2.1. The shared conception

While, as stated above, there is no consensus theory of content philosophers have subscribed to, we can point to a broad approach that is widely shared. Following Morgan and Piccinini, we can define “the shared conception” of existing theories of content:

In short, we define “the shared conception” as:

Of course, for the shared conception (*) to become an actual theory of content for neural representations, one would have to properly explicate how the notion of function and the notion of tracking are best understood, and this is where different theories diverge. While the broadly defined “shared conception” is not committed to any specific definition of these notions, it still carries some basic understanding of both. As for tracking—it is generally assumed to be some causal-informational relation that (on its own) does not allow for mistakes. Following Neander (Reference Neander2017a, 83) we can consider this as a commitment to some “natural-factive information relation,” such as Grice’s (Reference Grice1957) “natural meaning.” And as for the notion of function—on the shared conception it must enable precisely this talk of normativity and misrepresentation. We can again follow Neander and consider the shared conception as committed to a notion of “normal-proper function”:

But these broad, or ostensive, descriptions of tracking and function still have to be unpacked by actual theories of content. And if we take a look at how existing theories interpret the relevant notion of function, there seems to be a common thread among them—they all define neural representations by appealing to some historical notion of function. On the most popular approach, the function of a representation, or a representational vehicle, is defined by what it was selected for through some evolutionary process of selection (e.g., Millikan Reference Millikan1984, Reference Millikan1989, Reference Millikan2004; Neander Reference Neander1995, Reference Neander2017a). Others have also appealed to a natural process of learning (e.g., Dretske Reference Dretske1988, Reference Dretske1995) or to selection in virtue of “contribution to an organism’s persistence” (Shea Reference Shea2018). While these theories might differ in pointing out the relevant type of historical process, they still all define representations through a historical notion of function. The aim of this paper is to promote an appeal to a different notion of function—that of Cummins’s (Reference Cummins1975) account.Footnote ³

2.2. Cummins functions

Debates over what is the relevant notion of function are hardly limited to the issue of representations. There is a wider philosophical debate about understanding the relevant notion of function for scientific explanations in general, and physiological explanations in particular. And within that debate Cummins functions play a significant role. In what follows, Cummins defines when it is the case that a certain internal component x (e.g., the heart), within a containing system s (e.g., the circulatory system), has the function ϕ (e.g., to pump blood). Note that the definition is relative to an analytical account A (i.e., an explanation) of some capacity $\psi $ that is performed by the containing system.

Cummins functions.

On this view x has the function ϕ, not because it evolved to have that function or it was selected for that function in some other manner. Rather x has the function ϕ because the fact that x is capable of ϕ allows us to explain how the containing system achieves a more general capacity ( $\psi )$ . Importantly, this means that functions are defined relative to an explanation: “To ascribe a function to something is to ascribe a capacity to it which is singled out by its role in an analysis of some capacity of a containing system” (Cummins Reference Cummins1975, 765).

We should note that Cummins functions are sometimes referred to as role functions, causal-role functions, system functions, or systematic functions. The latter is also how Cummins himself refers to such functions. There are also some variations on how we should define such functions, but we will focus mainly on Cummins’s classic (Reference Cummins1975) account. Accordingly, we will continue to refer to these as Cummins functions. And with regards to the general debate over how to understand functional attributions in the physiological sciences, Cummins functions are often regarded as a viable candidate.Footnote ⁴

Importantly, though, this paper does not take a stance with regards to the interpretation of functional attributions in any general sense. I only wish to claim that Cummins functions can figure into an account of neural representations. Clearly, that would carry some commitment to Cummins’s account of function, but that does not necessarily mean that all functional attributions in neuroscience should be understood in this manner, and it definitely does not mean that functional attributions in the biological or physiological sciences should always be understood as Cummins functions. In this paper, we restrict our commitment to Cummins functions only to those functions that are relevant to defining representations. And I believe that, in this context of neural representations, Cummins functions have been grossly overlooked. Let us now see how an appeal to Cummins functions can feature in a theory of content for neural representations.

3.1. Definition

The Cummins functions approach (CFA) simply combines the shared conception (*) with Cummins’s (Reference Cummins1975) account of function. This will define a notion of internal representation that is dependent on explanatory context, just as Cummins functions are. But since we are currently only interested in the neural representations that are relevant to neuroscientific explanations, we will restrict our definition accordingly. While Cummins’s (Reference Cummins1975) account can define the function of any internal component, relative to some explanation of its containing system, CFA’s definition of neural representations will focus only on neural internal components, relative to some neuroscientific explanation of its containing system. Other than that, CFA directly follows Cummins’s definition of function quoted above. There, Cummins explained when some capacity ϕ of an internal component x can be regarded as its function. Now, if we look only at those cases where ϕ is a capacity of tracking, and assume (*) that representations are defined by a function of tracking, then Cummins’s definition of function also entails a notion of representation:

The Cummins functions approach (CFA) to neural representations .

Suppose that x is an internal neural component of s, that s has some capacity to $\psi $ , and that A is a neuroscientific explanation of s’s capacity to $\psi $ . Then,

x is a neural representation of X, relative to A, just in case:

1. 1. x is capable of tracking X, and

2. 2. A appropriately and adequately accounts for s’s capacity to $\psi $ by, in part, appealing to the capacity of x to track X.

There are a couple of points worth stressing here, before moving on. First, CFA does not amount to a well-defined notion of neural representation, since we have not offered a definition of tracking. This is why we consider this as an approach to neural representations, and not as an actual theory of content. This approach generalizes over various possible definitions of tracking, and as such it defines a family of possible theories of content.

Second, on this approach neural representations are only defined relative to a given explanation. And following Cummins (Reference Cummins1975, 762), we demand that the explanation A “appropriately and adequately accounts for s’s capacity to $\psi $ .” Yet, we will refrain from defining what makes an explanation “appropriate and adequate.” CFA is not committed to, and not dependent on, any account of explanation in general, or neuroscientific explanations in particular. Given a neuroscientific explanation, CFA will enable our understanding of the neural representations that are considered within it, under the assumption that this explanation “appropriately and adequately explains s’s capacity to $\psi $ .” Furthermore, CFA is also not committed to Cummins’s view of scientific explanations as functional analyses—whereby a complex capacity is explained by appealing to the systematic organization of simpler capacities. We will assume that the Cummins function of a particular component is defined by its contribution to the capacity of a containing system, relative to some explanation A, but we need not assume that the explanation A is itself a functional analysis.Footnote ⁵

3.2. Example

Let us now illustrate how CFA can account for the notion of representation found in neuroscientific explanations. We will consider the vestibulo-ocular reflex (VOR) which is the phenomenon of maintaining a steady eye gaze while the head moves. An explanation of this phenomenon, adapted from Robinson (Reference Robinson1989),Footnote ⁶ is very briefly described in Figure 1.

Figure 1. The explanation of VOR. On the right the canals transduce head velocity, $\dot H$ , and report it, coded as the modulation of the discharge rate, ${R_{V1}}$ , of primary vestibular afferents to the vestibular nucleus, vn. This signal becomes an eye velocity command for vestibular movements, ${\dot E_v}$ , which is sent directly to the motoneurons, mn, and to the neural integrator, NI, to provide the needed position signal E. These signals provide those needed by the motoneurons modulating by ${\rm{\Delta }}{R_m}$ (Robinson Reference Robinson1989, 35).

For the purposes of this paper, we need not delve into the details of this explanation of VOR. Suffice to say that it describes the mechanism underlying VOR as an information-processing mechanism, whereby internal neural components are regarded as carrying content or information. But still, if we wish to claim that these internal components truly are neural representations, then we need an account of what being a neural representation amounts to. For example, this explanation clearly regards the primary vestibular afferents as carriers of a head velocity signal, coded by their discharge rate ( ${R_{V1}}$ ). But in order to claim that these neurons are a neural representation of head velocity, we need to show that they meet some clearly defined criteria for being a representation of head velocity, and that is what theories of content aim to achieve.

CFA can account for the representational status of the primary vestibular afferents, though we will need to say something more about tracking. As defined above, CFA is uncommitted to any definition of tracking, but some assumption will be necessary in order to illustrate this approach on any specific example. In this case, we will assume that the fact that the discharge rate of the primary vestibular afferents correlates with head velocity (as described in Figure 1) can be taken as meaning that the primary vestibular afferents track head velocity. This is a nontrivial assumption but also an uncontroversial one. It is consistent with existing possible definitions of tracking, and I believe most philosophers would accept it. For us, the interesting question is not whether the vestibular afferents track head velocity, but whether it is their function to do so.

As stated, existing theories of content have dealt with such questions through some historical notion of function, such as claiming that the primary vestibular afferents have the function of tracking head velocity because they were naturally selected for this capacity in the evolutionary process. But now, let us fit this example to the definition of CFA given above. We have an internal neural component x, the primary vestibular afferents, of some system s, the brain^{Footnote 7}, where s has a certain capacity, $\psi $ , to maintain steady eye gaze while the head moves. We also have a neuroscientific explanation A of how the brain maintains its gaze, i.e. an analytical account of s’s capacity to $\psi $ . A is the explanation of VOR found in Robinson (Reference Robinson1989), and described in figure 1. Thus, according to CFA, we can state that:

The primary vestibular afferents (x) are a neural representation of head velocity (X), relative to A, just in case:

1. 1. The primary vestibular afferents are capable of tracking head velocity, and

2. 2. A appropriately and adequately accounts for the brain’s capacity to maintain steady eye gaze while the head moves (VOR) by, in part, appealing to the capacity of the primary vestibular afferents to track head velocity.

We have already assumed that condition 1 holds—that the primary vestibular afferents are capable of tracking head velocity. We also assumed that this tracking relation is identified by the neurons’ discharge rate correlating with head velocity. Therefore, I believe it is quite clear that condition 2 holds as well. The fact that the vestibular afferents track head velocity seems to be the only capacity of theirs that the explanation of VOR appeals to. Of course, their basic ability to transfer a signal via neural excitation or inhibition must also be relevant to the explanation, but that would clearly be true for all components of the mechanism. And it is not this ability that accounts for the role of this specific component—the primary vestibular afferents—in enabling VOR. Tracking head velocity is what accounts for that. Tracking head velocity is the capacity of the primary vestibular afferents which, in part, allows us to explain the functioning of the containing mechanism and understand the phenomenon of VOR. Thus condition 2 would hold, and we can conclude that, according to CFA, the primary vestibular afferents are a neural representation of head velocity, relative to the explanation of VOR.

I contend that CFA offers the right path towards understanding the notion of representation that is at use throughout current neuroscientific practice. Let us now scrutinize this claim through some possible challenges.

6.1. Compatibility with neuroscientific practice

To start, the previous section raised the possibility that CFA will define the same neural component x as being a representation of ${X_1}$ in one context, and a representation of some different ${X_2}$ in another context. Perhaps some would consider this a problem, or even claim it means CFA offers an indeterminate notion of representation. This latter claim is simply false, since CFA will only define representations relative to a given explanatory context. Thus, the type of possibility we are discussing is not a case of an indeterminate representation. It is not a case where CFA defines a representation with two contents ${X_1}$ and ${X_2}$ . Instead, this is a case where CFA defines the neural component x as a determinate representation of ${X_1}$ in one context, while also defining the same neural component x as a determinate representation of ${X_2}$ in a different context. And I believe that this is precisely the type of outcome we should want.

If CFA allows for a single neural component x to be a representation of two different contents (in two different contexts), then that can only occur in the following scenario. First, there is a single neural component x that has both the capacity to track ${X_1}$ and the capacity to track ${X_2}$ . Second, there are two neuroscientific explanations ${A_1}$ and ${A_2}$ , whereby ${A_1}$ explains some brain capacity ${\psi _1}$ by appealing to x’s capacity to track ${X_1}$ , and ${A_2}$ explains some brain capacity ${\psi _2}$ by appealing to x’s capacity to track ${X_2}$ . Finally, we assume that both ${A_1}$ and ${A_2}$ appropriately and adequately account for the capacities ${\psi _1}$ and ${\psi _2}$ , respectively. Only under these conditions would CFA define x as having two different contents, relative to the different contexts. And again, I believe that under these conditions, we should want a theory that allows us to state that x is a representation of ${X_1}$ in one context, and a representation of ${X_2}$ in the other. This is the simplest, most straightforward option. And it also seems consistent with contemporary neuroscience, where scientists often regard the same neural component as carrying different kinds of information, or “representing” different things.Footnote ¹⁰

In general, any fear that the context sensitivity of CFA will allow us to define neural representations that are in some sense “unwanted” seems to be entirely unfounded. We have already stressed that CFA’s dependence on explanatory context does not allow for some rampant subjectivity where scientists do as they please. If CFA defines x as a neural representation of X, relative to some explanation A, that means that there is some successful (or “appropriate and adequate”) neuroscientific explanation A according to which some phenomenon is achieved, in part, because x has the capacity to track X. Why, then, would we not want to consider x as a neural representation of X? Without assuming some alternative definition of neural representations, what reasoning could justify the claim that “the fact that CFA identifies x as a representation of X is bad or counterintuitive?” It is normally precisely these cases— where a successful neuroscientific explanation appeals to an internal component that tracks some distal entity—which representationalists regard as evidence for the explanatory significance of representations. Thus, we should want a theory that defines this internal component as a representation, and that is what existing theories of content have always attempted to do.

Still, some might insist that CFA’s dependence on explanatory context will necessarily mean that it is incompatible with the actual neuroscientific practice. That is because defining representations relative to explanations means defining them in a manner that is at least partially subjective and non-naturalistic. Meanwhile, neuroscientists are generally in the business of describing real and objective phenomena, and studies such as Bechtel (Reference Bechtel2016) and Thomson and Piccinini (Reference Thomson and Piccinini2018) have shown that this also applies to representations.Footnote ¹¹ Now, before I reply, it is worth noting that others have drawn very different conclusions from investigations of neuroscientific explanations. Egan (Reference Egan2014), for example, concludes that content ascriptions must be pragmatic.Footnote ¹² But I will not be making such a claim here. Nor do I intend to prove here that CFA is the correct view of neural representations, which offers the best possible account for representations in neuroscience. As I mention at the conclusion of this paper, that still remains to be seen. But in order to establish CFA as a viable approach to neural representations, I do wish to show that CFA certainly can be consistent with the naturalistic, objective practices of neuroscientists. To that end, let us highlight some aspects of this view that are usually associated with “realistic” approaches to representations.Footnote ¹³

For starters, according to CFA, a tracking relation will only define a neural representation if it is essential to a neuroscientific explanation. Hence, neural representations are essential components of neuroscientific explanations. That is, CFA is not a deflationary view of representations. Additionally, I would also claim that CFA is consistent with vehicle realism. Shea (Reference Shea2018, 15) states: “I will reserve the term ‘realism’ for accounts that are committed to there being real vehicles of content: individuable physical particulars that bear contents and whose causal interactions explain behaviour.” CFA, I would argue, is consistent with this demand. The primary vestibular afferents, for example, are individuable physical particulars that bear contents and whose causal interactions explain behavior, hence real vehicles of content. Now, it is possible some would want to insist that “vehicle realism” should also imply that the vehicle has its contents intrinsically. I think that would be a mistake. But frankly, whether you want to call it “vehicle realism” or not, the important point is that CFA regards real individuable physical particulars as carriers of content.

And it’s more than that. In defining a representation, CFA also demands that this real physical particular “track” some property X (its content). While we have not committed to any definition of “tracking,” CFA is certainly consistent with the (very likely) possibility that this is an objective relation in the world. In that case, neural representations, according to CFA, are dependent on scientists finding an objective “tracking” relation between a real physical particular and some property X. Furthermore, according to CFA, this tracking relation will only define a representation if it plays an essential role in enabling a more complex phenomenon. And it is again highly likely that scientists show this by pointing at an objectively real causal structure, such as the physical mechanism that enables VOR. All in all, we find that CFA is consistent with the possibility that neural representations are real physical particulars that exhibit an objective “tracking” relation, within an objectively real causal structure. All of which, of course, must be discovered empirically by neuroscientists.

Still, as discussed in section 5, CFA also maintains the necessity of the explanatory context in picking out these specific (objective) phenomena. But I do hope that the discussion above helps clarify why this type of explanatory dependence does not necessarily clash with the objective practices of neuroscientists. Again, we will still need to check and see which available theory best conforms to the scientific practice. But it would be a mistake to dismiss CFA just because we have stated that it is dependent on explanatory context. I think there is reason to believe that CFA can capture the notion of representation that is relevant to neuroscience, and we have seen some evidence for the advantages of this context-dependent view. But we have yet to touch upon the fact that CFA violates the naturalistic constraint.

6.2. Against the naturalistic constraint

The naturalistic constraint is the demand that a theory of content be specifiable in non-semantic, non-intentional terms. Thus, CFA violates the naturalistic constraint by defining representations relative to a given explanation and with respect to scientists’ explanatory aims. And since the project of accounting for neural representations is often considered synonymous with the project of finding a naturalistic theory of content, the possibility of violating the naturalistic constraint might seem atrocious to some. Obviously, I disagree.

Ultimately, the basic reasoning behind the naturalistic constraint is that if we are to have a theory that explains content, then it can’t itself be dependent on content (or semantics, or intentionality). But importantly, CFA makes no claims towards explaining content in any general sense. It only aims to account for the notion of representation we find in current neuroscientific practice. And I suppose that in this regard CFA may have strayed from the norm. Traditionally, theories of content have been considered in significantly different contexts. Through theories such as the Representational Theory of Mind (RTM), or the Representational Theory of Intentionality (RTI), philosophers of mind have attempted to account for the nature of thought and mental intentionality by positing the existence of internal mental representations. And if mental representations are used to explain intentionality, then intentionality can’t be used to explain mental representations. Thus, theories of content for such mental representations must be naturalistic.

Even theories of content which are specifically aimed at accounting only for subpersonal/nonconceptual/nonconscious representations (Shea Reference Shea2018; Neander Reference Neander2017a), are still normally regarded as ultimately serving some grander naturalization project in the philosophy of mind. As Shea puts it:

Neander (Reference Neander2017a) offers the same strategy. Her theory accounts solely for subpersonal, nonconceptual representations, but she also makes it abundantly clear that her ultimate goal is to account for mental intentionality in general.Footnote ¹⁴ Neander even states that it this ultimate goal—accounting for mental intentionality—that should lead us to exclude Cummins functions from theories of content: “It won’t do, for example, to claim that the relevant functions are ontologically grounded in the explanatory aims of researchers (as Cummins does) and then explain intentional mental phenomena, such as the explanatory aims of researchers, as grounded in such functions. That would be circular” (Neander Reference Neander2017a, 86).

But unlike Neander (Reference Neander2017a) and Shea (Reference Shea2018), CFA is truly not concerned, either directly or indirectly, with the attempts to “explain intentional mental phenomena, such as the explanatory aims of researchers.” CFA is not in any way intended to serve the naturalization project of thought, intentionality, or content. Not that I oppose such projects, but that’s simply not the goal of CFA. CFA is offered with the sole aim of accounting for neural representations, which we defined in section 1 as the representations that are posited in contemporary neuroscience. We only care about successfully characterizing the notion of representation that is relevant to current neuroscientific practice. Why, then, must our theory of content adhere to the naturalistic constraint?

There is a common assumption that an account of neural representations has to also (somehow) serve a more general account of intentionality. As Sprevak states: “it is widely assumed that neural representations are more fundamental than, and ground, other representations. Neural representations ground, and are somehow responsible for, personal-level thoughts such as beliefs, desires, and intentions. Personal-level representations in turn ground conventional representations such as signs, maps, and public language” (Reference Sprevak2013, 552). And this type of assumption can justify the naturalistic constraint. We cannot appeal to neural representations to ground personal level intentions if we appeal to the intentions of scientists to ground neural representations.

But I reject this line of thought. To be precise, I reject the assumption that neural representations must be able to play this type of “grounding” role. Again, we identified neural representations as the representations that are posited in contemporary neuroscience. Now, I have no intention of opposing RTM, or the idea that subpersonal representations can somehow ground personal level thoughts. But to assume that the representations that are posited in contemporary neuroscience are these subpersonal representations that will ultimately serve a naturalistic account of intentionality is a hopeful hypothesis at best. And a hopeful hypothesis cannot define a necessary constraint.

There might be good reasons for wanting neural representations to enable a naturalistic account of intentionality, but that doesn’t mean there are good reasons to assume that has to be the case. We cannot preemptively constrain a theory of neural representations to be naturalistic, just because we want it to be the case that the same representations that neuroscientists appeal to will also somehow ground the intentionality of thought. If we are looking for an account of neural representations, then that means we must look for the theory that best accounts for the notion of representation that is relevant to neuroscience. And whether or not a theory successfully accounts for representations in neuroscience is not dependent on whether or not this theory can also help us account for the intentionality of thought. Once the task of accounting for neural representations is adequately detached from “the problem of intentionality,” the naturalistic constraint loses its justification.

To be clear, it’s only the constraint that loses its justification. It’s only the claim that a theory of content for neural representations must necessarily be naturalistic that I reject here. I am not arguing against naturalistic theories of content. I am just saying that, when it comes to theories of neural representations, being naturalistic is not a necessary constraint. It could still be viewed as an advantage. I certainly agree, for that matter, that it would be nice to have a theory of neural representations that can also ground mental intentionality and content in general. And obviously CFA, at least as it is described in this paper, does not seem to achieve that. It is worth noting, though, that neither does any other existing account. No one has been able to actually make good on this promise to explain mental intentionality by grounding it in a naturalistic theory of content. As Sprevak notes: “Unfortunately, and despite a large investment of effort, an adequate theory of natural representation has not been forthcoming. Many contemporary philosophers suspect that representation simply cannot be naturalized” (Reference Sprevak2013, 547). So perhaps trying a different approach is not such a bad idea. Sprevak (Reference Sprevak2013) and Egan (Reference Egan, Margolis, Samuels and Stich2012, Reference Egan2014), for example, have also offered alternative theories of content for neural representations that do not accept the naturalistic constraint as defined above.^{Footnote 15}

Much more importantly though, and the point I have been trying to make throughout this section, is that it is wrong to let external considerations restrict our account of neural representations. The only question we should care about is whether or not CFA best accounts for the notion of representation that is relevant to contemporary neuroscientific explanations. And if it turns out that it does, we must conclude that neural representations really are what CFA says they are. The implications CFA might somehow carry for the naturalization program in the philosophy of mind, whether we like them or not, change nothing about that.