Deflating Mental Representations

I found this book illuminating. A font of fresh ideas, it challenges the currently dominant Representational Theory of Mind and develops an alternative account of mental states and processes that is worthy of close attention. I was not in the end persuaded by its main claims, but I came away feeling that I had learned a lot in thinking about those claims and evaluating them. The book also offers some ancillary benefits, such as an outstandingly clear account of the nature and scope of computational cognitive science.

The principal claim of the Representational Theory of Mind can be expressed as follows:

RTM: A wide range of mental items (including every propositional attitude, every concept, and every perceptual state) constitutively involve mental representations that are directed on, or represent, objects, properties, events, or states of affairs.

Adherents hold that RTM captures a key architectural assumption of folk psychology and contemporary cognitive science, and they advocate a robustly realistic reading of it, according to which the mind really is populated by mental representations, and representations really are the objects over which the cognitive processes postulated by folk psychology and cognitive science are defined. Contemporary versions of RTM allow that mental representations can take many forms—sentences, images, simulations, maps, place cells, and so on—but they agree that representations of all these kinds bear forms of the generic relation X represents Y to portions of extra-mental reality. Most advocates seek to naturalize this relation by explaining it reductively—that is, without using any semantic, intentional, or normative concepts. It is acknowledged on all sides that a fully satisfactory reduction has yet to be found, but many advocates think that a successful reduction is gradually emerging in the current literature (e. g., Neander, 2017; Shea, 2018).

Egan’s main objection to this picture is that attempts to explain the relation X represents Y are never naturalistic in the advertised sense. They aim to honor the intuition that each representational vehicle has a unique content, but there are always multiple candidates for the role of content, and selection among them is determined pragmatically—that is, by the explanatory interests of the relevant theorist—rather than naturalistically. Consider, for example, Neander’s attempt to show that a certain representational vehicle V in the toad’s optic tectum represents the complex of low-level properties moving worm-shaped object rather than a higher-level ecological property such as nutrient (Neander, 2017). In making her case, Neander points out that attributing the lower-level content to V better serves the purpose of describing the causal process that supports the toad’s feeding behavior. After all, experiments show that a toad will gobble up any moving worm-shaped object it perceives, regardless of whether the object is nutritious.

It follows that the properties moving and worm-shaped are triggers of the process that eventuates inV, and that nutrient makes no proprietary contribution to the process, but is merely co-instantiated with the properties that actually do the causal work. As Egan points out, however, a biologist whose dominant theoretical interest is in explaining the toad’s survival and flourishing in its ecological niche would have an equally strong reason to say that the vehicle V represents nutrient or a niche-equivalent property such as toad food. On what basis could a representationalist philosopher or scientist decide between these two competing hypotheses about the content of V? Egan’s answer appears to be that there are no naturalistic facts that provide such a basis. Rather, preference for one hypothesis over the other can only be justified by context-dependent pragmatic considerations, and in particular, by the explanatory goals of the person making the choice. Someone interested in explaining how the toad’s perceptual system works would prefer the moving worm-shaped object hypothesis, while someone interested in explaining how the toad manages to survive would prefer the nutrient hypothesis. This being so, she concludes, neither claim “will naturalize content in the advertised sense, inasmuch as the content attributed has a pragmatic rationale” (28).

According to Egan, then, representata are not determined naturalistically. Prima facie, this thesis calls the naturalistic credentials of the whole representationalist enterprise into question. Intuitively, however, representationalist explanations of behavior can be extremely illuminating and may even be conceptually necessary, given that naturalistic explanations in terms of neural states and processes are likely to be forbiddingly complex. Can Egan’s anti-naturalist conclusion be reconciled somehow with this observation about the value of representationalist explanations? She thinks it can. Her account of how this is possible begins with a distinction between a scientific theory proper and a gloss that scientists add to the theory.

To illustrate, in the case of the prey-catching behavior of the toad, a theory proper would consist of a naturalistic specification of a physical process involving a mechanism M in the toad’s brain that takes visually acquired information about a distal stimulus S as input and yields a protrusion of the toad’s tongue as output, where S is the target of the protruded tongue. A representationalist gloss on M would designate a state R of M as a mental representation and would add that R represents the external stimulus S. The gloss would therefore make contact with key components of the causal process running from S to the protruded tongue, and could therefore be said to track or model an illuminating physical explanation of the toad’s prey-capturing behavior.

But the gloss would not be unique. One possible gloss would claim that R represents S as a moving worm-shaped object, and another would claim that R represents S as a nutrient. Since, according to Egan, the physical facts of the case would be neutral between these two interpretations of them, we have an example of how representationalist glosses can have explanatory value without being naturalistically determined. A choice between them could only be motivated by the explanatory goals of scientists providing the glosses.

After providing an eagle’s eye view of her proposal in the first half of the book, Egan turns in the second half to examining representationalist accounts of three prominent cognitive domains, arguing in each case that it would be a mistake to view the account as literally true, but maintaining that each account is illuminating when understood to be a representationalist gloss on a set of physical facts. The domains are propositional attitudes, bodily sensations, and visual experiences. Unfortunately, space limitations preclude discussion of Egan’s views about more than one of these domains. I will focus here on what she has to say about propositional attitudes, but her ideas about bodily sensations and visual experiences are different and no less worthy of consideration.

Egan’s proposal about attitudes is complex, but for purposes of this review it can be simplified as follows. Consider an agent who possesses a functional architecture that conforms to the requirements of folk psychology. Thus, the agent has internal states that are linked to each other, and to external stimuli and types of behavior, by a system of causal relations that meets those requirements. Suppose also, however, that the agent only weakly realizes folk psychology because the states in question are individuated purely in terms of causal relations, with intentional contents playing no role. Given this restriction, it can’t be literally true that the states count as beliefs and other intentional states; but even so, we can use terms like “belief*” and “desire*” in describing the agent, provided that we thereby mean only that the so-designated states have the causal roles attributed to beliefs and desires by folk psychology. Egan would, I think, regard a theory of the agent that is guided by these assumptions as a theory proper of the agent’s mind. (Egan does not use starred terms like “belief*,” but they are useful tools for streamlining the exposition. Their use is warranted because we are considering a functionally defined structure with states that meet the causal conditions specified by folk psychology, which assigns differentiating causal roles to the class of beliefs, the class of desires, and so on.)

As I interpret Egan, a gloss on this theory would take the form of a function F mapping sentences of a spoken and written language to the agent’s belief*s, desire*s, and so on. The mapping would associate sentences with attitude*-states in such a way that inferential relations defined over the sentences would correspond uniquely to causal relations among the belief*-states and desire*-states. Thus, for example, if the F-images of sentences of the following forms were belief*-states:

If P, then Q.

P

those belief*-states would have a strong tendency to cause the belief*-state corresponding to Q under F. Here the sentences would have truth conditions and syntactic properties in the literal sense, but their images under F would not.

To appreciate that the theory proper we have been considering admits of multiple representationalist glosses, we need only recall that representational contents are not uniquely determined by causal relations. This is shown, for example, by the causal chain problem, which points out that if a mental state has one external cause, then it has many. In more positive terms, it is shown by the fact that friends of representationalism like Neander and Shea find it necessary to supplement appeals to causal relations with appeals to more complex properties in analyzing the relation X represents Y. (Readers of Neander and Shea will remember that Neander appeals to information functions in analyzing this relation, and that Shea appeals to stabilized task functions. (Neander, 2017; Shea, 2018)) Now as we saw, the internal states specified by our theory proper are individuated purely by causal relations to each other, to external stimuli, and behavior. It follows that if we want to map sentences with determinate truth conditions onto our agent’s attitude*s, we will have to make a choice among possible mappings that is not determined by the causal facts. Egan would say that pragmatic considerations provide the only possible basis for such a choice.

I have several reservations about the foregoing components of Egan’s position.

First, I worry about her initial argument that it is pragmatic considerations rather than naturalistic factors that determine the acceptability of hypotheses about representational contents. As becomes apparent on inspection, the argument presupposes that representationalism is committed to the following two theses:

(1) There is just one representation relation binding mental states to representata.

(2) Pairs of hypotheses about content like these:

(i) Mental state S represents the property-complex moving and worm-shaped object.

(ii) Mental state S represents the property nutrient.

are in competition to specify the unique representatum of a given mental state.

Given these presuppositions about the commitments of representationalism, a naturalistic version of representationalism would have to provide a naturalistic basis for choosing between the two hypotheses. As Egan sees the situation, however, there is no such naturalistic basis for choice. On her view, the principal difference between the hypotheses has to do with what they can be used to explain: the first hypothesis contributes to explaining how a visuo-behavioral mechanism works, and the second hypothesis contributes to explaining how the toad survives in its niche. Accordingly, giving preference to one of the hypotheses depends on the explanatory goals of a philosopher or toad scientist, which means that it depends on contextual pragmatic factors.

I have concerns about this argument because I see no reason to think that representationalism is committed to (1) and (2). It seems entirely compatible with the spirit of representationalism to allow that there are two or more representation relations linking internal states with properties of external objects, where each of these relations can be explained in naturalistic terms. To accommodate the case at hand, for example, we might distinguish between causal mechanism representation and ecological significance representation, perhaps explaining the first in terms of the teleo-causal apparatus that Neander presents (2017), and the second in terms of the ecological apparatus that Millikan constructs (1989). (Both of these proposals are fully in keeping with the requirements of naturalism.) It follows from this that hypotheses like (i) and (ii) are not in competition. Indeed, they can both be true, provided that they are understood to be concerned with different representation relations. To be sure, if a theorist chooses between them, the ground of her choice will no doubt be pragmatic, having to do with her explanatory interests, but she could choose both if her interests were sufficiently broad. After all, they can both be true. Moreover, there is nothing special about a situation in which a scientist bases a choice between true hypotheses on explanatory interests. That is in fact how scientists generally proceed when confronted with pairs of true hypotheses, and they do so without in any sense deflating their theories. (Representational pluralism is advocated at length in Hill, 2022.)

Second, I have apprehensions about Egan’s treatment of propositional attitudes. One reason for worry is that she assumes without explanation that the natural language sentences that model functionally defined belief*-states and desire*-states have a full share of meaning, including truth conditions. It seems that if the idea of mental states with representational contents is problematic, the idea of syntactically individuated sentences with truth conditions should be equally problematic. Also, functions like the function F cited five paragraphs ago are not well defined. As we saw, given an inference involving premises P1,…,Pn and conclusion C, F is supposed to map these sentences onto belief*-states B1,…,Bn+1 just in case B1,…,Bn cause Bn+1. But there will be many sets of n + 1 belief*-states that meet this weak causal condition.

Egan’s deflationary account of the representational contents of beliefs is quite different than her accounts of bodily sensations and perceptual experiences. The present concerns about the former account do not apply to either of the latter. However, the earlier point about multiple representation relations would seem to raise general questions about the motivation for Egan’s deflationary project as a whole.

REFERENCES

Hill, C. S. (2022). Perceptual Experience. (Oxford: Oxford University Press).

Millikan, R. G. (1989a). “Biosemantics,” Journal of Philosophy 86: 281–97.

Neander, K. (2017). A Mark of the Mental: In Defense of Informational Teleosemantics. (Cambridge, MA: MIT Press).

Shea, N. (2018). Representation in Cognitive Science. (Oxford: Oxford University Press).