2.1. Context Sensitivity

Even the most celebrated cases of localized neural function remain controversial. Consider the paradigmatic localization of speech production in humans to Broca’s area (Broadmann areas 44 and 45). Although speech production certainly depends on characteristic activation patterns in Broca’s area, recent research suggests at least four distinct classes of proposed functions that it also, ostensibly, realizes. These proposed functions vary along multiple theoretical dimensions, and there is substantial debate regarding both the extent to which they are compatible with one another and whether or not any of them are correct (Grodzinsky and Santi Reference Grodzinsky and Santi2008). Why should there be such controversy about the legitimacy of even the best cases of localized neural function? The answer, I submit, lies in the observation that many neurofunctional hypotheses make essential reference to transient environmental contexts.Footnote ¹

One might think that making reference to transient environmental contexts is precisely what neurofunctional hypotheses ought to do. Indeed, in section 2.3, I argue that such neurofunctional hypotheses play a crucial role in neuroscientific investigation. One scientific motivation for appealing to functions that essentially depend on reference to transient environmental properties is that such functions are relatively easy to relate to a lower level of description. Often researchers identify a neural mechanism that underlies some functional capacity, a strategy that requires one to descend at least one level (Piccinini and Craver Reference Piccinini and Craver2011). The prospects for success in this endeavor will improve if the range of environmental contexts within which one’s mechanistic hypothesis must be tested is relatively narrow. As the range of contexts expands, the number of processes that might impinge on a hypothesized mechanism will increase, making the task of theory confirmation (or model validation) more difficult. In order to make neurofunctional investigation tractable, therefore, neuroscientists often think in terms of context-sensitive functional hypotheses, in which the number of environmental variables relevant to understanding a neural mechanism is minimized. Jacqueline Sullivan has made a similar claim and supported it with detailed case studies. Her formulation reads: “Given Craver’s own commitment to the idea that mechanisms have specific components and temporal and spatial organizations, and given that such features of mechanisms are sensitive to what set that mechanism into motion, i.e., the events that preceded it, then it is entirely possible, indeed quite likely, that different experimental protocols will yield different mechanisms for potentially different phenomena” (Sullivan Reference Sullivan2009, 527).Footnote ²

Although context-sensitive neurofunctional hypotheses are legitimate and important aspects of neuroscientific investigation, I intend to show that, under some conditions, context sensitivity systematically prevents such hypotheses from explaining neural structure. The argument for that claim is presented in the second section. In the remainder of this section I introduce and define the concept of task-bound functional analysis, a scientific strategy built around the concept of ‘context-sensitive functional hypotheses’. Then I distinguish two logically distinct forms of explanation in which task-bound functional analysis plays a central role. Finally, I highlight one of these strategies and provide an example of how it works.

2.2. Defining TBFA

Task-bound functional hypotheses are a special case of context-sensitive hypotheses. In particular, they are hypotheses about the functional role of a brain area in which the kind of function posited has the property of being task bound, which is defined as follows.

A function F is task bound just in case:

• (i) its description makes essential reference to a behavioral task invoked by an experimental paradigm in which environmental properties P ₁ … P _n are manipulated, and

• (ii) it is ascribed to brain areas that are anatomically individuated.

Each component of this definition deserves clarification. In the first component, variables P ₁ … P _n serve as partial identity criteria for an experimental paradigm. (I say partial because other factors may also be relevant to concerns about individuation, such as the content of the hypothesis under scrutiny.) The values of variables P ₁ … P _n cannot be included in the definition of task boundedness without sacrificing generality. This is because the criteria that determine the paradigm to which a particular experiment belongs are local to the empirical domain in question. Sullivan defines the term experimental paradigm as “a standard method or procedure for producing an effect of a certain type” (Sullivan Reference Sullivan2009, 514). The exact set of environmental properties to be manipulated will depend on the nature of that intended effect. Typically these properties will include the kind of stimuli used and their temporal presentation, the shape and spatial layout of objects in close proximity to the organism, the ambient lighting and acoustic conditions, and many other factors.Footnote ³ The crucial thing to note about the first component of the definition is that, among the identity criteria for a given experimental paradigm will be a set of environmental properties, the manipulation of which is required by the behavioral task invoked by that paradigm.

The second component of the definition rules out analytic functional designations. Consider a claim of the following form: area x, new evidence suggests, does not realize function F₁; in fact, it realizes function F₂. If realizing function F₁ is both necessary and sufficient for being counted as a token of type ‘area x’, then this kind of claim is a matter of definition rather than empirical discovery. I am only interested in the empirical truth of task-bound functional hypotheses, so analytic statements must be set aside. Given the definition of task-bound functional hypotheses, task-bound functional analysis, or TBFA, may be defined as a scientific strategy that attempts to taxonomize and explain the workings of the central nervous system (in any species) via the application of task-bound functional hypotheses.

2.3. The Direction of Explanation

I claim that TBFA is explanatorily limited. But the limitation I intend to identify only manifests itself when TBFA is deployed in the service of potential explanations with a particular logical form.

In neuroscience, as in other fields, functional claims may be invoked either as explanans or explanandum. When invoked as explanandum (probably the more widely discussed strategy in philosophy of biology) the existence of some function is already known, and then a structure is discovered that realizes the function, explaining how one particular system type manages to achieve it. An example of this strategy in neuroscience is the explanation of photoreception in terms of the isomerization of retinal and its subsequent liberation from opsin proteins in the rod and cone cells of the retina. It had long been known that the retina functions to detect and transduce light, but the structural realization of this function remained a mystery until the discovery that only one of two possible conformations of retinal binds to opsin proteins (Wolf Reference Wolf2001).Footnote ⁴ In this case, structural properties of retinal explain how the initial stage of transduction is realized.

Let us refer to explanations of this kind, in which the discovery of structural properties explains how a system type achieves some function, as functionally oriented explanations. TBFA often makes possible the discovery of functionally oriented explanations and, in this capacity, plays a crucial role in neuroscientific practice.Footnote ⁵ Much of the literature on mechanistic explanation illustrates exactly how this role has been, and ought to be, played (see, e.g., Craver Reference Craver2006).

In other kinds of explanation, the logical positions occupied by functional claims and structural claims are reversed, such that the discovery of a function explains why structural properties are present. For example, it was long known that some neural projections are covered with a layer of fatty tissue, known as myelin. But it was not until 1949 that Huxley and Stämpfli gathered the electrophysiological evidence required to demonstrate that sheaths of myelin, interrupted by the nodes of Ranvier, serve to create local circuits that propagate short electrotonic potentials down the length of the axon, making the transmission of current faster than it would be if the only mechanism of propagation were the chain reaction of action potentials, the speed of which is capped by the refractory period of voltage-gated sodium channels (Huxley and Stämpfli Reference Huxley and Stämpfli1949). This discovery explained not only why myelin sheaths are interrupted by the nodes of Ranvier (to create miniature circuits at internode intervals), but also why those sheaths have the material properties they do (in order to block current from escaping). In this example, functional properties explain the presence of structural properties.

Nonscientific explanations of human artifacts often proceed in a similar logical fashion. Pinker (Reference Pinker2009) relates a vignette in which a customer, browsing at an antique shop, stumbles upon a perplexing device composed of a stand, a metal arm bolted perpendicularly to the stand with a horizontally oriented ring on its end, a sharp x-shaped blade capable of pivoting vertically through the ring, and a lever to control the motion of the blade. The customer cannot understand why anyone would assemble such a bizarre collection of parts, until the shopkeeper explains that the device is an old-fashioned olive pitter. The ring holds an olive, and the lever drives the x-shaped blade through it vertically, pushing the pit through the flesh on the bottom of the olive and into a container below. Given that functional insight, the structural properties of the device suddenly make sense. Moreover, the sense of understanding achieved is not merely subjective. A person who understands the function of the olive pitter would almost immediately be in a position to know what sorts of interventions would affect the performance of the device and what sorts would not. Since the facts about whether an olive has been pitted are objective, information about the functional effects of potential interventions on the device constitutes a form of objective knowledge (as opposed to mere belief, or the subjective feeling of understanding that is sometimes elicited by the acquisition of knowledge). Let us refer to explanations of this form, in which the discovery of function explains the presence of structural properties, as structurally oriented explanations.

In this article, I restrict myself to the claim that TBFA faces limitations when used to generate structurally oriented explanations; I prefer to set aside discussion of the potential limitations of TBFA when deployed in the service of functionally oriented explanations.Footnote ⁶ In order to forestall confusion, I should emphasize that both ‘functionally oriented’ and ‘structurally oriented’ explanations are meant to denote forms of explanation that, one way or the other, include functional claims.

2.4. An Example of TBFA

With the above qualification in mind, it will be helpful to consider a well-known example of TBFA that has generated an impressive amount of scientific controversy. The hippocampus is an anatomically defined structure in the medial temporal lobe of mammals and birds. It is widely agreed that the hippocampus serves a memory-related function, but consensus breaks down where more specific hypotheses are requested. As early as 1984, Nestor Schmajuk, a specialist in theoretical hippocampus research, was able to provide a list of 20 distinct theories of hippocampal function, and more have been added since (Schmajuk Reference Schmajuk1984; Morris Reference Morris, Andersen, Amaral, Bliss and O’Keefe2006). Proposals include recording of spatial map coordinates, transitive inference, behavioral inhibition, and consolidating autobiographical memory (Andersen et al. Reference Andersen, Morris, Amaral, Bliss and O’Keefe2006).

Perhaps what is most striking about the list of proposed functions is that it is far from obvious which properties, if any, all items on the list have in common. One reason for the diversity of proposed functions is, no doubt, the diversity of species upon which experiments are conducted: experimental paradigms designed for rats hardly resemble those designed for humans. Nevertheless, on the basis of widespread anatomical and physiological conservation, it is believed that the function of the hippocampus remains constant across species. Indeed, the structural similarities have prompted researchers to claim that a hypothetical surgery, replacing a human hippocampus with that of a macaque, could be entirely successful (Morris Reference Morris, Andersen, Amaral, Bliss and O’Keefe2006). This impressive degree of evolutionary conservation, combined with the fact that the organization of hippocampal substructures is characterized by an intricate form of layering that is unique within the mammalian brain, provides some reason to think that there exists a function, at some level of description, that the hippocampus is expressly designed to perform. It is therefore puzzling, prima facie, that the existing hypotheses of hippocampal function are so wildly divergent. Moreover, it is puzzling that no unifying functional theories have emerged that show how the various task-bound theories are related to one another at a higher level of description.

By stepping back from the details of particular experiments, and focusing instead on shared methodology, it is not difficult to see how the application of TBFA makes the multiplicity of hippocampal theories more or less inevitable. The spatial coordinate mapping hypothesis is an attempt to explain behavioral phenomena observed in Morris water maze paradigms, in which a properly functioning hippocampus is required for remembering the location of a stable platform hidden below the surface of opaque water. The hypothesis that hippocampal function has to do with transitive inference is an attempt to explain behavioral phenomena in an odor-based paired association task, in which associations between bins of food cannot be consolidated over time without an intact hippocampus (Gilbert and Kesner Reference Gilbert and Kesner2003). The hypothesis that hippocampal function has to do with behavioral inhibition is an attempt to explain the fact that rats with lesions to dentate gyrus cells fail to become still in the presence of what is normally perceived as danger (Takahashi Reference Takahashi1995).

The list of task-bound functional hypotheses could be extended, but these few examples suffice to illustrate the feature of TBFA that presently concerns us. It is this: each functional hypothesis in the literature is tailor-made to account for phenomena generated by one family of behavioral experimental paradigms. As the following section demonstrates, this kind of tailoring has the unfortunate consequence of desensitizing functional hypotheses to the data collected in paradigms that differ substantially from the one upon which a particular hypothesis is based.

3.1. Unipotent Structures

The first part of the argument is simple. Consider the strong assumption that each neuroanatomical structure—under a given description—realizes only one genuine function. Bastardizing terminology from developmental biology, I’ll refer to this as the unipotent case.

If a neuroanatomical structure were unipotent, then the diversity of functional hypotheses proposed to explain the presence of structural properties would contradict one another. Since truth does not accommodate contradiction, and explanation requires truth, no explanation will be forthcoming until the contradictions between proposals can be resolved.Footnote ⁷ Moreover, in order to resolve those contradictions under the assumption of unipotency, any given anatomical area would have to be describable by one, uniquely true task-bound functional hypothesis. But no task-bound functional hypothesis is uniquely true. Since task-bound functions refer essentially to the environmental properties manipulated in an experimental paradigm, each task-bound hypothesis describes only a narrow range of the spectrum of functional contributions associated with an anatomical area. Given unipotency, then, task-bound functional analysis always yields contradiction and, for that reason, fails to explain the presence of structural properties.

3.2. Pluripotent Structures

The unipotency assumption is controversial.Footnote ⁸ It will be objected that neuroscientists do not always take incompatible functional hypotheses about the role of an anatomical structure to be logically contradictory. Sometimes neuroscientists circumvent the incompatibility by indexing functional hypotheses to different environmental contexts. In light of this observation, it is necessary to consider the status of TBFA when applied to pluripotent neuroanatomical structures, where pluripotent means that the structure may serve multiple genuine biological functions over time. If a structure is pluripotent, there need not be any contradiction between what might otherwise appear to be competing functional hypotheses. Which particular function is realized at a time will depend on the occurrent behavioral goals of the organism, as well as the particular kind of environment the organism happens to be in. Howard Eichenbaum, a leader in hippocampus research, often expresses this ecumenical view: “Hippocampal sequence representation underlies a range of memory performance capacities, including episodic recall, cognitive mapping, sequence prediction, and inferential memory expression” (Fortin, Agster, and Eichenbaum Reference Fortin, Agster and Eichenbaum2002).

However, the possibility of multiple diachronically realized functions does not suffice to redeem the explanatory status of TBFA. To see this, recall the definition of task-bound functions: a function F is task bound just in case its description essentially refers to a behavioral task in an experimental paradigm in which environmental properties P ₁ … P _n are manipulated. Consider now what happens if the functional activity of the structure in question is investigated within a related experimental paradigm in which the set of manipulated environmental properties are different: P _r … P _s. Since task-bound functions are defined relative to experimental paradigms, a change in paradigm entails a change in the identity of the task-bound function. This has the consequence that new task-bound functions are all too easy to generate. Since there is no principled upper limit on the kinds of behavioral paradigms that might be invented, the list of purported task-bound functions for any given anatomical structure is unbounded: creative variations in the task description yield increasingly many new functions.

The possibility of an arbitrarily long list of functional ascriptions is the source of the explanatory weakness inherent in TBFA. There are two reasons for this. To see the first, assume that functions are defined relative to a containing system, as suggested by Cummins (Reference Cummins1975). A containing system is a system that exhibits a complex capacity, relative to which the functions of its parts are defined.Footnote ⁹ (For example, the containing system for the pumping action of the heart is the circulatory system.) If we hold the containing system fixed, and if we are realists about functional roles, it should always be possible to construct a true (or approximately true) claim specifying the functional profile for a given neuroanatomical area. By functional profile, I mean a complete list of functional capacities. Faced with the specter of an unbounded list, however, such a construction is impossible. Any finite candidate list could be falsified by adding just one more functional ascription.Footnote ¹⁰

Note that TBFA faces this problem even if neuroscientists have no empirical motivation to extend the list of hypotheses beyond its current length. For suppose that TBFA were the only method of discovering structurally oriented explanations. Then there could be no grounds for claiming of any list of task-bound functions that it is complete, even if empirical adequacy had been achieved.Footnote ¹¹ Any functional contribution to behaviors yet to be taken into account would demonstrate the incompleteness of the list.

The second reason that the possibility of an arbitrarily long list of functions entails that TBFA is explanatorily limited is that there is a particular kind of understanding associated with the simplicity of functional insights that TBFA cannot deliver. Consider again the discovery of the function of myelin sheaths. That discovery was powerful in part because the functional insight it represents is extraordinarily compact and yet can account for the presence of at least two otherwise puzzling structural properties. An arbitrarily long list of functions would lack that conceptual compactness, which, although hard to quantify, seems essential to achieving significant explanatory depth.Footnote ¹²

3.3. Objection I: Levels of Abstraction

One might suspect that the possibility of an unbounded list can be avoided by ascending to a more abstract level of description. If we characterize environmental properties abstractly, it is possible to derive a generalized functional hypothesis that stands in relation to specific task-bound functions as determinable stands to determinate. The generalized functional hypothesis will itself be more abstract than any given task-bound function it determines and, therefore, invariant with respect to a larger class of changes in environmental properties. As the number of admissible changes increases, it becomes more difficult and less trivial to generate new functional hypotheses. For example, sequence prediction refers to a broad class of behavioral tasks used in learning and memory studies (Lisman and Redish Reference Lisman and Redish2009). A wide variety of paradigms are counted as tests of sequence prediction processing capacities: a sequence of stimuli can be generated over spatial or temporal metrics; it can be either one or two dimensional; stimuli can be visual, auditory, or olfactory, and so forth. We can define a neural function, F _S, as the function of carrying out sequence prediction processing. One specific instantiation of sequence prediction is a task in which the animal must visually detect nonrandom strings of stimuli buried within longer, random sequences (Dunnett et al. Reference Dunnett, Fuller, Rosser and Brooks2012). By appealing to this more specific task, we can define another neural function, F _R, such that F _R is a determinate of the determinable F _S. Since F _S ranges over a broader class of environmental properties than F _R, it is surely less susceptible to the charge of explanatory weakness. So why not iterate the process of abstraction until a level is reached at which the number of possible functions is suitably limited?Footnote ¹³

Although this iterative process of abstraction is, logically speaking, quite legitimate, it cannot redeem the explanatory status of TBFA. The amount of explanatory strength to be gained by implementing such an abstraction process is severely limited by an additional constraint on structurally oriented explanations: they must be capable of discriminating between distinct neural structures. If a purported explanation were true of multiple brain areas, each of which exemplified a unique form of structural organization, then, whatever else might be explained, it would not be the unique structural properties of one of those areas in particular. The need for this structural discrimination constraint can be clarified by analogy with the lock-and-key mechanisms of the immune system. Consider a purported explanation of the peculiar shape of an antibody that cites the fact that the antibody supports immune system function. Compare it to a purported explanation that cites the following fact instead: the antibody must be shaped as it is in order to ensure compatibility with the surface proteins on the antigen it serves to neutralize. All antibodies promote immune system function. In order to explain the structural properties of the antibody, a functional hypothesis must include specific information about the capacities of that particular molecule. The second of the two aforementioned hypotheses does include such information and, for that reason, is clearly superior to the first. Similarly, all neuroanatomical areas participate in whatever function(s) are served by the nervous system as a whole (i.e., survival), but that information is irrelevant to understanding the structural eccentricities of any particular area.

The explanatory limitations of very abstract functional hypotheses derive from the fact that they will be satisfied by many, and in some cases, all neuroanatomical structures in the brain. To return to our example: when it is claimed that the function of the hippocampus is sequence prediction, we say something which perhaps is true but which fails to provide insight into its unique structural properties. Although originally proposed as a generalized theory of hippocampal function, sequence prediction is so general a concept that some researchers have ascribed a virtually identical function to the brain as a whole. Friston (Reference Friston2010) and other proponents of the predictive coding theory of brain function argue that the most adequate general theory of brain function is the prediction of sequences of perceptual stimuli, given the memory of immediately past sequences. As we have already seen, however, if a functional theory applies so generally, it cannot shed light on the structural properties of the hippocampus in particular.

The lesson of this section is this: if we attempt to construct an abstract functional theory by generalizing over a large class of task-bound functions, the threat of an unbounded list is inevitable. In section 4, I sketch an alternative approach to functional analysis that avoids both the explanatory limitations associated with TBFA and those of hypotheses that generalize over task-bound functions. Before moving on to that discussion, one additional objection should be considered.

3.4. Objection II: Containing Systems Are Chosen by Convention

Another objection to the argument that TBFA cannot explain the presence of structural properties has to do with the distinction between proximal and distal effects. It can be framed as a reductio: the argument is absurd, one might think, because it threatens not only task-bound functions but functional analysis generally. Compare the functional profile of a gene. If it is stipulated that the containing system for some activity of a gene is the cell nucleus, then the function of that gene will be specified in terms of proximal effects, such as the initiation of RNA transcription. If it is stipulated that the entire nervous system is the containing system, the function of the same gene will be specified in terms of distal effects, such as the regulation of neurogenesis. Some philosophers hold that containing systems are stipulated, rather than discovered. On this view, it is always possible to construct an arbitrarily long list of functions by stipulating additional containing systems. If such stipulation is legitimate, then this case seems parallel to the case of TBFA. If the two cases are parallel, then the argument in section 2.2 ought to apply just as well to genetic explanation and, therefore, gives us reason to think that no amount of functional information could provide insight into genetic structure. But if that is the case, the argument proves far too much and should therefore be rejected.

This objection is mistaken for two reasons. First, in the gene case, the generation of new functions is constrained by the biological pathways that exist between gene and container. There may be many, but the set is finite.Footnote ¹⁴ In the case of TBFA, no such biological constraints exist. The number of unique behavioral tasks in which a neuroanatomical structure can be shown to participate is limited only by the creativity of the experimental designer. Second, and more importantly, in order to show that functional profiles generated by TBFA can be made arbitrarily long, it is not necessary to vary the containing system itself, as it is in the gene case. Rather, it is only necessary to vary the set of environmental properties P ₁ … P _n referenced in the definition of the experimental paradigm. So the variability in the TBFA case radically outstrips the variability in the gene case. Since the two cases are not parallel after all, the objection fails.

Arbitrarily long lists of function cannot explain the presence of structural properties. We have seen that TBFA makes impossible the identification of a principled upper limit on the list of functions associated with a given brain area. Any actual list generated by the application of TBFA will therefore have an arbitrary length (whether the magnitude of that length is great or not). Therefore, TBFA is incapable of explaining the presence of structural properties.

In the next section, an alternative concept of functional analysis in neuroscience is discussed, and some speculative remarks are made about the research programs that appeal to it. I argue that the form of functional analysis suggested by this alternative concept is not subject to the explanatory limitations of TBFA and therefore deserves further philosophical investigation.

4.1. A New Concept of Neural Function

If the argument in section 2 is correct, then the explanatory status of certain neuroscientific theories that depend on TBFA is threatened. Rather than trying to identify those theories, I propose a more positive, philosophical project: I propose to provide a rough sketch of a new kind of functional analysis that avoids the explanatory weakness of TBFA.

Given the vast amount of information yet to be learned about neural structure, this synthetic project must remain at least partially speculative. Nevertheless, it can be made systematic if it is framed in terms of a search for desiderata that a new, more explanatory concept of neural function ought to satisfy. The first of these desiderata is this: the new functional concept must denote a property capable of being realized in neural processes that span multiple behavioral tasks. This follows immediately from the argument in section 2. As suggested above, in order to be explanatorily robust, neurofunctional hypotheses must not be susceptible to falsification by the replacement of one set of environmental properties with another. In order to accommodate alternative environmental properties, a function must be realizable in processes that range across tasks.

4.2. Multiple Cognitive Domains

Recent work in functional neuroanatomy suggests two additional desiderata. The first of these is inspired by the work of Michael Anderson, who has argued for a claim similar to the conclusion reached in section 2, albeit on very different, mostly evolutionary grounds (Anderson Reference Anderson2007, Reference Anderson2010).Footnote ¹⁵ Although Anderson does not attempt to define a novel concept of neural function, he does make some suggestive remarks. “Insofar as our approach to discovering the specific functional role of a given brain area involves modeling its activity across different cognitive domains, then it makes little sense to try and characterize the contribution of the area using domain specific terms” (Anderson Reference Anderson2007, 339).

Anderson is suggesting that the concept of neural function appropriate to capturing the functional role of anatomical areas must be one that is realizable in processes that span multiple cognitive domains.Footnote ¹⁶ This desideratum is stronger than the first, which only necessitated that our new concept of function be realizable in processes that span multiple behavioral tasks. To see that the former demand is stronger than the latter, one must only note the very small number of cognitive domains relative to the number of (possible) behavioral tasks. One standard taxonomy of cognitive domains includes just six: attention, skill learning, semantic memory, language, episodic memory, and working memory (Cabeza and Kingstone Reference Cabeza and Kingstone2001). While finer-grained distinctions between cognitive domains are sometimes used, they are not nearly as fine grained as the behavioral tasks invoked by standard experimental paradigms. Consequently, there will be a greater number of conditions under which one brain area is involved in two behavioral tasks than conditions under which that brain area is involved in two cognitive domains. Therefore, the requirement that a functional concept be realizable in processes that span domains is more stringent than the requirement that it be realizable in processes that span behavioral tasks.

The motivation for imposing this condition is a body of empirical evidence from imaging meta-analyses that suggest that brain areas traditionally associated with one cognitive domain in fact participate in processes that span multiple domains (Cabeza and Nyberg Reference Cabeza and Nyberg2000; Wager, Lindquist, and Kaplan Reference Wager, Lindquist and Kaplan2007). Such meta-analyses do not suggest that every brain area is multifaceted in this way, but they do show that multidomain activity is common. If so, it would be useful to have a concept of neural function that can capture the general contribution of an area across cognitive domains. On this point, it is worth quoting Cabeza and Nyberg (Reference Cabeza and Nyberg2000) at length, from a passage in which they interpret the section of their results in which the data are modeled by anatomical region.

As Cabeza and Nyberg suggest, domain-specific functions are useful but explanatorily limited. If we want a more general theory about the functional role of an anatomical area, we require a concept of neural function that is divorced from any essential connection to a particular cognitive domain.

4.3. Intrinsic Function

The third desideratum, which is also the strongest of the three, is this: our new concept of neural function must be capable of eschewing reference to behavior altogether. This desideratum is motivated by a relatively new body of evidence suggesting that much neural activity is designed to achieve ends that are—in a sense that requires explication—intrinsic to the central nervous system. This formulation is inspired by the term ‘intrinsic neural function’, which was recently coined by Marcus Raichle and which he defines as “ongoing neural and metabolic activity not directly associated with the performance of a task” (Raichle Reference Raichle2010, 180). Raichle has marshaled evidence suggesting that at least a large proportion of neural function must be intrinsic in this sense. If he is right, then given the lack of association between this alleged class of neural functioning and the particularities of the behavior with which it is concurrent, we require a concept of neural function that can be articulated without mention of those particularities.

One line of evidence for Raichle’s proposal comes from his own work on the default mode network—a network that is differentially activated whenever an organism is unencumbered by goal-oriented activity. For example, neurometabolic activity in the resting state (in humans) consumes approximately 20% of the body’s total energy budget. Relative to resting state energy consumption, additional activity associated with momentary changes in brain activity is only about 5%, even during arousing perceptual and motor activity.Footnote ¹⁷ On the reasonable assumption that neurometabolic rates are correlated with functional operation, Raichle concludes that much of our normal brain function is driven by internal demands rather than the transient, environmentally modulated demands of whatever task an organism happens to be engaged in. Raichle’s work focuses primarily on whole brain analysis and contains little explicit discussion of the prospects for localization to anatomically defined areas. Nevertheless, Raichle’s claims are based on calculations of glucose consumption, a process that is fundamental to the metabolic capacities of all neurons. Raichle’s findings, therefore, are relevant to any theory of neural function, whether it concerns activities that are localized to small anatomical structures or activities that are distributed across them.

If, as this evidence suggests, the majority of localized neural functions are intrinsic, then we require a concept of neural function that is compatible with intrinsic functionality. I therefore suggest that the third desideratum be formulated as follows: the new concept must denote properties that are realizable in intrinsic processes.

It is important to note that, despite a superficial similarity, this condition differs from the first—the one imposed by the result obtained in section 2. That argument only shows that if we want structurally oriented explanations, the neural functions we appeal to must be realizable in processes that range across behavioral tasks. In order to capture the contributions of intrinsic neural activity, the concept of neural function must be capable of eschewing reference to behavior altogether, which is a considerably stronger condition.

It should also be noted that the second and third desiderata constrain the concept of neural function in substantially different ways. Consider the (dubious) claim that the amygdala is a fear-processing center. Here no reference to behavior is made, but the hypothesis still lacks explanatory power to the extent that the amygdala participates in processes that span other domains.Footnote ¹⁸ Because this case satisfies the second, but not the third desideratum, it serves to demonstrate the substantive difference between them.

4.4. Containing Systems inside the Brain

Unfortunately, our three desiderata are difficult to satisfy without immediately running up against the limitation on abstraction discussed in section 2.3. In order to provide nontrivial explanations of structural properties (why are the inputs to the hippocampal formation mossy fibers, rather than pyramidal cells?), functional hypotheses must supply information that is both detailed and free from reference to particular tasks. The challenge posed by this requirement can be reconceptualized as a trade-off between the explanatory limitations of task boundedness, on the one hand, and the explanatory limitations of ungrounded abstraction, on the other.

Fortunately, it may be possible to avoid the trade-off, at least in part, by reconsidering the kinds of containing system that are compatible with our desiderata. Since reference to behavior has been ruled out by the third desideratum, any containing system to which the new concept of neural function refers must be internal to the organism. In particular, the new concept of function will be defined with respect to containing systems inside the brain.Footnote ¹⁹

This restriction forces us to think of neural function in terms of the proximal effects a brain area has on the circuit of which it forms a part. Since cognitive and behavioral vocabularies have been ruled out, we are forced to resort to a more abstract, mathematical level of description to capture those proximal effects. Rather than framing functional hypotheses about an anatomical area in terms of its contribution to behavior, or its capacity to realize a cognitive process, the new concept of neural function allows us to frame functional hypotheses in terms of signal transformation. Each anatomical hub in a large-scale brain circuit exhibits a unique set of structural properties that determines the kinds of transformation it is capable of carrying out. As a signal flows through the circuit, it will be transformed in a particular way at each anatomical hub. In principle, that transformation can be modeled mathematically. If so, functional ascriptions take on a logical form that is very different from those underlying most functional ascriptions in biology. Consider the following schema: the function of brain area A is to transform signal S in manner Ψ. If transformation Ψ can be modeled by a mathematical function, then the functional role identified by this schema provides information that is both detailed and specific. If so, the trade-off between task boundedness and abstraction has been circumvented.

The two-variable formulation of the schema above is designed to make possible a relatively fine-grained form of functional analysis. Whether or not a brain region instantiates Ψ will be determined by local microanatomy. But the individuation conditions on S may incorporate information about the sensory modality from which it originates, as well as information about the site at which the transformed signal is to be consumed. So even in cases in which two brain areas are indistinguishable with respect to local anatomy, they can nevertheless realize distinct functional roles.Footnote ²⁰

In the past decade, multiple new lines of research have adopted a signal transformation based view of functional analysis. What they have in common is their foundation in quantitative analysis of spiking neuron models. For example, Levy, Hocking, and Wu (Reference Levy, Hocking and Wu2005) offer a theory of hippocampal function that overcomes the limits of TBFA by appealing to the purely intrinsic concept they call the “associator of last resort.” Briefly, the idea is that the hippocampus is designed to form associations between input patterns that are, statistically, so dissimilar from one another that other anatomical regions cannot “find” the associations.

Its ability to do this depends on the structural properties that define the network used to model the brain area, such as percentage connectivity and the degree of feedback inhibition from interneurons. These anatomical properties determine which functions (in the mathematical sense) a region is able to approximate.Footnote ²¹ In this functional hypothesis, the capacity that explains the presence of particular structural properties in the brain region is specified in terms of intrinsic statistical properties that remain constant over time and across environmental changes. Moreover, Levy’s “associator of last resort” concept makes no reference to experimental tasks, particular cognitive domains, or particular behaviors and thereby satisfies the three desiderata we have imposed on an explanatory concept of neural function.

I hope to have made two overarching points thus far. The first was that TBFA fails to explain the presence of structural properties. The second point was that the new concept of neural function suggests ways of overcoming that limitation. The philosophical upshot of these two points is simply that, when it comes to the brain, functional analysis will have to be carried out in abstract terms that are very different from those discussed in philosophy of biology. Localization of neural function is possible, but we will find the kinds of functions that can be localized quite surprising. Successful functional analysis will become integrated with what is currently known as “theoretical neuroscience” and will most likely draw on fields such as information theory, statistics, genetics, and perhaps physics, rather than directly from experimental paradigms in wet laboratories. Moreover, functions will be defined with respect to container systems within the brain itself.

These points return us to the claim made at the outset—that no brain area has a behavioral task as its principal function. Since task-bound functions cannot explain structural properties, they also cannot identify the unique contribution an anatomical area makes to the capacities of the brain as a whole. Since task-bound functions clearly fail to approximate that unique contribution, they ought not be used to label anatomical areas: the selection of any one label would be arbitrary. If we want our neurofunctional designations to be principled, they will have to satisfy the desiderata outlined above.