategy participants use to engage in it (evidence compatible with multiple views). Similarly, if the attenuation disappears, this could be because a distinct neural subpopulation in the same region was being recruited (apparent evidence against reuse); because in the two tasks the same neural populations are being asked to represent different things (compatible with reuse); or because the first task induced neural changes that interfere with performance of the second task (apparent evidence for reuse; see Glenberg et al. [2008a] for one such example, and Grill-Specter et al. [Reference Grill-Specter, Henson and Martin2006] for a discussion of various ways to interpret repetition suppression).

For these reasons, I think that fMRIa might be of limited usefulness in teasing apart “real” neural reuse from the possibility that neighboring neural populations are responsible for the different uses to which individual regions are apparently put. As noted above, better techniques for this include cross-domain interference and use-induced plasticity paradigms (Glenberg & Kashak Reference Glenberg and Kaschak2002; Glenberg et al. Reference Glenberg, Sato and Cattaneo2008a), and I certainly hope to see more such work in the near future.

Of course, the possibility that the limited spatial resolution of fMRI might be hiding some functional segregation isn't Klein's only worry about that data set. He also wonders whether fMRI activations are particularly good at identifying which brain regions are making genuinely functional contributions to a task in the first place. Rather, activation may spread around the brain network, leading to false positives: regions that are activated only as a side effect of their connectivity, and not because they are making a functional contribution to the task under investigation. He is right, of course; this is a possibility (although not one that cuts against neural reuse in particular). That is why it is crucial to have not just imaging data, but also behavioral data and, where possible, results from techniques like Transcranial Magnetic Stimulation (TMS). If applying TMS over a region thought to be functionally involved in two different tasks in fact disrupts both of those tasks, that is evidence that the activation there is not just a side effect of connectivity, but is making a genuine functional contribution. The target article purposely included data from all of these sources, but here again I would encourage and welcome more studies along these lines.

R3.4. How do workings work?

Even those willing to entertain the possibility that the brain might actually have local workings had some questions about how best to understand what they are. Ritchie & Carruthers, for instance, ask whether they ought to be understood in computational or intentional terms, and express some skepticism that they could be both computational and multi-use, since it is hard to see how the same computations could realize distinct representational properties on different occasions of use. Rather than repeat or expand upon the arguments from the target article on this matter, I would like instead to refer the reader to the very interesting suggestions made by Bridgeman; Speed, Verzi, Wagner, & Warrender [Speed et al.]; and Donnarumma, Prevete, & Trautteur [Donnarumma et al.] Bridgeman quite succinctly describes the representational power and flexibility of even fairly simple computational elements, and Donnarumma et al. offer a specific proposal for how this representational flexibility might be harnessed for multiple uses via programmable neural networks. One especially noteworthy aspect of their proposal is a solution to one apparent problem for neural reuse, mentioned also by Klein and Ritchie & Carruthers, that reused brain components might send their outputs to all their consumer complexes all the time, which would presumably result in a great deal of noise and behavioral confusion. That this does not appear to be the outcome means either that there is little neural reuse, or that the brain has managed a solution to this problem. Donnarumma et al. offer one model for how selective routing of outputs could be achieved even in a network with fixed connections.

Equally interesting is the proposal made by Speed et al. that reuse might be enabled by a mechanism similar to that employed in optical holography. Here, it is somewhat harder to understand what form local workings would take (as these commentators note, in optical holography every piece of the plate encodes the entire image, and nothing like this appears to obtain in the brain), but that massive reuse is possible on this model is quite clear; and the proposal is notable for detailing the high-level functional predictions that emerge from taking a holographic perspective. Whether these solutions resemble the ones implemented by the brain for managing reuse remains to be seen, of course, but knowing that there exist possible solutions is certainly a positive step forward. I look forward to seeing how these diverse research programs evolve.

R5.1. Some additional classes of reuse

Foglia & Grush suggest that one way to further specify the theory is to distinguish between neurophysiological reuse – the physical use of neural circuits to support multiple tasks – and representational reuse – the reuse of physical and mental stand-ins for multiple purposes. They further divide the latter category into “domain” and “functional” reuse: the reuse of a model in one domain (space) to support tasks in another domain (time) versus the reuse of a model in a single domain for multiple purposes (visual representations for both on-line processing and imagining). From their perspective, what is most striking is the brain's capacity to use and reuse models across diverse contexts. Foglia & Grush suggest that this may be the more general principle driving instances of neurophysiological reuse, and that the latter occurs when the models in question are neutrally instantiated.

I think the distinction between neurophysiological and representational (model) reuse is a good one, and I agree that our ability to reuse models across contexts is a crucial cognitive ability (Landy et al., in press; Landy & Goldstone Reference Landy and Goldstone2007a; Reference Landy and Goldstone2007b). However, I don't think it is right that the latter category falls under the former. Instead, I think these are largely but not entirely overlapping sets: There is model reuse without neural reuse (using a chessboard to play chess, and as a calendar); model reuse with neural reuse (using perceptual systems for imagination); and, I would argue, neural reuse without model reuse. For an example of the last category, consider again the case of using the fingers to represent numbers, raised by Foglia & Grush and Michaux et al. One way to use the fingers to support mathematical reasoning is to use them as models of numbers, and Butterworth (Reference Butterworth1999c) argues that this results in and explains the observed neural overlap between neural circuits involved in finger-representation and number-representation. But I think the evidence points to a different explanation for the observed overlap: infrastructural reuse without model reuse (Penner-Wilger & Anderson Reference Penner-Wilger, Anderson, Love, McRae and Sloutsky2008; submitted; Penner-Wilger Reference Penner-Wilger2009). Here, the idea is that part of the finger representation circuit just happens to have a functional structure that lends itself to supporting certain aspects of number representation. It is not because the fingers are or can be used as models (although they certainly are and can), nor is the neural circuit being used as a model of anything; it is simply being used because it can serve the purpose. Although this example involved reuse of neural infrastructure, one imagines there could be important cases of non-neural infrastructural reuse – the use of the hands as an alternative representational resource that aids learning (Goldin-Meadow Reference Goldin-Meadow2003) may be one such case.

Thus, it seems there are three large classes of cognitively relevant reuse: neural reuse, model reuse, and infrastructural reuse. None of these classes entirely reduces to the others. Foglia & Grush further divide model reuse into cross-domain and intra-domain reuse (I drop their term “functional” here, since all of this reuse seems functional to me), and, following Rozin and Fishbein et al., we can divide infrastructural reuse into structural token reuse, physiological process reuse, and developmental plan reuse. A cross-cutting category is reuse resulting in semantic inheritance, which Kiverstein has suggested has an especially important role in cognitive bootstrapping. Naturally, the target article was centrally concerned with neural reuse, and with understanding when such reuse is (merely) infrastructural, when it involves reuse of models, and when it results in semantic inheritance. But Foglia & Grush are quite right to draw our attention to the cognitive importance of non-neural reuse as well.

R5.2. Does reuse rule out modularity?

One side-effect of the apparent underspecification of neural reuse theory is that it seemed to several commentators – including Toskos Dils & Flusberg, Jungé & Dennett, and Ritchie & Carruthers – to leave neural reuse closer to modularity than my own rhetoric on the matter would indicate. For example, Jungé & Dennett suggest that a software theory of modularity that posits a modular organization at the abstract level, with no commitment about neural implementation, could survive the critique offered by neural reuse. And Ritchie & Carruthers argue that their specific version of massive modularity is in fact compatible with neural reuse. Indeed, one might argue that no largely functionalist account of the mind, insofar as it leaves open the many possibilities for implementation, would have any intellectual friction with an account of how neural resources are deployed in the service of cognitive functions. Although I can see the attraction of this position, I think it doesn't apply to the case of modularity. Any theory of modularity worthy of the name must have modules, of course, and these modules need to have some specific functional characteristics, such as relative encapsulation, or functional separability. These characteristics in fact place limits on the way such modules can be implemented and, in my view of how our brain is organized, this means they cannot be implemented there.

Ritchie & Carruthers try two (apparently incompatible) tacks to avoid this conclusion: First, they suggest that it would be possible to functionally separate two part-sharing modules via double-dissociation, just so long as one did this by disrupting parts that were not shared between them; and, second, they suggest that maybe modules don't share parts after all, since my imaging evidence is indeed compatible with there being distinct neural regions, smaller than the spatial granularity of fMRI, dedicated to individual modules. I have already discussed why I think this second possibility is not likely to obtain, but will note here that even if the brain were like that, this argument would fail to demonstrate the compatibility of massive modularity and neural reuse. For what it would show is not that these two theories were compatible, but that neural reuse was false.

Unfortunately for one of our two theories, I think the first argument fares no better. It is certainly true (and not in conflict with neural reuse) that there will be pairs of functional complexes that are functionally dissociable because they share no parts. And it is also true that even functional complexes that share some parts can be distinguished by their reactions to disruptions of the parts not shared. But although the claim that for all modules A and B it will be possible to functionally dissociate them by disrupting any of their respective parts X and Y may be too strong for even the most dedicated modularist, surely a claim of the following logical form is too weak to distinguish between any competing theories of brain organization: that there exist some modules A and B that can be distinguished by disrupting some of their respective parts X and Y. Wouldn't that be true on just about anyone's theory of brain organization? So the fact that we both accept that statement doesn't make for a particularly strong alliance. Yet, I don't see that Ritchie & Carruthers have anything stronger to offer here. And if it is the case, as reuse predicts, that in disrupting region X one might not disrupt functional complex B, but would disrupt some complex C (and often many complexes C, D, …), then even though specific pairs of functional complexes will be functionally separable, it would appear that functional separability will not be a general characteristic of the brain. But this is exactly the core claim of massive modularity. I am forced to conclude once again that the two theories are incompatible and, as noted in the target article, in fact point cognitive science in very different empirical directions.

This being said, it is important to note that the term module is used in many different ways in many different disciplines, and many of these senses of module are compatible with neural reuse. For instance, in graph theory the term module is often used to refer to a set of nodes that are highly connected to one another, and less connected with other parts of the graph. Note that this is a structural rather than a functional characterization. Modules are defined in terms of features of the abstract topology of the representational vehicle: the graph. Nevertheless, one of the reasons graphs have proven a useful representational format is that these abstract structures often identify functionally relevant features of the underlying system. In molecular and developmental biology, for instance, a “module” is a set of interacting elements – genes, gene networks, proteins, brain regions, and so forth – that make a specific, relatively context-insensitive contribution to some developmental process (Rives & Galitski Reference Rives and Galitski2003; Spirin & Mirny Reference Spirin and Mirny2003; Tong et al. Reference Tong, Lesage, Bader, Ding, Xu, Xin, Young, Berriz, Brost, Chang, Chen, Cheng, Chua, Friesen, Goldberg, Haynes, Humphries, He, Hussein, Ke, Krogan, Li, Levinson, Lu, Ménard, Munyana, Parsons, Ryan, Tonikian, Roberts, Sdicu, Shapiro, Sheikh, Suter, Wong, Zhang, Zhu, Burd, Munro, Sander, Rine, Greenblatt, Peter, Bretscher, Bell, Roth, Brown, Andrews, Bussey and Boone2004; von Dassow & Munro Reference von Dassow and Munro1999) wherever it is instantiated. This sense of module is roughly equivalent to what I have been calling a functional complex, and is perfectly compatible with the notion that the elements of the functional “module” cooperate with different sets of partners to support other outcomes in other circumstances. And, indeed, we know that developmental modules share parts and are often nested as components of larger modules (Jablonka & Lamb Reference Jablonka and Lamb2006; Schlosser & Wagner Reference Schlosser and Wagner2004).

This is a perfectly viable use of the term module, but note that these modules are individuated in ways quite distinct from the mental modules posited by evolutionary psychology (e.g., Pinker 1997). Mental modules are entities with abstract functional characteristics (encapsulation, domain specificity, functional separability, etc.) and flexible structural characteristics. In contrast, typical biological modules (as found in gene networks or neural co-activation graphs, for example) are entities with well-defined abstract structural characteristics but flexible functional characteristics. As tidy as it would be for neuroscience if the modules in neural co-activation graphs identified brain structures with the functional features of mental modules, that is not the way the brain is organized. Therefore, it is important in debates about brain organization and function to try to keep the different senses of “module” distinct; it is all too easy to let them blur into one another.

R5.3. Reuse and evolutionary theory

One place where neural reuse theory is somewhat short on specifics involves its precise fit with existing evolutionary theory. The massive redeployment hypothesis, for instance, is based in part on an overly simplified, armchair-evolutionary story. I think we can and should do better than this. Thankfully, several of the commentators point the way to a better integration of reuse and evolution.

Moore & Moore and Bergeron both suggest that the concept of homology can serve as an organizing framework for the further exploration of reuse in evolutionary and developmental context. Bergeron argues that we ought to search for cross-species cognitive homologies – workings with the same phylogenetic origins serving different uses in different animals – and he offers some evidence that the search will prove fruitful. Such discoveries would not only help further specify the evolutionary mechanisms behind neural reuse, but could also offer some unique insights into the cognitive relatedness of various species.

Naturally, such a project would involve a great deal of comparative work. Katz and Niven & Chittka are rightly dismayed by the dearth of comparative data (in my defense, I plead lack of both expertise and space). These authors offer many examples of reuse in other species, and Katz in particular offers evidence for just the sorts of cognitive homologies that Bergeron suspects should exist. One interesting upshot from the commentaries of both Katz and Niven & Chittka is that invertebrates may prove the most promising class of animals for initial investigations. All I can say is I think that sounds like a great idea, and hope that someone – if not these authors, then some enterprising young graduate students – will successfully take up the challenge.

Moore & Moore have a somewhat different take on the same general idea. They argue that the concept of homology can also be applied in developmental context, whenever two (or more) psychological traits or processes share a neural circuit that has been put to different uses. In this case, we may have identified a developmental homology, a shared ontogenetic “ancestor” circuit being used in different ways. I agree that this perspective offers the possibility of leveraging useful analogies from the evolutionary literature to form hypotheses about features of developmental change, and think that it can help continue the important process of integrating these two perspectives. There are some disanalogies to be careful of as well, however. Chief among these is the fact that traditional evolutionary and cognitive homologies in different species are far less functionally entangled than developmental homologies in a single animal. Whatever limitations imposed by the nature of the inheritance, the use one species subsequently makes of that inheritance does not affect the uses made by others. This is not the case with developmental homologies, where subsequent use can affect the functional properties of other uses, if only because of increased processing demand on the shared circuit. Thus, in cross-species homologies it is more possible to alter the properties of the underlying inheritance, whereas in a developmental homology changing the nature of the shared circuit could have deleterious consequences. Nevertheless, when properly noted, I think both the analogies and the disanalogies will prove a fruitful source of developmental hypotheses for future investigations.

Exploring the parallels with reuse in genetics offers another very promising avenue both for hypothesis generation and for coming to better understand the mechanisms of neural reuse. As Reimers details, there are many examples of reuse in molecular evolution: Protein domains can be used for multiple purposes; novel metabolic pathways are often assembled by reusing and adapting parts from existing pathways; and signaling pathways are widely reused throughout development. That there is a such a pattern of structure-to-function mapping at the molecular level suggests, among other things, that the neural overlaps I uncovered by reviewing fMRI experiments are not going to go away with higher-resolution imaging techniques. There is too much to be gained functionally by taking advantage of reuse and recombination for this strategy, evident at the micro level, to be absent from the macro level.

R5.4. A history of reuse

Bridgeman and Lia both do the community a service by placing the neural reuse hypothesis in historical context, pointing out some intellectual forebears of the idea in addition to those identified in the target article. Awareness of history is important to scientific progress, for while intellectual cycles are an inevitable by-product of the social structure of science, we can at least try to notice whether we are on a spiral staircase or a high-school track. Right now, the cognitive sciences feel to me more like the former than the latter – and neural reuse theory seems a genuine advance – but others should of course make their own judgments.