2.1. Anatomical modularity

Anatomical modularity is functional modularity plus a strong thesis about how the functional modules are implemented in the brain. Functional modularity is (minimally) the thesis that our cognitive systems are composed of separately modifiable (or “nearly decomposable”; Simon Reference Simon1962/1969) subsystems, each typically dedicated to specific, specialized functions (see sect. 3.1 for a discussion). Anatomical modularity is the additional thesis that each functional module is implemented in a dedicated, relatively small, and fairly circumscribed piece of neural hardware (Bergeron Reference Bergeron2007).

Simply put, neural reuse theories suggest anatomical modularity is false. According to the picture painted by reuse, even if there is functional modularity (see sect. 3.1), individual regions of the brain will turn out to be part of multiple functional modules. That is, brain regions will not be dedicated to single high-level tasks (“uses”), and different modules will not be implemented in separate, small, circumscribed regions. Instead, different cognitive functions are supported by putting many of the same neural circuits together in different arrangements (M. L. Anderson Reference Anderson2008a). In each of these arrangements, an individual brain region may perform a similar information-processing operation (a single “working”), but will not be dedicated to that one high-level use.

Although there are few defenders of a strong anatomical modularity hypothesis, Max Coltheart (Reference Coltheart and Rapp2001) goes so far as to include it as one of the fundamental assumptions guiding cognitive neuropsychology. The idea is that the success of neuropsychological research – relying as it does on patients with specific neurological deficits, and the discovery of double-dissociations between tasks – both requires and, in turn, supports the assumption that the brain is organized into anatomical modules. For if it were not, we wouldn't observe the focal deficits characteristic of some brain injuries, and nor would we be able to gather evidentiary support for double-dissociations between tasks.

If this argument were sound, then the success of neuropsychology as a discipline would itself be prima facie evidence against neural reuse. In fact, the inference is fairly weak. First, it is possible for focal lesions to cause specific functional deficits in non-modular systems (Plaut Reference Plaut1995), and double-dissociations do not by themselves support any inference about the underlying functional architecture of the brain (Van Orden et al. Reference Van Orden, Pennington and Stone2001). In any event, such deficits are the exception rather than the rule in human brain injuries. Even some of the patients most celebrated for having specific behavioral deficits often have multiple problems, even when one problem is the most obvious or debilitating (see Bergeron Reference Bergeron2007; Prinz Reference Prinz and Stainton2006 for discussions). The evidence coming from neuropsychology, then, is quite compatible with the truth of neural reuse. But is neural reuse compatible with the methodological assumptions of cognitive neuropsychology? Section 7 will discuss some of the specific methodological changes that will be needed in the cognitive neurosciences in light of widespread neural reuse.

2.2. Optimal wiring hypotheses

The layout of neurons in the brain is determined by multiple constraints, including biomorphic and metabolic limitations on how big the brain can be and how much energy it can consume. A series of studies by Christopher Cherniak and others has reported that the layout of the nervous system of C. elegans, the shape of typical mammalian neuron arbors, and the placement of large-scale components in mammalian cortex are all nearly optimal for minimizing the total length of neurons required to achieve the structure (Cherniak et al. Reference Cherniak, Mokhtarzada, Rodrigues-Esteban and Changizi2004; see also Wen & Chklovskii Reference Wen and Chklovskii2008). The last finding is of the highest relevance here. Cherniak et al. examined the 57 Brodmann areas of cat cortex. Given the known connections between these regions, it turns out that the Brodmann areas are spatially arranged so as to (nearly) minimize the total wiring length of those connections.

This is a striking finding; and even though this study examined physical and not functional connectivity, the two are undoubtedly related – at least insofar as the rule that “neurons that fire together wire together” holds for higher-level brain organization. In fact, Cherniak et al. (Reference Cherniak, Mokhtarzada, Rodrigues-Esteban and Changizi2004) predict that brain areas that are causally related – that co-activate, for instance – will tend to be physically adjacent. The data reviewed above did not exactly conform to this pattern. In particular, it seems that the neural regions supporting more recent cognitive functions tended to be less adjacent – farther apart in the brain – than those supporting older cognitive functions. Neverthless, neural reuse and the global optimization of component layout appear broadly compatible, for four reasons. First, wiring length can hardly be considered (and Cherniak et al. do not claim that it is) the only constraint on cortical structure. The total neural mass required to achieve the brain's function should also be kept minimal, and reuse would tend to serve that purpose. Second, it should be kept in mind that Cherniak et al. (Reference Cherniak, Mokhtarzada, Rodrigues-Esteban and Changizi2004) predict global optimization in component layout, and this is not just compatible with, but also positively predicts that subsets of components will be less optimal than the whole. Third, there is no reason to expect that all subsets will be equally suboptimal; global optimality is compatible with differences in the optimality of specific subsets of components. Fourth, when there is a difference in the optimality of component subsets, neural reuse would predict that these differences would track the evolutionary age of the supported function. That is, functionally connected components supporting recently evolved functions should tend to be less optimally laid out than those supporting older functions. More specifically, one would expect layout optimality to correlate with the ratio of the age of the cognitive function to the total evolutionary age of the organism. When functional cooperation emerged early in the evolution of the cortex, there is a greater chance that the components involved will have arrived at their optimal locations, and less chance of that for lower ratios, as overall brain morphology will not have had the same evolutionary opportunity to adjust.

This notion is not at all incompatible with the thesis of global (near-) optimality and indeed might be considered a refinement of its predictions. Certainly, this is a research direction worth pursuing, perhaps by merging the anatomical connectivity data-sets from Hagmann et al. (Reference Hagmann, Cammoun, Gigandet, Meuli, Honey, Wedeen and Sporns2008) with functional databases like BrainMap (Laird et al. Reference Laird, Lancaster and Fox2005) and the NICAM database (M. L. Anderson et al. Reference Anderson, Brumbaugh, Şuben, Chaovalitwongse, Pardalos and Xanthopoulos2010). In fact, I am currently pursuing a related project, to see whether co-activation strength between regions predicts the existence of anatomical connections.

3.1. Massive modularity

As noted above, functional modularity is minimally the thesis that the mind can be functionally decomposed into specialized, separately modifiable subsystems – individual components charged with handling one or another aspect of our mental lives. Carruthers (Reference Carruthers2006) follows this formulation:

Massive modularity, which grows largely out of the modularity movement in evolutionary psychology (Pinker Reference Pinker1997; Sperber Reference Sperber1996; Tooby & Cosmides Reference Tooby, Cosmides, Barkow, Cosmides and Tooby1992) is the additional thesis that the mind is mostly, if not entirely, composed of modules like this – largely dissociable components that vary independently from one another. Is such a vision for the mind's architecture compatible with widespread neural reuse? Carruthers (Reference Carruthers2006) certainly thinks so:

As much as I appreciate Carruthers' swift adoption of the redeployment hypothesis, I am troubled by some aspects of this argument. First, it appears to contain a false premise: Energetic constraints predict more compact or localized, not necessarily fewer brain systems. Second, it may be logically invalid, because if functions must be underlain by separately modifiable systems, then they cannot be built from shared parts. That is, it appears that this apparently small concession to neural reuse in fact undermines the case for massive modularity.

Consider Carruthers' hi-fi system analogy. There it is true that the various components might share the amplifier and the speakers, the way many different biological functions – eating, breathing, communicating – “share” the mouth. But if neural reuse is the norm, then circuit sharing in the brain goes far beyond such intercommunication and integration of parts. The evidence instead points to the equivalent of sharing knobs and transistors and processing chips. A stereo system designed like this would be more like a boom-box, and its functional components would therefore not be separately modifiable. Changing a chip to improve the radio might well also change the performance of the tape player.Footnote ⁶

To preview some of the evidence that will be reviewed in more detail in section 4, the brain may well be more boom-box than hi-fi. For instance, Glenberg et al. (Reference Glenberg, Sato and Cattaneo2008a) report that use-induced motor plasticity also affects language processing, and Glenberg et al. (Reference Glenberg, Sato, Cattaneo, Riggio, Palumbo and Buccino2008b) report that language processing modulates activity in the motor system. This connection is confirmed by the highly practical finding that one can improve reading comprehension by having children manipulate objects (Glenberg et al. Reference Glenberg, Brown and Levin2007). And of course there are many other such examples of cognitive interference between different systems that are routinely exploited by cognitive scientists in the lab.

This does not mean that all forms of functional modularity are necessarily false – if only because of the myriad different uses of that term (see Barrett & Kurzban Reference Barrett and Kurzban2006 for a discussion). But it does suggest that modularity advocates are guided by an idealization of functional structure that is significantly at odds with the actual nature of the system. 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” (Mitchell Reference Mitchell2006). Rather than approach a complex system by breaking functions into subfunctions and assigning functions to proper parts – a heuristic that has been quite successful across a broad range of sciences (Bechtel & Richardson Reference Bechtel and Richardson1993; Reference Bechtel and Richardson2010) – network thinking suggests one should look for higher-order features or patterns in the behavior of complex systems, and advert to these in explaining the functioning of the system. The paradigm exemplars for this sort of approach come from the discovery of common, functionally relevant topological structures in various kinds of networks, ranging from human and insect social networks to the phone grid and the Internet, and from foraging behaviors to the functioning of the immune system (Barabási & Albert Reference Barabási and Albert1999; Barabási et al. Reference Barabási, Albert and Jeong2000; Boyer et al. Reference Boyer, Miramontes, Ramos-Fernández, Mateos and Cocho2004; Brown et al. Reference Brown, Liebovitch and Glendon2007; Jeong et al. Reference Jeong, Tombor, Albert, Oltvai and Barabási2000; Newman et al. Reference Newman, Barabasi and Watts2006). Although it is hardly the case that functional decomposition is an ineffective strategy in cognitive science, the evidence outlined above that patterns of neural co-activation distinguish between cognitive outcomes more than the cortical regions involved do by themselves suggests the need for a supplement to business as usual.

Even so, there are (at least) two objections that any advocate of modularity will raise against the picture of brain organization that is being painted here: Such a brain could not have evolved, because (1) the structure would be too complex, and (2) it would be subject to too much processing interference and inefficiency.

Carruthers (Reference Carruthers2006) follows Simon (Reference Simon1962/1969) in making the first argument:

The argument from design set forth here is more convincing when it is applied to the original emergence of a complex system than when it is applied to its subsequent evolutionary development. What the argument says is that it must be possible for development to be gradual, with functional milestones, rather than all-or-nothing; but neural reuse hardly weakens the prospect of a gradual emergence of new functions. And the possibility that new functionality can be achieved by combining existing parts in new ways – which undermines independent variation and separate modifiability, as Carruthers (Reference Carruthers2006) admits, here – suggests that a modular architecture is only one possible outcome from such gradualism.

Moreover, the strong analogy between natural selection and a designer may not be the most helpful conceptual tool in this case. When one thinks about the brain the way a human designer would, the problem that neural reuse presents is one of taking a given concrete circuit with a known function and imagining novel uses for it. That this process can be very difficult appears to place a heavy burden on reuse theories: How could such new uses ever be successfully designed? But suppose instead that, in building a given capacity, one is offered a plethora of components with unknown functions. Now the task is quite different: Find a few components that do something useful and can be arranged so as to support the current task – whatever their original purpose. Thus is a problem of design imagination turned into a problem of search. Evolution is known to be quite good at solving problems of the latter sort (Newell & Simon Reference Newell and Simon1976), and it is useful to keep this alternate analogy for the evolutionary process in mind here.

This brings us to the second objection, that non-modular systems would suffer from disabling degrees of interference and processing inefficiency. Here, it may be useful to recall some of the main findings of the situated/embodied cognition movement (M. L. Anderson Reference Anderson2003; Chemero Reference Chemero2009; Clark Reference Clark1997; Reference Clark, Bechtel and Graham1998). Central to the picture of cognition offered there is the simple point that organisms evolve in a particular environment to meet the particular survival challenges that their environment poses. Situated/embodied cognition emphasizes that the solutions to these problems often rely in part on features of the environments themselves; for example, by adopting heuristics and learning biases that reflect some of the environments' structural invariants (Gigerenzer et al. Reference Gigerenzer and Todd1999; Gilovitch et al. 2002). One such useful feature of most environments is that they don't pose all their problems all at once – inclement weather rarely comes along with predator abundance, pressing mating opportunities, and food shortages, for instance. And often when there are competing opportunities or challenges, there will be a clear priority. Thus, an organism with massively redeployed circuitry can generally rely on the temporal structure of events in its environment to minimize interference. Were this environment-organism relationship different – or if it were to change – then neural reuse does predict that increased interference will be one likely result.

Interestingly, contemporary humans encounter just such a changed organism-environment relationship in at least two arenas, and the effect of reused circuitry can often be seen as a result: First, in the labs of some cognitive scientists, who carefully engineer their experiments to exploit cognitive interference of various sorts; and, second, at the controls of sophisticated machinery, where the overwhelming attentional demands have been observed to cause massive processing bottlenecks, often with dangerous or even deadly results (Fries Reference Fries2006; Hopkin Reference Hopkin1995). It is no coincidence that, in addition to designing better human-machine interfaces, one important way of minimizing the problems caused by processing bottlenecks is to engineer the environment, including, especially, changing its task configuration and social structure, for instance by designing more efficient teams (Hutchins Reference Hutchins1995).

3.2. ACT-R

At the core of ACT-R is the notion of a cognitive architecture, “a specification of the structure of the brain at a level of abstraction that explains how it achieves the function of the mind” (J. R. Anderson Reference Anderson2007, p. 7). ACT-R is explicitly modular. As of ACT-R 6.0, it consisted of eight functionally specialized, domain-specific, relatively encapsulated, independently operating, and separately modifiable components. Given J. R. Anderson's definition of a cognitive architecture, it might seem to directly follow that ACT-R is committed to the notion that the brain, too, consists of functionally specialized, domain-specific, relatively encapsulated, independently operating, and separately modifiable regions that implement the functional modules of the ACT-R model. Certainly, recent experiments meant to associate ACT-R components with specific brain regions encourage this impression (J. R. Anderson Reference Anderson2007; J. R. Anderson et al. Reference Anderson, Qin, Junk and Carter2007). As he argues:

Given that neural reuse implies that anatomical modularity is false (see sect. 2.1), success in assigning ACT-R modules to specific brain regions would seem to be a problem for neural reuse, and evidence for neural reuse would appear to create problems for ACT-R. But the conclusion does not follow quite so easily as it seems. First, ACT-R does not strictly imply anatomical modularity. ACT-R is committed to the existence of functional modules, and to the existence of elements of the brain that implement them. If it turned out that activity in the ACT-R goal module was a better fit to the coordinated activity of some non-contiguous set of small brain regions than it was to the anterior cingulate (to which they currently have the goal module mapped), then this would count as progress for ACT-R, and not a theoretical setback. Similarly, if it turned out that some of the brain regions that help implement the goal module also help implement the imaginal module, this would pose no direct challenge to ACT-R theory.Footnote ⁷

Therefore, although J. R. Anderson is at pains to deny he is a functionalist – not just any possible mapping of function to structure will count as a success for ACT-R – there is a good deal of room here for alternatives to the simple 1:1 mapping that he and other ACT-R theorists are currently exploring. For its part, neural reuse predicts that the best fit for ACT-R modules, or any other high-level functional components, is much more likely to be some cooperating complex of multiple brain regions than it is a single area, and that brain regions involved in implementing one ACT-R function are likely to be involved in implementing others as well. Interestingly, this is more or less what J. R. Anderson et al. (Reference Anderson, Qin, Junk and Carter2007) found. For every task manipulation in their study, they found several brain regions that appeared to be implicated. And every one of their regions of interest was affected by more than one factor manipulated in their experiment. Thus, despite their methodological commitment to a 1:1 mapping between modules and brain regions, J. R. Anderson et al. (Reference Anderson, Qin, Junk and Carter2007) are quite aware of the limitations of that approach:

Here, the regulative idealization promoted by decomposition and localization may have unduly limited the sorts of methodological and inferential tools that they initially brought to bear on the project. As noted already in section 1, one of the contributions neural reuse may be able to make to cognitive science is an alternate idealization that can help guide both experimental design and the interpretation of results (M. L. Anderson et al. Reference Anderson, Brumbaugh, Şuben, Chaovalitwongse, Pardalos and Xanthopoulos2010).

Going forward there is at least one other area where we can expect theories of neural reuse and modular theories like ACT-R to have significant, bidirectional critical contact. Right now, ACT-R is not just theoretically, but also literally modular: it is implemented as a set of independent and separately modifiable software components. It does not appear, however, that separate modifiability is theoretically essential to ACT-R (although it is no doubt a programming convenience). Therefore, implementing overlaps in ACT-R components in light of the evidence from neuroimaging and other studies of the sort recounted here is likely to offer scientific opportunities to both research communities (see Stewart & West Reference Stewart and West2007 for one such effort). For example, overlaps in implementation might offer a natural explanation and a convenient model for certain observed instances of cognitive interference, such as that between language and motor control (Glenberg & Kaschak Reference Glenberg and Kaschak2002) or between memory and audition (Baddeley & Hitch Reference Baddeley, Hitch and Bower1974), helping to refine current hypotheses regarding the causes of the interference.

The ACT-R community is already investigating similar cases, where different concurrent tasks (dialing the phone while driving) require the use of the same ACT-R module, and thus induce performance losses (Salvucci Reference Salvucci2005). Altering ACT-R so that different modules share component parts might enable it to model some cognitive phenomena that would otherwise prove more difficult or perhaps impossible in the current system, such as the observation that object manipulation can improve reading comprehension (Glenberg et al. Reference Glenberg, Brown and Levin2007). Finally, observations of interference in a modified ACT-R but not in human data, might suggest that the ACT-R modules did not yet reflect the correct division of the mind's functions. Such conflicts between model and data could be leveraged to help ACT-R better approximate the high-level functional structure of the mind.

3.3. Classic parallel distributed processing

It is of course true that from a sufficiently abstract perspective, the idea of neural reuse in cognitive functioning is nothing new. It has been a staple of debates on brain architecture at least since the advent of parallel distributed processing (PDP) models of computation (Rummelhart & McClelland Reference Rumelhart and McClelland1986). For one widely cited example, consider the following from Mesulam (Reference Mesulam1990). He writes:

Broadly speaking, neural reuse theories are one of a family of network approaches to understanding the operation of the brain. They share with these an emphasis on cooperative interactions, and an insistence on a non-modular, many-to-many relationship between neural-anatomical sites and complex cognitive functions/behaviors. But there are also some important differences that set neural reuse apart.

First is a better appreciation of the computational work that can be done by very small groups of, or even individual, neurons (Koch & Segev Reference Koch and Segev2000). Neural reuse theories all agree that most of the interesting cognitive work is done at higher levels of organization, but they also emphasize that local circuits have specific and identifiable functional biases. In general, these models make a strong distinction between a “working” – whatever specific computational contribution local anatomical circuits make to overall function – and a “use,” the cognitive purpose to which the working is put in any individual case. For neural reuse theories, anatomical sites have a fixed working, but many different uses.

In contrast, note that in Figure 2 “neural computations” are located at Plane 2, parallel distributed processing. This reflects the belief that computational work can only be done by fairly large numbers of neurons, and that responsibility for this work can only be assigned to the network as a whole. Put differently, on PDP models there are no local workings. Classic PDP models are indeed a powerful way to understand the flexibility of the brain, given its reliance on relatively simple, relatively similar, individual elements. But the trouble for PDP models in this particular case is that there is no natural explanation for the data on increasing scatter of recently evolved functions, nor for the data on the cross-cultural invariance in the anatomical locations of acquired practices (see sect. 6.3). Indeed, on PDP models, investigating such matters is not even a natural empirical avenue to take. This represents a significant distinction between PDP and neural reuse.

Figure 2.

Detail of Figure 3 from Mesulam (Reference Mesulam1990). Reprinted with permission of the author.

Other important differences between neural reuse and classic PDP models flow from the above considerations, including the way neural reuse integrates the story about the cognitive architecture of the brain into a natural story about the evolution and development of the brain. In a sense, neural reuse theories make some more specific claims than generalized PDP – not just that the brain is a kind of network, but that it is a kind of network with functional organization at more levels than previously thought. As can be seen already in the evidence outlined above, and will be seen in greater detail in sections 5 and 6, this specificity has led to some interesting and empirically testable implications for the brain's overall functional organization.

3.4. Contemporary parallel distributed processing models

More contemporary versions of network models, such as Leabra (O'Reilly Reference O'Reilly1998; O'Reilly & Munakata Reference O'Reilly and Munakata2000) tend to be composed of densely connected, locally specialized networks that are sparsely connected to one another (see Fig. 3).

Figure 3.

Overview of the Leabra architectural organization. Reprinted from Jilk et al. (Reference Jilk, Lebiere, O'Reilly and Anderson2008) with permission of the authors.

In one sense, Leabra appears to be more compatible with neural reuse than classic PDP models are, as Leabra explicitly allows for regional functional biases. But insofar as this new architecture reflects the influence of the selectivity assumption, and represents a more modularist approach to understanding the brain, then there are potentially the same points of conflict with Leabra as there are with those theories. Consider the following, from a recent paper describing Leabra:

There is nothing here that explicitly commits the authors to the idea that large brain regions are dedicated to specific tasks or cognitive domains – something the data presented here throw into question – although that is certainly one possible reading of the passage. Moreover, O'Reilly (Reference O'Reilly1998) tends to focus more on modeling processes over modeling parts, an approach that need not commit one to a specific story about how and where such processes are implemented in the brain – it needn't be the case that individual brain regions implement the processes being modeled, for instance.

And yet, O'Reilly and his collaborators have assigned these processes to specific regions:

And, in fact, the Leabra team has gone further than this by recently integrating Leabra with ACT-R to form the SAL architecture:

So it is not clear just what commitments Leabra has to modularity and localization. As with ACT-R, there doesn't seem to be anything essential to Leabra that would prevent it from explicitly incorporating neural reuse as one of its organizing principles. In particular, the functional specializations being ascribed to the brain regions mentioned are general enough to plausibly have many different cognitive uses, as predicted by neural reuse theories. But, as with ACT-R, more research will be needed before it becomes clear just how compatible these visions for the functional organization of the brain in fact are. The notion of neural reuse cuts across some old divisions – localization versus holism; modular versus connectionist – and whether theories falling on one or another side of each dichotomy are compatible with the notion of neural reuse will ultimately depend on how their advocates interpret the theories, and how flexible their implementations turn out to be.

4.1. Reuse of motor control circuits for language

A great deal of the effort to discover the specific neural underpinnings of higher cognitive functions has focused on the involvement of circuits long associated with motor control functions. In a typical example of this sort of investigation, Pulvermüller (Reference Pulvermüller2005) reports that listening to the words “lick,” “pick,” and “kick” activates successively more dorsal regions of primary motor cortex (M1). The finding is consistent both with the idea that comprehending these verbs relies on this motor activation, insofar as the concepts are stored in a motoric code, and also with the idea that understanding these verbs might involve (partial) simulations of the related actions. Either interpretation could easily be used as part of the case for concept empiricism.

Similarly, Glenberg and Kaschak (Reference Glenberg and Kaschak2002) uncover an interesting instance of the entanglement of language and action that they call the “action-sentence compatibility effect” (ACE). Participants are asked to judge whether a sentence makes sense or not and to respond by pressing a button, which requires a move either toward or away from their body. In one condition “yes” is away and “no” is toward; another condition reverses this. The sentences of interest describe various actions that would also require movement toward or away, as in “put a grape in your mouth,” “close the drawer,” or “you gave the paper to him.” The main finding is of an interaction between the two conditions, such that it takes longer to respond that the sentence makes sense when the action described runs counter to the required response motion. More striking, this was true even when the sentences described abstract transfers, such as “he sold his house to you,” which imply a direction without describing a directional motor action. Following the reasoning originally laid out by Sternberg (Reference Sternberg1969), an interaction between two manipulated factors implies at least one shared component between these two different processes – movement and comprehension. A likely candidate for this component would be a neural circuit involved in motor control, a supposition confirmed by Glenberg (2008b).Footnote ¹¹ Thus, this seems another clear case in which motor control circuits are involved in, and perhaps even required for, language comprehension, whether via simulation (e.g., in the concrete transfer cases), metaphorical mapping (e.g., in the abstract transfer cases), or by some other mechanism. Glenberg has suggested both that the effect could be explained by the activation of relevant action schemas (Glenberg et al. Reference Glenberg, Sato, Cattaneo, Riggio, Palumbo and Buccino2008b) and by the activation and combination of appropriate affordances (Glenberg & Kaschak Reference Glenberg and Kaschak2002; Glenberg et al. Reference Glenberg, Becker, Klötzer, Kolanko, Müller and Rinck2009). Whatever the precise mechanism involved, the finding has been widely interpreted as support for both concept empiricism and for conceptual metaphor theory (although see M. L. Anderson Reference Anderson, Calvo and Gomila2008c for a dissent).

4.2. Reuse of motor control circuits for memory

Another interesting description of the motor system's involvement in a different cognitive domain comes from Casasanto and Dijkstra (Reference Casasanto and Dijkstra2010), who describe bidirectional influence between motor control and autobiographical memory. In their experiment, participants were asked to retell memories with either positive or negative valence, while moving marbles either upward or downward from one container to another. Casasanto and Dijkstra found that participants retrieved more memories and moved marbles more quickly when the direction of movement was congruent with the valence of the memory (upward for positive memories, downward for negative memories). Similarly, when participants were asked simply to relate some memories, without prompting for valence, they retrieved more positive memories when instructed to move marbles up, and more negative memories when instructed to move them down. Because the effect is mediated by a mapping of emotional valence on a spatial schema, the finding seems to most naturally support conceptual metaphor theory. The fact that the effect was bidirectional – recounting memories affected movement and movement affected memory retrieval – is a striking detail that seems to suggest direct neural support for the mapping.Footnote ¹²

4.3. Reuse of circuits mediated by spatial cognition

Many of the apparent overlaps between higher-order cognition and sensorimotor systems appear to be mediated by spatial schemas in this way. For example, Richardson et al. (Reference Richardson, Spivey, Barsalou and McRae2003) report that verbs are associated with meaning-specific spatial schemas. Verbs like “hope” and “respect” activate vertical schemas, whereas verbs like “push” and “argue” activate horizontal ones. As the authors put it, “language recruits spatial representations during real-time comprehension.” In a similar vein, Casasanto and Boroditsky (Reference Casasanto and Boroditsky2008) suggest that our mental representations of time are built upon the foundations of our experience with space. These findings appear to provide strong and relatively unproblematic support for conceptual metaphor theory, and perhaps also for a generic theory of concept empiricism, according to which the content of our concepts is grounded in (but does not necessarily constitute a simulation or reactivation of) sensorimotor experiences.

On the other hand, even when simulation is an important aspect of the reuse of resources between different domains, it does not always play the functional role assigned it by concept empiricism or conceptual metaphor theory. For some time, there has been growing evidence that doing actions, imagining actions, and watching actions done by others all activated similar networks of brain regions (Decety et al. Reference Decety, Sjoholm, Ryding, Stenberg and Ingvar1990; Decety et al. Reference Decety, Grezes, Costes, Perani, Jeannerod, Procyk, Grassi and Fazio1997; Jeannerod Reference Jeannerod1994). This has suggested to many that social cognition – understanding the actions and intentions of others – could involve simulating our own behaviors, a notion that attracted even more widespread interest after the discovery of mirror neurons (Decety & Grèzes Reference Decety and Grèzes1999; Gallese et al. Reference Gallese, Fadiga, Fogassi and Rizzolatti1996; Gallese & Goldman Reference Gallese and Goldman1998; Rizzolati et al. 1996). The trouble for concept empiricism and conceptual metaphor theory is that the logic governing the reuse of resources for multiple purposes is quite different in this case. Here, the idea is that circuits associated with behavioral control can be used to build predictive models of others, by inputting information about another agent into the system that would normally be used to guide one's own actions (and reactions). Although it could be argued that using simulation in support of such “mindreading” (Gallese & Goldman Reference Gallese and Goldman1998) requires a kind of metaphorical mapping (he is like me in relevant ways), in fact this is simply a necessary assumption to make the strategy sensible, and does not play the role of a domain-structuring inheritance.

Even some of the evidence for the reuse of spatial operations in other cognitive domains – which has been a mainstay of research into concept empiricism and conceptual metaphor theory – suggests the existence of more kinds of reuse than can be accounted for by these theoretical frameworks. Consider just a few of the various manifestations of the spatial-numerical association of response codes (SNARC) effect (Dehaene et al. Reference Dehaene, Bossini and Giraux1993): (1) When participants are asked to judge whether numbers are even or odd, responses are quicker for large numbers when made on the right side of space (canonically with the right hand, although the effect remains if responses are made while hands are crossed) and quicker for smaller numbers when responses are made on the left side of space. (2) Participants can accurately indicate the midpoint of a line segment when it is composed of neutral stimuli (e.g., XXXXX), but are biased to the left when the line is composed of small numbers (e.g., 22222 or twotwotwo) and to the right when the line is composed of large numbers (e.g., 99999 or nineninenine). (3) The presentation of a number at the fixation point prior to a target detection task will speed detection on the right for large numbers and to the left for small numbers. Hubbard et al. (Reference Hubbard, Piazza, Pinel and Dehaene2005) hypothesize that the SNARC effect can be accounted for by the observed reuse in numeric cognition of a particular circuit in left inferior parietal sulcus that plays a role in shifting spatial attention. Briefly, the idea is that among the representational formats we make use of in numerical cognition there is a mental “number line,” on which magnitudes are arrayed from left to right in order of increasing size. Once numerals are arrayed in this format, it is natural to reuse the circuit responsible for shifting spatial attention for the purpose of shifting attention between positions on this line. The resulting magnitude-influenced attentional bias can explain the SNARC effect.

This redeployment of visuo-spatial resources in support of alternate cognitive uses is somewhat difficult to explain from the standpoint of either concept empiricism or conceptual metaphor theory. In these examples, the effects would not be accounted for by the fact that numbers might be grounded in or involve simulations of basic sensorimotor experience, nor is it immediately obvious what metaphorical mapping might be implicated here. In fact, if the reuse of spatial schemas were in support of some semantically grounding structural inheritance from one domain to the other, we would expect the numbers to be arrayed vertically, with magnitude increasing with height. Instead, the reuse in this case appears driven by more abstract functional considerations. When doing certain numerical tasks, a number line is a useful representational format, and something like the visuo-spatial sketchpad (Baddeley Reference Baddeley1986) offers a convenient and functionally adequate storage medium. Similarly, reusing the spatial shifting mechanism is a sensible choice for meeting the functional requirements of the task, and need not ground any semantic or structural inheritance between the domains.

4.4. Reuse of circuits for numerical cognition

In fact, several such examples can be found in the domain of numerical cognition. Zago et al. (Reference Zago, Pesenti, Mellet, Crivello, Mazoyer and Tzourio-Mazoyer2001) found increased activation in the premotor strip in a region implicated in finger representation during multiplication performance compared to a digit reading condition. Similar findings were reported by Andres et al. (Reference Andres, Seron and Oliver2007), who found that hand motor circuits were activated during adults' number processing in a dot counting task. That these activations play a functional role in both domains was confirmed by Roux et al. (Reference Roux, Boetto, Sacko, Chollet and Tremoulet2003), who found that direct cortical stimulation of a site in the left angular gyrus produced both acalculia and finger agnosia (a disruption of finger awareness), and by Rusconi et al. (Reference Rusconi, Walsh and Butterworth2005), who found that repetitive Transcranial Magnetic Stimulation (rTMS) over the left angular gyrus disrupted both magnitude comparison and finger gnosis in adults.

Here again, this reuse of a basic sensorimotor function in an alternate cognitive domain does not seem to follow the logic of conceptual metaphor theory or concept empiricism. These theories are not making the claim that magnitudes inherit their meanings from finger representations, nor is any mathematical metaphor built in any straightforward way on our finger sense. Rather, the idea is that this neural circuit, originally developed to support finger awareness, is offering some functionally relevant resource in the domain of numerical cognition. For instance, Butterworth (Reference Butterworth1999c) suggests that the fingers provide children a useful physical resource for counting, with the neural result that the supporting circuits now overlap, while Penner-Wilger and Anderson (Reference Penner-Wilger, Anderson, Love, McRae and Sloutsky2008; submitted) suggest instead that the circuit in question might itself offer useful representational resources (such as a storage array).Footnote ¹³ This is not to question the notion that mathematical concepts and procedures are in some way grounded in sensorimotor experience (Lakoff & Núñez Reference Lakoff and Núñez2000), but this specific overlap in neural circuitry isn't straightforward to explain in the context of such grounding, nor is it anything that would have been predicted on the basis of either conceptual metaphor theory or concept empiricism. In fact, proponents of conceptual metaphor theory in mathematics tend to focus on relatively higher-level concepts like sets and investigate how our understanding of them is informed by such things as our experience with physical containers.

A similar argument can be made when considering the interrelations of speech and gesture, and the cognitive importance of the latter (see, e.g., Goldin-Meadow Reference Goldin-Meadow2003). According to Goldin-Meadow (Reference Goldin-Meadow2003), gesture is typically used not just to signal different moments in the learning process (e.g., to index moments of decision or reconsideration in a problem-solving routine), but also appears to have utility in advancing the learning process by providing another representational format that might facilitate the expression of ideas currently unsuited (for whatever reason) to verbal expression. The motor control system is here being used for a specific cognitive purpose not because it is performing semantic grounding or providing metaphorically guided domain structuring, but because it offers an appropriate physical (and spatiotemporal) resource for the task.

4.5. Reuse of perceptual circuits to support higher-order cognition

There are examples of the reuse of circuits typically associated with perception that also make the same point. Although there have certainly been studies that appear to unproblematically support concept empiricism – for example, Simmons et al. (Reference Simmons, Ramjee, Beauchamp, McRae, Martin and Barsalou2007) report the discovery of a common neural substrate for seeing colors, and for knowing about (having concepts for) color – other studies suggest that such cases represent only a small subset of a much broader phenomenon. Consider one of the earliest and most discussed cases of the reuse of neural circuits for a new purpose, the Baddeley and Hitch model of working memory (Baddeley & Hitch Reference Baddeley, Hitch and Bower1974; Reference Baddeley and Hitch1994; Baddeley Reference Baddeley1986; Reference Baddeley and Gazzaniga1995). One strategy for remembering the items on a grocery list or the individual numbers in a phone number involves (silently) saying them to one's self (producing a “phonological loop”), which engages brain areas typically used both in speech production and in audition.

A pattern of findings supports the existence of a phonological loop, and the engagement of both inner “speaking” and inner “hearing” to support working memory (see Wilson Reference Wilson2001 for a review). First, there is poor recall of similar sounding terms; second, there is poor recall of longer words; third, there is poor recall if the subject is made to speak during the maintenance period; and fourth, there is poor recall when the subject is exposed to irrelevant speech during the maintenance period. Moreover, imaging studies have found that such memory tasks cause activation in areas typically involved in speech production (Broca's area, left premotor cortex, left supplementary motor cortex, and right cerebellum) and in phonological storage (left posterior parietal cortex) (Awh et al. Reference Awh, Jonides, Smith, Schumacher, Koeppe and Katz1996).

In this interesting and complicated case, we have something of a triple borrowing of resources. First is the use of a culturally specific, acquired representational system – language – as a coding resource, and second is the application of a particular learned skill – silent inner speech – as a storage medium. These two together imply the third borrowing – of the neural resources used to support the first two functions. And note that all of this borrowing is done in support of what is likely an enhancement of a basic evolved function for storing small amounts of information over short periods. This raises the obvious question of whether and to what degree evolutionary pressures might have shaped the language system so that it was capable of just this sort of more general cognitive enhancement (Carruthers Reference Carruthers2002). In any case, it seems clear that this sort of borrowing is very hard to explain in terms of concept empiricism or conceptual metaphor theory. In the case of sensorimotor coding in working memory, the phonological loop is not metaphorically like speech; rather, it is a form of speech. In this, it is another instance of a straightforward functional redeployment – the reuse of a system for something other than its (apparent) primary purpose because it happens to have an appropriate functional structure.

4.6. Reuse is not always explained by conceptual metaphor theory or concept empiricism

These various examples suggest something along the following lines: One of the fundamental principles guiding reuse is the presence of a sufficient degree of functional relatedness between existing and newly developing purposes. When these functional matches result in the reuse of resources for both purposes, this history sometimes – but not always – reveals itself in the form of a metaphorical mapping between the two task domains, and sometimes, but not always, results in the inheritance or grounding of some semantic content. This way of thinking makes conceptual metaphors and “grounded” symbols into two possible side-effects of the larger process of reuse in cognition. It also muddies the causal story a bit: Planning is like locomotion because it inherits the structure of the existing domain via neural overlap; but planning also overlaps with the neural implementation base of locomotion to the degree that it is like locomotion.

The suggestion here is not that planning or communication or any other cognitive function has some predetermined Platonic structure that entirely reverses the causal direction typically supposed by conceptual metaphor theory. Rather, the idea is to point out the need to be open to a more iterative story, whereby a cognitive function finds its “neural niche” (Iriki & Sakura Reference Iriki and Sakura2008) in a process codetermined by the functional characteristics of existing resources, and the unfolding functional requirements of the emerging capacity (Deacon Reference Deacon1997).

Consider, in this regard, the particular phonemic character of human speech. A phoneme is defined by a certain posture of the vocal apparatus, and is produced by moving the apparatus toward that posture while making some noise (Fowler et al. Reference Fowler, Rubin, Remez, Turvey and Butterworth1980). Why should speech production be this way? In an article outlining their discoveries regarding the postural organization of the motor-control system, Graziano et al. (Reference Graziano, Taylor, Moore and Cooke2002b) write:

There are certainly functional characteristics that a unit of acoustic communication must have in order to adequately perform its communicative purpose, and not just any neural substrate would have had the required characteristics. But there remain degrees of freedom in how those characteristics are implemented. Speech production, then, developed its specific phonemic character as the result of the circuits on which it was built. Had the motor control system been oriented instead around (for example) simple, repeatable contractions of individual muscles – or had there been some other system with these functional characteristics available for reuse as acoustic communication was evolving – the result of the inheritance might have been a communication code built of more purely temporal elements, something closer to Morse code.Footnote ¹⁴

Finally, consider what may be a case not of the reuse of a basic sensorimotor area for higher cognitive functions, but rather the reverse. Broca's area has long been associated with language processing, responsible for phonological processing and language production, but what has recently begun to emerge is its functional complexity (Hagoort Reference Hagoort2005; Tettamanti & Weniger Reference Tettamanti and & Weniger2006). For instance, it has been shown that Broca's area is involved in many different action- and imagery-related tasks, including movement preparation (Thoenissen et al. Reference Thoenissen, Zilles and Toni2002), action sequencing (Nishitani et al. Reference Nishitani, Schürmann, Amunts and Hari2005), action recognition (Decety et al. Reference Decety, Grezes, Costes, Perani, Jeannerod, Procyk, Grassi and Fazio1997; Hamzei et al. Reference Hamzei, Rijntjes, Dettmers, Glauche and Weiller2003; Nishitani et al. Reference Nishitani, Schürmann, Amunts and Hari2005), imagery of human motion (Binkofski et al. Reference Binkofski, Amunts, Stephan, Posse, Schormann, Freund, Zilles and Seitz2000), and action imitation (Nishitani et al. Reference Nishitani, Schürmann, Amunts and Hari2005). Note that Müller and Basho (Reference Müller and Basho2004) suggest that these functional overlaps should not be understood as the later reuse of a linguistic area for other purposes, but are rather evidence that Broca's area already performed some sensorimotor functions that were prerequisites for language acquisition, and which made it a candidate for one of the neural building blocks of language when it emerged. That seems reasonable; but on the other hand, Broca's area is also activated in domains such as music perception (Tettamanti & Weniger Reference Tettamanti and & Weniger2006). While it is possible that this is because processing music requires some of the same basic sensorimotor capacities as processing language, it seems also possible that this reuse was driven by functional features that Broca's acquired as the result of its reuse in the language system, and thus by some more specific structural similarity between language and music (Fedorenko et al. Reference Fedorenko, Patel, Casasanto, Winawer and Gibson2009). Whatever the right history, this clearly represents another set of cases of functional reuse not explained by conceptual metaphor theory or concept empiricism.

Assuming the foregoing is sufficient to establish the existence of at least some cases of neural reuse that cannot be accounted for by these theoretical frameworks alone, the question naturally arises as to whether these anomalous cases should be dealt with by post-hoc elaborations of these theories (and/or by generating one or a few similarly specific theories), or whether this is a situation that calls for a global theory of reuse that supersedes and at least partially subsumes these existing frameworks. Far be it from me to argue a priori that one tack must be the correct one to take – science works best when we pursue multiple competing research paths – but one thing it might be useful to know when deciding how to spend one's research time is exactly how widespread neural reuse is. That is, the more widespread reuse appears, and the more instances of reuse that can be identified that do not involve the sensorimotor system, the stronger the justification would seem for trying to formulate a more global theory of neural reuse.

6.1. Neural exploitation hypothesis

The neural exploitation hypothesis is a direct outgrowth of conceptual metaphor theory and embodied cognition, and largely sits at the intersection of these two frameworks. The main claim of the framework is that “a key aspect of human cognition is … the adaptation of sensory-motor brain mechanisms to serve new roles in reason and language, while retaining their original function as well.” (Gallese & Lakoff Reference Gallese and Lakoff2005, p. 456) This claim is the conclusion of an argument about the requirements of understanding that runs roughly as follows:

1. 1 Understanding requires imagination. In the example most extensively developed by Gallese and Lakoff (Reference Gallese and Lakoff2005), understanding a sentence like “He grasped the cup” requires the capacity to imagine its constituent parameters, which include the agent, the object, the action, its manner, and so on.

2. 2 Imagination is simulation. Here, the neural exploitation hypothesis dovetails with concept empiricism in arguing that calling to mind individuals, objects, actions, and the like involves reactivating the traces left by perceiving, doing, or otherwise experiencing instances of the thing in question.

3. 3 Simulation is therefore neural reuse. Simulation involves reuse of the same functional clusters of cooperating neural circuits used in the original experience(s).

As much of the evidence for these claims has been laid out already in earlier sections, it won't be recounted here. The reader will of course notice that the theory as stated is limited to the adaptation of sensorimotor circuits, and we have already seen that reuse in the brain is much more broad-based than this. This is indeed a drawback of the theory, but it is nevertheless included here for two reasons: first, because it has been expanded to include not just the case of concept understanding, but also of human social understanding (Gallese Reference Gallese2008); and, second, because it incorporates a detailed computational model for how the reuse of circuitry might actually occur, based on work by Feldman and Narayanan (Reference Feldman and Narayanan2004). This model has broader applicability than is evidenced in the two main statements of the neural exploitation hypothesis (Gallese Reference Gallese2008; Gallese & Lakoff Reference Gallese and Lakoff2005).

The core of the computational model is a set of schemas, which are essentially collections of features in two layers: descriptions of objects and events and instructions regarding them. These two layers are systematically related to one another and to the sensorimotor system, such that event schemas can be used both to recognize events and to guide their execution, and object schemas can be used both to recognize objects and also to guide actions with respect to them.Footnote ²³ . The schemas are also connected to the conceptual system, such that the contents of our concepts are built from the same features that form the schemas. The general idea is that the features' connections to the sensorimotor system give semantic substance to the concepts, as well as a natural model for understanding as the activation of neurally (or, in the current case, neural-network-ly) instantiated features and schemas. Like Gallese and Lakoff (Reference Gallese and Lakoff2005), Feldman and Narayanan (Reference Feldman and Narayanan2004) focus primarily on cases of understanding that can be directly (“He grabbed the cup”) or metaphorically (“He grabbed the opportunity”) mapped to basic perception-action domains.

But there is no reason in principle that the model need be limited in that way. As the authors note, by adding layers of abstraction, one can move from concrete action execution plans to abstract recipes like mathematical algorithms. Given this flexibility, it seems that action schemas need not be limited to providing guidance for the manipulation of independent objects (whether concrete or abstract) but could presumably also become control systems for the manipulation of neural circuits. That is, the same action schema that might normally be used to control rhythmic speech production, could be reused to guide silent memory rehearsal, and more abstract schemas might form the basis of control systems for predictive modeling or other applications.Footnote ²⁴

Of course, this emendation would constitute a significant departure from the model as originally formulated.Footnote ²⁵ . In particular, it would turn a system in which neural reuse was driven by grounding – the inheritance of semantic content from one level to another – into one in which reuse was driven by the need to create control systems for functionally relevant outcomes. Although it is far from clear that this switch precludes the possibility that grounding plays a role in driving neural reuse, it certainly moves it from center stage, which may have undesirable theoretical consequences for the theory as a whole, and for the way it interfaces with related ideas in linguistics, philosophy, and psychology. On the other hand, without some emendation that significantly broadens the kinds of cases that it can cover, the neural exploitation hypothesis risks being inadequate to the full range of available empirical evidence. We will return to these issues when we come to our general discussion of the four candidate theories.

6.2. The shared circuits model

The shared circuits model (Hurley Reference Hurley, Hurley and Chater2005; Reference Hurley2008) is organized around five control layers of similar structure, which are differentiated by the increasing abstraction of inputs and outputs. Each layer consists of an adaptive feedback loop that takes state information as input and generates control information as output. The first, lowest layer is a simple perception-action feedback loop that monitors progress toward action goals (reaching a target) and adjusts motor output in light of perceptually generated state information. It is, in this sense, a model of the simplest sort of thermostat; and the idea is that behavioral control systems might consist, at the most basic level, of multiple such control systems – or circuits. Layer 2 takes input from the external world, but also from layer 1, and becomes in essence an adaptive feedback loop monitoring the original layer. That is, layer 2 is in essence a forward model of layer 1. As is well known, incorporating such models into adaptive control systems tightens overall control by allowing for the prediction of state information, so appropriate action can be taken without waiting for the (typically slower) external feedback signal.Footnote ²⁶ . The more hysteresis in the system – the longer it takes control interventions to produce expected results – the more improvement forward models can offer.

Circuit sharing really begins with layer 3, in which the same control circuits described by layers 1 and 2 take as input observations of the actions (or situations) of other agents. Hurley's suggestion is that the mirror system (Decety & Grèzes Reference Decety and Grèzes1999; Gallese et al. Reference Gallese, Fadiga, Fogassi and Rizzolatti1996; Rizzolati et al. 1996) should be modeled this way, as the activation of basic control circuits by state information relevant to the situations of other agents. Layer 3 also implements output inhibition, so agents don't automatically act as if they were in another agent's situation whenever they observe another agent doing something. Layer 4 incorporates monitoring of the output inhibition, supporting a self-other distinction; and layer 5 allows the whole system to be decoupled from actual inputs and outputs, to allow for counter-factual reasoning about possible goals and states and about what actions might follow from those assumptions. The idea is that the same circuits normally used to guide action in light of actual observations can also be fed hypothetical observations to see what actions might result; this can be the basis of predictive models. By the time we achieve the top layer, then, we have the outline for a model both of deliberation about possible actions, and also of multi-agent planning, which could serve as the basis for high-level social awareness and intelligence.

Like the neural exploitation hypothesis, one of the main explanatory targets of the shared circuits model is the possibility of mindreading and intelligent social interaction. And like the neural exploitation hypothesis, it is built entirely on the foundation of sensorimotor circuits. However, unlike the neural exploitation hypothesis, the shared circuits model does not revolve around the inheritance of semantic content from one level to another, but rather around the inheritance of function. The core capacities of the higher layers are based on exploiting the functional properties of the lower layers; all the layers are essentially control loops containing predictive models because they are reusing the basic implementation of the lowest levels. This is an advantage in that it is easier to see how the shared circuits model could be used to explain some of the specific instances of function-driven inheritance canvassed above; for, although Hurley models layer 1 on low-level sensorimotor circuits, there seems no reason in principle that the general approach couldn't allow for other kinds of basic circuits, on which other functional layers could be built.Footnote ²⁷ It is also a potential weakness, in that it is less easy to see how it could be used to account for the central findings of concept empiricism or conceptual metaphor theory; can the sort of functional inheritance allowed by this model also allow for semantic inheritance? The inheritance of a basic feedback structure does not seem to lend itself to any obvious examples of this sort. This is not a criticism of the model as it stands – it was meant only to account for our understanding of instrumental actions; but it suggests that there is no very simple way to generalize the model to a wider set of cases. On the other hand, there seems no fundamental conflict between inheriting a function and thereby inheriting semantic content or domain structure.

I mentioned at the outset that the hierarchy of levels was characterized by an increasing abstraction of input and output. Consider layer 3 in this regard – at this level, input will be both impoverished and abstract as compared with lower layers. It will be impoverished because it will be missing a great deal of the richness of embodied experience – tactile experience, proprioceptive feedback, and efference copy are all absent when observing as opposed to acting. One is left with the visual experience of an action. And note that an action viewed from the first-person perspective looks different from the same action viewed from the third-person perspective. This points to one reason that the information must be abstract: since the visual experience of another agent's action will differ in most, if not all, of its low-level particulars, the system must be sensitive not to these, but to high-level features of the action that are common to the two situations.Footnote ²⁸ Moreover, by hypothesis, layer 3 responds not just to actions, but to situations in which actions are possible – not just to another agent reaching for a banana, but to the banana being within the reach of another agent. This requires imputing possible goals to the observed agent, as well as encoding the high-level features of situations (relations between other agents, their capacities, and the objects in a scene). Here, the shared circuits model may need to be supplemented with something like the feature schemas from the neural exploitation model, itself expanded to allow for situation schemas, and not just object-action ones.

Similarly, if layer 4 is to appropriately and selectively inhibit the control outputs, it must take as input information about the relationships among the actions, agents, goals, and situations – who is in which situation doing what – which requires at least a rudimentary self/other distinction. And if layer 5 is going to be useful at all, the predictions it provides as output must be abstract, high-level action descriptions, not low-level motor commands.

These facts might seem to be nothing more than interesting and functionally useful features of the model, but in fact the requirement for abstraction at higher levels raises a puzzle: If low-level circuits respond to high-level features as inputs, and can produce high-level commands as outputs, might this not imply that layers 1 and 2 are more abstract than the model assumes? The trouble this raises is not with the coherence of the model, but with the evidence for it: All the evidence for layer 1 and 2 type controllers comes from on-line control systems dealing with real-time effector-specific, low-level feedback and control information, and not with abstract, feature-based information.

One obvious way to address this puzzle is to say that each layer is in fact a separate control structure that takes input from and delivers output to the layer below it, but this would undercut the entire premise of the model, since it would no longer be clear in what sense circuits were being “shared.” That high-level control systems are structurally like low-level ones is a perfectly reasonable hypothesis, but this is not the hypothesis put forward by this model, nor is it one for which there is a great deal of biological evidence.

A different approach would be to retain the central hypothesis that control circuits are shared among layers – that layer 3 reuses the control circuit defined by layers 1 and 2, and layer 5 reuses the control circuit defined by layers 1–4 – but suggest that the inputs between layers must be mediated by translators of various kinds. That is, layer 3 takes high-level feature information and translates this into the low-level information favored by layers 1 and 2 before passing it on. Indeed, one might hypothesize it does this by reusing other circuits, such as those that translate abstract plans into successive low-level motor actions. Similarly, layer 5 accepts the low-level motor commands natively output by layer 1, but translates them into high-level action descriptions. This picture is pretty plausible in the case of layer 3 observations of abstract action features, but it is much less clear how situations might get translated appropriately; and it is especially unclear how the reverse inference from low-level motor commands to high-level action descriptions might work. Just as a high-level action might be implemented any number of ways, a specific motor movement might be put in the service of innumerable high-level actions. The fact that part of the sensory information used to retroduct the action/intention from motor movement is the observed effect of the motor movement will help somewhat, but the basic problem still remains: There is a many-to-many relationship between movement and actions, so the valid deduction of a movement from an intention, and the valid retroduction of an intention from a movement need not follow the same paths in opposite directions.

These are hard problems to address; and because they originate from the fact that the shared circuits model requires that different kinds of inputs be fed to the same neural circuits, they may be problems that will surface for any theory of neural reuse (see discussion in section 6.4). Hence, it seems that the best approach to this puzzle may be to bite the bullet and say that, in at least some cases, circuit reuse is arranged such that different data – both information pertaining to different targets, as well as information about the same targets but at different levels of abstraction – can be fed without translation to the same circuits and still produce useful outputs.Footnote ²⁹ Many sorting algorithms can just as easily sort letters as numbers; and if you feed a given algorithm pictures instead, it will do something with them. Naturally, this raises some pressing questions that seem ready-made for an enterprising theorist of neural computation: Under what conditions might useful things be done by circuits working with non-standard data? What kinds of implementations increase the chances of functionally beneficial outcomes given the fact of reuse? We will return to these issues in sections 6.4 and 7.

At its core, the shared circuits model offers an approach to understanding how high-level function could possibly be enabled by low-level circuits – and specifically by the reuse of low-level circuits for various purposes. Unfortunately, it is left fairly unclear exactly how they might actually be so enabled, given the different input-output requirements for each level; I have tried to sketch a solution that does the least damage to the intentions of the model, but I have to admit that some deep puzzles potentially remain.

Nevertheless, the model is interesting as an example of what might come from adopting a fairly classical “boxological” approach to cognitive modeling – understanding information processes via decomposition and interrelation – but without the underlying assumption of anatomical modularity.Footnote ³⁰ If neural reuse is indeed a pervasive feature of the functional organization of the brain – as the current article is arguing – we will need to see more such work in the future.

6.3. The neuronal recycling hypothesis

The neuronal recycling hypothesis (Dehaene Reference Dehaene, Dehaene, Duhamel, Hauser and Rizolatti2005; Dehaene & Cohen Reference Dehaene and Cohen2007) originates from a set of considerations rather different from those motivating the two theories just discussed (i.e., the neural exploitation hypothesis and the shared circuits model). While those are neutral on the question of how and over what timescales the brain organization they propose came about, Dehaene is interested specifically in those cognitive capacities – such as reading and mathematics – that have emerged too recently for evolution to have generated cortical circuits specialized for these purposes. Such cultural practices must be learned, and the brain structures that support them must therefore be assigned and/or shaped during development.

There are two major ways to explain how recent cultural acquisitions, which emerge and are maintained in a population only by learning and not via genetic unfolding, can be supported by neural structures, as of course they must partly be. One way is to take our capacity to acquire such practices as reading and arithmetic as evidence for domain-general learning mechanisms (Barkow et al. Reference Barkow, Cosmides and Tooby1992) and fairly unconstrained neural plasticity (Quartz & Sejnowski Reference Quartz and Sejnowski1997). The other way is to suggest that cultural acquisitions must find a “neuronal niche”– a network of neural structures that already have (most of) the structure necessary to support the novel set of cognitive and physical procedures that characterize the practice. The neuronal recycling hypothesis is of the latter sort.

Note the interesting implication that the space of possible cultural acquisitions is partly constrained by cortical biases. The phrase “neuronal niche” is clearly meant to echo the idea of an ecological niche, and suggests both that acquired cognitive abilities “belong” in specific neural locations (i.e., can only survive where the neural climate is appropriate) and that the neural ecology may partly determine the characteristics that these cultural acquisitions possess, by limiting what is even possible to learn (and therefore which cognitive animals survive). Assuming the set of evolutionarily determined cortical biases is consistent across the species, we should expect to find evidence of at least three things: First, the neural manifestations of acquired abilities should be relatively consistent across individuals and even cultures; second, these practices should have some common cross-cultural characteristics; and third, the same sorts of cortical biases, as well as some of the same possibilities for learning, should be present in nonhuman primates.

As evidence for the first expectation, Dehaene and Cohen (Reference Dehaene and Cohen2007) note that the visual word form area, functionally defined as a region specifically involved in the recognition and processing of written words, appears in the same location in the brain across participants, whether the participants in question are using the same language and writing system or using different ones. Similarly, the intraparietal sulcus has been implicated in numeric tasks, regardless of the culture or number representation system used by the participants. As evidence for the second expectation, they point to work by Changizi and colleagues (Changizi & Shimojo Reference Changizi and Shimojo2005; Changizi et al. Reference Changizi, Zhang, Ye and Shimojo2006) that writing systems are characterized by two cross-cultural invariants: an average of three strokes per written letter; and a consistent frequency distribution for the types of contour intersections among the parts of those letters (T, Y, Z, etc.). Finally, the third expectation has been supported by some interesting and groundbreaking work by Atsushi Iriki and colleagues (Iriki Reference Iriki, Dehaene, Duhamel, Hauser and Rizzolati2005; Iriki & Sakura Reference Iriki and Sakura2008) who uncover evidence for real-time neural niche construction in primate brains (specifically Macaca fuscata) as the result of learning to use simple tools. The location of the observed neuro-morphological changes following tool training is roughly homologous to the regions associated with tool-use in the human brain (Culham & Valyear Reference Culham and Valyear2006). Thus, the theory suggests a novel pathway by which Homo sapiens may have achieved its current high-level cognitive capacities.

The neuronal recycling hypothesis outlines a universal developmental process that, although illustrated with specific examples, is meant to describe the way any acquired ability would come to have a neural instantiation. In this sense, it is broader in conception than the neural exploitation and shared circuits theories described in sections 6.1 and 6.2, respectively (although as noted their scope might well be increased with a few modifications). How much neural plasticity would be required in any given case will vary with the specifics of the acquisition, but one strength of the neuronal recycling theory is it makes clear some of the limits and costs that would be involved. The greater the distance between the function(s) required by a given practice, and the existing cortical biases, the harder the learning process will be, and the more likely that the learning process will disrupt whatever other functions the affected brain regions support.

On the other hand, the more the requirements of the acquisition match what is already possible, the less novel and potentially less valuable the cultural practice is likely to be – unless, that is, it is possible to combine existing capacities in new ways, to use old wheels, springs, and pulleys to form new machines. It is interesting to note in this regard that while the neural exploitation and shared circuits theories outlined earlier tend to envision neural circuits being put to fairly similar new uses – for example, forward models in motor control being used to support forward models in social interaction – the neuronal recycling hypothesis suggests that neural circuits might be put to uses quite other than the ones for which they were originally developed. As already noted above, this notion is central to the massive redeployment hypothesis, which we will briefly review next.

6.4. The massive redeployment hypothesis

Since the massive redeployment hypothesis has already been discussed in section 1.1, I will only review the main idea here. The primary distinction between massive redeployment and neuronal recycling is the time course over which each is supposed to operate. Massive redeployment is a theory about the evolutionary emergence of the functional organization of the brain, whereas neuronal recycling focuses on cognitive abilities for which there has been insufficient time for specialized neural circuits to have evolved. Both, however, suggest that the functional topography of the brain is such that individual circuits are put to various cognitive uses, across different task domains, in a process that is constrained in part by the intrinsic functional capacities (the “workings” or “cortical biases”) of local circuitry.

It is worth noting that the concepts of a “working” and of a “cortical bias” are not identical. The workings envisioned by the massive redeployment hypothesis commit that theory to the existence of cortical biases – that is, limitations on the set of functions it is possible for the circuit to perform in its present configuration. However, Dehaene is not committed to the notion of a local working in virtue of belief in cortical biases. Although it would be natural to understand cortical biases as the result of fixed local workings, a region could in fact perform more than one working and still have a cortical bias. However, the more flexible regions are, the less their individual biases will differ, and the harder it will be to explain the findings that recently evolved cognitive functions use more and more widely scattered neural components. On the other hand, as noted already above, the data are consistent with a number of functionally relevant constraints on local operation. For example, it could be that the dynamic response properties of local circuits are fixed, and that cognitive function is a matter of tying together circuits with the right (relative) dynamic response properties (for a discussion, see Anderson & Silberstein, submitted). In this sense, “cortical bias” is perhaps useful as a more generic term for denoting the functional limitations of neural regions.

In any event, both theories are committed to the notion that putting together the same neural bits in different ways can lead to different – in some cases very different – functional outcomes. In the discussion of the shared circuits model (sect. 6.2), I raised the issue of whether and how a single circuit could be expected to deal with various different kinds of data, as reuse theories seem to require. The question arises here as well: Exactly how is such reuse possible? It must be considered a weakness of both the massive redeployment and the neuronal recycling hypotheses that they lack any semblance of a functional model. In describing my theory (M. L. Anderson Reference Anderson2007a; Reference Anderson2007b; Reference Anderson2007c), I have used the metaphor of component reuse in software engineering, which may be useful as a conceptual heuristic for understanding the proposed functional architecture but cannot be taken as a model for the actual implementation of the envisioned reuse. In software systems, objects are reused by making virtual copies of them at run-time, so that there can be multiple, separately manipulable tokens of each object type. With wetware systems, no such process is possible. What is reused is the actual circuit.

In general, how such reuse is actually effected must be considered an open question for the field. Going forward, supporters of recycling and redeployment need to provide at least three things: specific models of how information could flow between redeployed circuits; particular examples of how different configurations of the same parts can result in different computations; and a more complete discussion of how (and when and whether) multiple uses of the same circuit can be coordinated. Penner-Wilger and Anderson (Reference Penner-Wilger, Anderson, Love, McRae and Sloutsky2008; submitted) have taken some tentative steps in this direction, but much more such work is needed. It is to the credit of both Hurley and Gallese that they each offer a (more or less concrete) proposal in this regard (see Gallese 1996; Reference Gallese2008; Hurley Reference Hurley, Hurley and Chater2005; Reference Hurley2008). That neither seems wholly adequate to the task should not be surprising nor overemphasized; the neurosciences are replete with what must be considered, at best, partial models of the implementation of function by neural structure. More important by far is that neural reuse offers a unique guide to discovery – a sense of what to look for in understanding brain function, and how to put the pieces together into a coherent whole. If neural circuits are put to many different uses, then the focus on explaining cognitive outcomes should shift from determining local circuit activity and single-voxel effects to uncovering the complex and context-dependent web of relations between the circuits that store, manipulate, or otherwise process and produce information and the functional complexes that consume that information, putting it to diverse purposes.Footnote ³¹

One way this effort might be abetted is via the formulation of an even more universal theory of neural reuse than is offered by any of the four theories canvassed above. As should be clear from the discussion, none of the four proposals can explain all the kinds of reuse in evidence: reuse supporting functional inheritance, reuse supporting semantic inheritance, reuse that occurs during development, and reuse that occurs during evolution. In fact, each is strongest in one of these areas, and weaker in the others. This opens the obvious possibility that the four theories could be simply combined into one.Footnote ³² While it is true that there seems no obvious barrier to doing so, in that none of the theories clearly contradicts any of the others, this lack of conflict is in part an artifact of the very under-specification of the theories that leaves them covering distinct parts of the phenomenon. As mentioned already, it may turn out that the kind of functional inheritance required by the shared circuits model precludes the kinds of semantic inheritance required by the neural exploitation hypothesis, or that the schemas envisioned by neural exploitation cannot be modified and expanded along the necessary lines. Likewise, it could turn out that the processes driving massive redeployment are in tension with those driving neuronal recycling; or that one, or the other, but not both can explain semantic and/or functional inheritance.

Should such problems and conflicts arise, no doubt solutions can be found. The point here is simply: We don't yet even know if there will be problems, because no one has yet even tried to find a solution. I would encourage all those interested in the general topic of brain organization to ponder these issues – how does the fact of reuse change our perspective on the organization, evolution, development, and function of the brain? Within what framework should findings in neuroscience ultimately be placed? There is enough work here for many hands over many years.