Are Sensory Concepts Learned by “Abstraction” from Experience?

Abstract

In recent years, many philosophers and scientists have argued or accepted that it is impossible to learn primitive sensory concepts like “blue” and “red”. This paper defends a more qualified picture. I try to show that some received characterisations of “learning” are nonequivalent and point towards different learning-nonlearning distinctions. And, on some ways of specifying such a distinction, it might be correct that we do not and cannot “learn” a concept of blue. But on other ways of specifying such a distinction, we can and do sometimes “learn” a concept of blue from experiences of blue. The latter part of the argument connects with some traditional “abstractionist” views, and I defend the present claims in view of some widely circulated concerns about “abstracting” concepts from experience. I close with some reflections on how one might, in view of all this, think about “the learning-nonlearning distinction”.

1 Introduction

As I will use terms, to “have a concept” of something is to be able to think about that thing. To have a concept of horses is to be able to think about horses; to lack a concept of horses is to lack that ability.¹

The question for this paper is whether one can learn sensory concepts from experiences of that to which these concepts apply; for example, whether one can learn to think about blue from experiences of blue. It may be natural to suppose that this is not just possible but that we humans in fact sometimes do learn sensory concepts this way. However, the view that we do not ever learn a concept of blue from experiences of blue, and that this is not even possible, has in recent years been widely accepted in science and philosophy. The main root of this state of the art is a line of argument by Jerry Fodor (in, among other places, 1975, chap. 2, 1981, 2008, chap. 5).

Fodor defends to begin with this claim:

PrimitiveNotLearned: It is impossible to learn a primitive concept.

That a concept is primitive means for Fodor that it lacks analysis. For example, the concept “white horse” can plausibly be analysed in terms of the components “white” and “horse”. It is more controversial whether either of the concepts “horse” or “white” can be analysed in terms of more primitive components. If they cannot they are “primitive”.

To learn a concept is according to Fodor one way of coming to possess it; but just one. As Fodor uses terms, there is both (a) concept acquisition (and cognitive development more generally) that amounts to learning, and (b) concept acquisition (and other cognitive development) that does not amount to learning. Here is one passage that draws this distinction:

…concept learning is some sort of rational process. It contrasts sharply with kinds of concept acquisition where, for example, a concept is acquired by surgical implantation; or by swallowing a pill; or by hitting one’s head against a hard surface, etc. Intuitively, none of these is concept learning, but any or all of them might eventuate in concept attainment (2008, 135).

This passage appeals to an “intuitive” separation between concept learning and concept acquisition that is not learning. It also characterises concept learning as a “rational” process. In other passages, Fodor characterises concept learning as “computational”, “intentional”, “cognitive” and “in the domain of cognitive psychology” (1975, 93, 2008, 136–7, 139, 151–2), and concept acquisition that is not learning as “automatic”, “noninferential”, jumpy,² “brute-causal” (1981, 273), and as “merely triggered” by circumstances (1981, 275). These different characterisations will be central in what follows.

In addition to PrimitiveNotLearned, Fodor defends:

ManyPrimitive: Many of our concepts are primitive.

ManyPrimitive is vague, but the idea is that primitive concepts include sensory concepts like “blue” and “red” and much more besides; Fodor mentions in these contexts concepts like “horse”, “tree”, “carburettor”, and “umbrella” (2008, 129).

The conjunction of PrimitiveNotLearned and ManyPrimitive entails:

ManyNotLearned: Many of our concepts are such that it is impossible to learn them.

This conclusion, ManyNotLearned, is widely contested. However, critics of it have typically objected specifically to the idea that it is impossible to learn concepts like “horse” and “carburettor”. Many of them have accepted Fodor’s view that it is impossible to learn primitive sensory concepts, like “blue” and “red”.³

Such theorists come in two types. The first type is exemplified by Stephen Pinker. Pinker maintains that, contrary to ManyPrimitive, few of our concepts are primitive, and “carburettor”, for example, is not among these (2007, 107). But Pinker accepts that there are some primitive concepts, and that primitive concepts are innate (and therefore, presumably, not learned), and he thinks that “red” is “patently” among those (ibid., 92–3).⁴

A second type of theorist accepts ManyPrimitive but urges, contrary to PrimitiveNotLearned, that primitive concepts come in two different types: Some primitive concepts, like “horse” and “carburettor”, can be learned. But others cannot. And the learning of the former builds upon the latter. This kind of view is held by Eric Margolis (1998), Stephen Laurence and Margolis (2002), Margolis and Laurence (2011), Susan Carey (2009, 2014, 2015), Daniel Weiskopf (2008), and Kim Sterelny (1989), all of whom more-or-less clearly suggest that the class of primitive concepts that cannot be learned includes sensory concepts like “blue”.⁵

Thus, theorists of both types agree with Fodor that primitive sensory concepts like “red” and “blue” cannot be learned. They disagree with Fodor in maintaining that “carburettor” can be learned. And they disagree with each other about the reason why “carburettor” can be learned: the former theorist maintains that “carburettor” can be learned because it is nonprimitive; the latter maintains that “carburettor” can be learned despite being primitive.

I shall here leave aside concepts like “carburettor” and focus solely on whether primitive sensory concepts can be learned.

I will accept some established frames of the debate. To begin with I will—at least up until the last few paragraphs of the paper—rather uncritically use Fodor’s characterisations, cited above, as guides as to whether a given process of cognitive development amounts to learning. I think these characterisations are widely accepted. They are not questioned by any of the above-mentioned critics of Fodor’s argument, even though they are often highlighted.⁶ And they are sometimes echoed by others.⁷ I shall also, at one juncture, rely on an argument by Fodor for “computationalism” about the mind. This argument too is, I think, widely accepted and even more widely not-questioned both in this debate and beyond.

I will try to make plausible that sensory concepts are sometimes acquired from experience. I will also try to make plausible that this acquisition process has some but not all of the characteristics associated by Fodor—and others—with “learning”. And therefore, whether it is correct to say that this acquisition process amounts to learning depends on the exact precisification of the learning-nonlearning distinction.

It is an interesting possibility that the case of experience-based sensory concept acquisition is exceptional. It may well be that processes of cognitive development typically have all or none of the characteristics standardly associated with “learning”, and therefore fall more clearly on one side or the other of the learning-nonlearning divide. For example, Sect. 3 below discusses a case that arguably has all the characteristics standardly associated with “learning” (like being “rational”, “computational” or “cognitive”). The typical development of our mnemonic abilities arguably has none of them. I shall not defend these suggestions; I just note the possibility that experience-based sensory concept acquisition is an interestingly special case where a set of characteristics that are typically all present together or all absent together fails to be so. And, anyway, whether it is exceptional or not, I think this case, where a set of characteristics standardly associated with “learning” are (a) not all present and (b) not all absent provides an illuminating perspective on the distinction between learning and nonlearning.

I shall give the name ‘abstraction*’ to the process by which primitive sensory concepts can be, and are sometimes, acquired and in a sense “learned” from experience. I choose the ‘abstraction’-part because the process has some characteristics that have historically been associated with that label. I choose the ‘*’-part because the process has some characteristics that have historically not been associated with ‘abstraction’. The historical connections are not central to my interests here, but I shall make a few remarks along the way about how the present picture compares with some historically held views.

My ambitions are limited. I mentioned some frames of the debate that I shall accept. I will also not argue for anything stronger than that primitive sensory concepts can be and sometimes are acquired and in a sense learned from experience. I will, for example, not argue for any of the following claims: (1) that sensory concepts can be acquired from experience without the prior possession of other concepts; (2) that sensory concepts in some sense form the “basis” of our acquisition of all our other concepts; or (3) that all our concepts are primitive sensory concepts or combinations of such. All the central claims to be made will be compatible with each of these claims. Some of the below may also be of use for someone who wishes to defend either of these claims. But it is not part of my present aim to defend any one of them.⁸

Section 2 specifies some central terms. Section 3 presents Fodor’s main argument for PrimitiveNotLearned; aspects of this argument will serve as an important foil in later sections. Sections 4 and 5 contain the core of the present argument. Section 4 argues that we can acquire a concept of blue from experiences of blue, and Sect. 5 argues that, drawing on some received ways of characterising “learning” and “nonlearning”, this concept acquisition is an instance of learning. Section 6 addresses an objection deriving from Geach and Sect. 7 a worry concerning attention and circularity. Section 8 summarises and spells out four different ways of thinking about “the learning-nonlearning distinction” in view of the preceding.

2 Terminology

2.1 “Having a Concept”

I said that to “have a concept” of something is to be able to think about that thing. That was a bit imprecise. More precisely, to have a concept of something at a time t is to be able to think about that thing at t without first doing something else and without anything first happening to one. To illustrate both the content and the naturalness of this “immediacy” restriction: There may be some mathematical property, P, that I can think about if—but only if—I first go through some mathematical training. If in my current state I am able go through that mathematical training there is a sense in which it is now correct to say that I “am able” to think about P: I am able to think about it in a sense that a cat, say, is not. But I am not in the present “immediate” sense able to think about P, because I am not able to do it without first doing something else. And therefore I do not now “have a concept” of P in the present sense of the phrase. Similarly, there may be something that I can think about if—but only if—I first have a certain experience. If I am now able to have that experience, there is a sense in which I am now able to think about that thing. But I do not have the thinking ability in the present “immediate” sense since I can employ it only if I first have the experience. And therefore I do not now “have a concept” of that thing, in the present sense.⁹

This specifies what it is to be able to think about something. It does not specify what it is to be able to think. I shall be minimally committal about this. I assume that there is a familiar and ordinary use of the term ‘thinking’ that clearly enough points towards a relatively natural cognitive kind that different philosophers and psychologists have offered competing or supplementing theories of. I have in mind the use of ‘thinking’ in sentences like, ‘I am now thinking about Tutankhamun’ and ‘the neighbour’s dog cannot think about Tutankhamun’. About this cognitive kind, some have held that nonhuman animals do not instantiate it (Davidson 1975) while others have held that some such animals do (Bermudez 2003; Carruthers 2004). Some have held that its instantiation in us involves the instantiation of sensory “ideas” (Hume 1739); others that it involves nonsensory symbols in a language of thought (Fodor 1975); and yet others that it involves acceptance of certain inferences or transitions (Peacocke 1992). To “have a concept”, as that phrase is understood here, is to be able to engage in this kind of cognitive act, however that kind should ultimately be understood.¹⁰

From this understanding of ‘having a concept’, there is a straightforward extension to related phrases. Thus, to “acquire a concept” of something is to acquire an ability to think about that thing, to “learn a concept” is to learn an ability to think about something, and so on. I believe, and will assume, that the just-explained understanding of such phrases agrees with how they are understood in all the literature discussed in this paper.¹¹

Philosophers often highlight the possibility of having multiple abilities to think about one and the same thing (compare Frege 1892). But issues about the individuation of thinking abilities will be of marginal importance in what follows. I will therefore allow myself locutions that are arguably inappropriate, like ‘acquiring the concept “so-and-so”’. These locutions can be paraphrased in terms like, ‘acquiring a certain, recognisable ability to think about so-and-so’.

2.2 “Blue”

My question is whether one can learn “sensory concepts” from experiences of that to which they apply. We can distinguish at least two types of sensory concept: (1) concepts of specific sensory qualities, like the particular shade of blue that I right now experience, and (2) concepts of generic sensory qualities, like the concept “blue” that encompasses many different specific shades. I will argue that it is most clearly the case that we can acquire—and in a sense learn—concepts of specific sensory qualities from experiences of that to which the concepts apply. For reasons of space, I must set aside a discussion of whether this conclusion can be generalised to other types of sensory concept. Since it would clutter the exposition to keep track of different sensory quality concepts throughout, I will in many contexts use imprecise talk like “sensory concept” and “concept of blue”.

3 Fodor’s Argument for PrimitiveNotLearned

To repeat: Fodor and many others maintain that primitive concepts cannot be learned. Yet others maintain that this is true at least for primitive sensory concepts, like “blue”. Why might one believe either of these things? A central line of thought in Fodor’s work can be summarised thus:

• (1) If concepts are learned, they are learned by a “process of projecting and confirming hypotheses about what the things that the concept applies to have in common” (2008, 132).

• (2) But concepts—or at least primitive concepts—cannot be learned in this way.

• (3) Therefore concepts—or at least primitive concepts—cannot be learned.

This summary states two different arguments: a weaker one with a conclusion restricted to primitive concepts and a stronger one with a conclusion about concepts generally. We will get to the difference between these arguments shortly.¹²

Fodor develops the argument with a focus on the following “concept learning paradigm” in psychology. A test subject is provided stimuli consisting of different combinations of shapes and colours, for example red triangles, green triangles, black squares etc. The subject’s task is to figure out what it is for something to be “flurg” on the basis of data points like: “this one is flurg”, “that one is not”. In a given case, being “flurg” may be the same as being green or triangular.

It may be right—relatively uncontroversial even—that when a subject in such an experiment comes to understand what it is to be “flurg” then there is something that she has learned. It seems clear that there is something that she has come to know that she did not know before, and it may well be right that her acquisition of new knowledge is a learning process. In any case, it seems clear that she has acquired new knowledge through a reasoning processes of this kind: “To be flurg cannot be the same as being red-and-triangular since the green triangle was flurg. It cannot be the same as being triangular since the green square was flurg. But the available data allow that being flurg is the same as being green-or-triangular”.

It is however debatable whether there is a concept that one learns—or even, typically, acquires—in this kind of experiment. If there is a concept that one typically acquires it is presumably the concept “flurg”. But that concept is plausibly one and the same as the concept “green or triangular”, and subjects in this experiment typically can think green or triangular even at the outset. And if they can do that, they do not acquire that concept in the course of the experiment.

If they do not learn a concept, what do the test subjects learn in this experiment? One possibility is that they learn a new linguistic term (‘flurg’) for something that they could already think—and, for that matter, say in other terms. That would be to say that subjects increase their vocabularies but not their mental or linguistic “expressive power”.

I believe Fodor was always favourable to this conclusion: that subjects never learn—and typically do not acquire—any concept in the “concept learning paradigm”, and that this illustrates that concepts generally cannot be learned.¹³ He was not consistently adamant about this strong conclusion. Fodor sometimes concedes that complex concepts, like “flurg”/”green or triangular”, might in some sense be acquired, or even learned, in the experiment. For example, in one passage he floats the idea that a concept that is at the outset available only for “hypothesis formation” comes to be available for a wider range of cognitive tasks, like reasoning, organisation of memory and perceptual identification and that there is therefore a sense in which the concept may be said to be “acquired”.¹⁴

But whether or not the complex concept “flurg”/”green or triangular” may be in some sense acquired or learned in the experiment, Fodor was always adamant that this can plausibly be maintained at most for such complex concepts. There is no plausibility at all in maintaining it for the case of primitive concepts like “green”. The reason is that these are concepts one needs in order to at all go through the reasoning that the learning process consists in; they are the concepts one reasons with when one considers what is compatible and not compatible with one’s data.

I am happy to accept that no concept—not “flurg”, nor “green”—is ever learned, and that none is typically acquired in the kind of process of “projecting and confirming hypotheses” that Fodor describes. This is to say that I am happy to accept premise (2) of Fodor’s argument, both the strong and the weak version. To the extent that I shall question Fodor’s view, I shall question premise (1), that this kind of projecting and confirming of hypotheses is the only way in which one can learn a concept. The next two sections will develop this.

4 We Can and Do Sometimes Acquire a Concept of Blue from Experience

I shall assume that it is possible to acquire a concept of blue from experiences of blue. And I shall assume that human subjects in fact sometimes do acquire a concept of blue that way.

I think these assumptions are extremely plausible. The most convincing cases concern, I believe, acquisition of specific sensory quality concepts, like concepts of specific shades of colour. Take some specific shade of blue that I have never experienced, do not know whether anybody else has experienced, whether it has been named, whether anything instantiates it, or what its physical correlate is. It is plausible that, unless something first happens to me or I first do something else, I cannot now think about that specific shade. Having never experienced it I cannot think about demonstratively as “that shade”; not knowing whether it has been named I cannot think about as “the shade called so-and-so”; not knowing whether anybody else has experienced it I cannot think about it as “the shade that so-and-so experienced then and there”; not knowing whether anything instantiates it I cannot think about it as “the shade I would experience if I perceived such-and-such”; and not knowing its physical correlate I cannot think about it as “the shade with such-and-such physical correlate”. And, even if I have general grasp of colour space and its dimensions of difference and similarity, I plausibly cannot on that basis alone cognitively “determine” that particular shade and isolate it, in thought, from other nearby shades.

However, it is plausible that I could now experience that shade. And it is plausible that, if I did experience the shade then I could think about it by drawing on that experience. For example, as a result of having the experience I might be able to think that (as I may naturally put it) “ripe tomatoes do not have that shade of colour”. Moreover, it seems plausible that this kind of thing does in fact happen now and then. I believe it has happened to me for example.

Besides being plausible, the present assumptions are common ground, at least within a sociological frame that is significant for present purposes. For example, Fodor clearly accepts that we sometimes acquire a concept of blue from experiences of blue. He says:

it is not part of the Rationalist program to claim that your repertoire of primitive concepts is independent of the character of your experience. On the contrary … there are presumably triggering experiences that are nomologically necessary for the acquisition of sensory-primitive concepts like RED (1981, 274; see also 2008, 145).

Similarly, Pinker accepts that “patently atomic” concepts, like “red”, are “triggered directly by the eyes and visual system” (2007, 93), which I take to mean that these concepts are at least sometimes acquired as a consequence of visual experiences. And, as we shall see in Sect. 6 below, Peter Geach, in the context of his criticism of “abstractionism”, seems also to accept that a concept of blue can be acquired from an experience of blue.¹⁵

I will on the above grounds assume that some sensory concepts—at least concepts of specific sensory qualities—can be and are sometimes acquired from experience.

As I said, I call the process whereby a concept of (a specific shade of) blue is acquired from experience ‘abstraction*’. The process has some characteristics that have historically been associated with the label ‘abstraction’. For one thing, abstraction* is a process in which a sensory concept is acquired from experience of that to which the concept applies. Abstraction* is also, and more specifically, a process in which, typically, an ability to think about something, x, is in a certain sense “distilled” from an experience of x and something else. In a given case, an ability to think about a specific shade of blue may be acquired from an experience of something—a ball, say—as having that specific shade and such-and-such shape.

But abstraction* also has some characteristics that have not historically been associated with ‘abstraction’. For example, abstraction* is a process in which a concept can be acquired from a single experience. This is in contrast with for example Russell’s idea that “abstraction” builds upon many experiences:

When we see a white patch, we are acquainted, in the first instance, with the particular patch; but by seeing many white patches, we easily learn to abstract the whiteness which they all have in common, and in learning to do this we are learning to be acquainted with whiteness (1912, 101).

Further, abstraction* is a process by which concepts of specific shades of colour can be acquired. By contrast, it is natural to understand Russell in the quote above as having in mind a process that outputs “acquaintance” with generic whiteness (on the basis of acquaintance with “patches” that exemplify different shades of white).¹⁶ Note 18 below registers yet another difference between abstraction* and abstraction as traditionally understood.

5 Is Acquiring a Concept of Blue from Experience an Instance of Learning?

Given that one can, and that we sometimes do, acquire a concept of blue from experience, the central question for us is whether this process, abstraction*, is or is not a matter of learning the relevant concepts. The present guidelines on this issue will, to repeat, be Fodor’s characterisations of learning processes as “rational”, “computational”, “intentional”, “cognitive”, and “in the domain of cognitive psychology”, and nonlearning processes as “automatic”, “noninferential”, jumpy, “brute-causal”, and “merely triggered” by circumstances.

I think these characterisations suggest a way of distinguishing “learning” from “nonlearning” on which abstraction* is not a learning process.

To begin with, it is not clear that abstraction* includes any thinking or reasoning. In any event, it is not a process that is made up solely by thinking or reasoning: it begins with an experience and ends with a capacity for thought. To this extent and in this sense, abstraction* is nonrational and noncognitive. It stands in contrast, for example, to the learning process that takes place in the “concept learning paradigm” described in Sect. 3 above. Recall that Sect. 3 never questioned that this process is a learning process; it just discussed what its end result is—possession of a new concept or possession or of a new term for what the subject could already think and say. But whatever the end result is, the process of reaching it is, as was noted, a reasoning process.

Somewhat connected with the nonrational and noncognitive character of abstraction* is the fact that this process is not—or not solely—a personal-level activity. Gaining the ability to think about blue from experiences of blue is, at least in part, something that just happens to one, at the personal level.¹⁷ Of course, to use this ability is to engage in a personal-level action, namely thinking. But gaining the ability is not a personal-level action. As a consequence of having the experience, one comes to possess the new thinking ability just as passively as one comes to have the experience itself. To this extent and in this sense, the process is an “automatic jump” and it stands again in contrast to the learning process of the “concept learning paradigm”.¹⁸

Given these nonrational, noncognitive, and automatic-jumpy features, there is ground for taking abstraction* to be an instance of nonlearning, given Fodor’s characterisations of learning and nonlearning. However, there are also grounds for regarding abstraction* as a learning process, if we focus on certain other of Fodor’s characterisations.

To begin with, there is reason to regard abstraction* as a computational process. At least, there is reason to do so if we accept Fodor’s main line of argument for computationalism, which I will now briefly review.

Fodor is impressed by the fact that we can and often do explain action and cognition by appealing to psychological laws that are general both insofar as they quantify over subjects of mental states and insofar as they abstract from the contents of mental states. As an example of such a law, Fodor cites: “If x wants that P and x believes that –P unless Q, and x believes that it is within his power to bring it about that Q, then ceteris paribus x tries to bring it about that Q”, where P and Q are schematic letters that may mean anything (1987, 13). Fodor sometimes refers to such laws as “intentional generalizations” (1994, 4).

How do processes that fall under intentional generalisations work? How does one for example transition from (a) a state of wanting that P and believing that –P unless Q and that is within one’s power to bring it about that Q to (b) a state of trying to bringing it about that Q? According to Fodor, the computational theory of the mind offers the best—or only—available answer to these questions. This theory says that the transition from (a) to (b) is a process of symbol-transformation that is sensitive only to “syntactic” properties of symbols and not to the meanings or contents of these symbols, and yet reliably “respects” these meanings or contents. To illustrate, suppose that two symbols, O and E, are involved in a state of: wanting something that is symbolised by O and believing that what is symbolised by O will not come about unless what is symbolised by E comes about and that it is within one’s power to bring about what E symbolises. And suppose that O and E can be identified in the mind (as can letters on paper) by their shapes. The relevant mind may then be set up in such a way that, by forces that operate solely on the shapes of O and E, it generates a state that involves the symbol E and amounts to trying to bringing about what E symbolises. If O in fact means that an omelette be made, and E that one breaks an egg, this process will have amounted to a transition from (1) a state of wanting that an omelette be made and believing that an omelette will not be made unless one breaks an egg and that it is within one’s power to break an egg to (2) trying to bringing it about that one breaks an egg. It will have amounted to this even though the mechanism driving this causal process was “blind” to these meanings. A parallel story can be told should O and E be identified in the mind by, say, frequency of a certain kinds of neural firing rather than by shape properties.

Let us now, against the background of this argument, consider the process, abstraction*, of acquiring a concept of blue from an experience of blue, as this process now and then occurs in us. It seems clear that this process falls under a psychological law with the kind of generality—expressible by quantification over both subjects and contents of mental states—that tends to impress Fodor. That is just to say that the present focus on blue is arbitrary, as of course it is. It is not only the case that we humans can and sometimes do acquire a concept of blue from experiences of blue. We can and do sometimes also acquire a concept of yellow from experiences of yellow. More generally, it seems clear that some kind of “intentional generalisation” along roughly the following lines is true: We humans are such that, for all x’s in a somewhat extensive domain: if one of us lacks a concept of x and has an experience of x then that person acquires an ability to think about x. For present purposes it can be left open exactly how extensive the relevant domain is, but it presumably includes at least many (specific) sensory qualities. We can also leave open whether some further condition should be added to the antecedent of the just-stated generalisation, for example something about direction of attention.¹⁹ That would not make a difference that matters for present purposes. What matters for present purposes is just that some principle that in the above way quantifies over both subjects and contents of mental states is correct.²⁰

It is worth noting that it is not trivial that we sometimes acquire sensory concepts from experience by a process, abstraction*, that falls under this kind of “intentional generalisation”. A creature could be such that it tended not to acquire a concept of blue from experiences of blue, or a concept of yellow from experiences of yellow, but instead tended to acquire the former concept from experiences of white triangles and the latter from itches on the nose.²¹ It is a nontrivial fact that we are not creatures of that kind.

Now, insofar as there is reason to accept Fodor’s general argument that processes that fall under “intentional generalisations” are computational, there is reason to suppose that the process abstraction* is computational. And if, as Fodor suggests, being computational is a distinctive characteristic of cognitive development that amounts to learning, that is in turn a reason to regard abstraction* as a learning process.²²

Furthermore, abstraction* falls, I believe, in a distinctive sense “within the domain of cognitive psychology”. There is, I think, a certain “autonomy” to the explanation of someone’s ability to think about blue in terms of (a) that person’s having experienced blue, and (b) the fact that she is the kind of subject that tends to acquire an ability to think about blue from an experience of blue, and (c) the theory that this process is implemented by a “computational” mechanism. This explanation is, I think, “autonomous” in the sense that the process will not be much further illuminated by an account of the neural realisation of the relevant computational process. Suppose for example we were to find out that the postulated “syntactic” features of the symbols involved in this process are numbers of pulses rather than, say, frequencies of firing of certain neurons. That would add detail to the already given computational theory, but it would not provide more illumination than that. Something similar is true, I think, if we do not assume computationalism but explain someone’s ability to think about blue just in terms of (a) and (b).

Consider for contrast the imaginary case of some subject that is disposed to acquire an ability to think about blue from an experience of a white triangle. Here an account of neural realisation might well provide a greater degree of illumination. We could admittedly make some safe assumptions about why the subject has the relevant disposition, even without knowing details of neurology. Assuming that the mental at least nomologically supervenes on the physical, we could safely suppose that some kind of “hard wire” connects, in this subject, (1) the neural basis of the state of experiencing a white triangle with (2) the neural basis of the state of being able to think about blue. But the details of neurology could still explain why that hard wire exists. For example, its existence could be explained by the proximity of these two neural states. (It is easier for a “wayward” wire to accidentally grow between proximal regions than between remote regions; to that extent, proximity can explain the existence of such a wire.) It is hard, I think, to imagine a similar degree of illumination from a neural account added to the explanation of an ability to think about blue in terms of (a) an experience of blue and (b) the relevant intentional generalisation, perhaps together with (c) a computational implementation hypothesis.

If this is correct, then there is a sense in which abstraction* is more exclusively “within the domain of cognitive psychology” than is the process, of some possible subject, of acquiring a concept of blue from an experience of a white triangle. And if being “within the domain of cognitive psychology” is, as Fodor suggests, distinctive of learning processes, that is in turn another reason to regard abstraction* as a learning process.

6 Geach on “Abstractionism”

It is often supposed that Geach (1957) presents an important case against the idea that sensory concepts can be learned by “abstraction” from experience. I shall now consider whether Geach provides any serious concern for the proposal developed here.

Geach takes it as given that concepts essentially form a network of capacities to make various judgments. The view is illustrated with an analogy:

the ability to make an appropriate [chess] move … presupposes … the capacities to carry out the moves and captures that are lawful for the pawns and the various pieces. As these skills are related to the chess-move, so concepts are related to the act of judgment (13).

Applied to a sensory concept like “red”, Geach’s idea is, I think, that to have this concept is not merely to be able to form judgments like, “this is red”. It is a more complex ability that involves also being able to form judgments like, “this is not red” and “something could be red and triangular”.

Geach claims that this kind of complex ability cannot be learned from experience. He does not in fact provide much argument for this claim. He appears to regard it as more-or-less obvious. The two passages that contain his most developed thoughts behind this assertion read as follows:

if I possess the concept red, then I can perform acts of judgment expressible in sentences containing the word “red”. This ability, however, certainly cannot be learned by any kind of attention to red patches for any length of time; even if after a course of attending to red patches the ability turned out to be present, we should still be justified in refusing to say it had been learned that way. We can say this just as confidently as we can say that the ordinary use of the word ‘red’ cannot be learned by hearing the word ‘red’ uttered ceremonially in the presence of a red object—simply on this account, that such a ceremony is not the ordinary use of the word ‘red’ (1957, 34).

it is integral to the use of a general term that we are not confined to using it in situations including some object to which the term applies; we can use the terms ‘black’ and ‘cat’ in situations not including any black object or any cat. How could this part of the use be got by abstraction? (ibid., 34–5).²³

However, even if one grants Geach’s view that (1) having a given concept is to have a network of capacities to make a variety of judgments, one cannot, I think, derive from these passages a strong challenge to what has been defended above, namely that (2) drawing on at least some received characterisations of “learning” and “nonlearning”, we can and do sometimes learn sensory concepts on the basis of experience.

Note to begin with that—as I indicated earlier—Geach seems specifically to aim for the conclusion that sensory concepts cannot be learned from experience; he does not question that they can be acquired from experience. This seems clear from the following passage, where Geach defends his a priori approach to the topic of abstractionism by appealing to a limitation in his aims:

the connexion between learning and being able to do something is not just an empirical one; for we should plainly not be willing to call any process that ended in ability to do something “learning” to do it. Of course we cannot work out in our heads how abilities are acquired; but there are conceivable ways of acquiring them to which we should unhesitatingly refuse to apply the term “learning”. If … a boy could come to know Latin by submitting to a brain operation, he would not have learned Latin from the surgeon. Now abstractionism is a theory of how certain mental performances are learned; the discussion I am entering upon is designed to show that the processes alleged by abstractionists would not be a learning of how to make these mental acts (18–19).

Recall also Geach’s previously quoted words: “even if after a course of attending to red patches the ability turned out to be present, we should still be justified in refusing to say it had been learned that way”.

Moreover, there is, as Sect. 4 above argued, good reason for Geach to aim for this limited conclusion because it is very plausible that sensory concepts can be acquired from experience (indeed, they can presumably be acquired from anything, including pills) and even strong reason to think that sensory concepts sometimes are acquired thus.

But given this, it is clear that Geach’s thoughts do not put in question anything in the present proposal. Nothing in Geach’s passages challenge the grounds given in Sect. 5 for regarding the process of acquiring a concept of blue from experiences of blue as a learning process. The passages do not put in question (a) that this process, abstraction*, falls under an intentional generalisation and that there is to that extent reason to regard it is “computational”, or (b) that abstraction* falls in a distinctive way “within the domain of cognitive psychology”, or (c) that being computational or in the domain of cognitive psychology are characteristics of learning processes.

7 Worries About Circularity and Attention

“Abstractionist” views of concept learning or acquisition have often been subject to worries about circularity and attention. The general character of such worries is that: the idea of abstracting a concept from experience involves vicious circularity because (1) the presumed abstraction would require attention to relevant features in experiences, but (2) this attention would in turn require application and therefore prior possession of the concept allegedly abstracted.²⁴

I think there is no serious concern in this neighbourhood for the present proposal. To see this, it is helpful to first consider whether worries like the above concern concept learning specifically or concept acquisition more broadly. The answer to that must be, I think, that they concern the latter. (What would the alternative construal be? That acquiring a concept of blue from experience does not require concept-involving attention, but that this acquisition process is not an instance of learning because…?)

When that much is clear it is also, I think, clear that there is no serious concern of the present kind for the proposal made here. In Sect. 4 above, I tried to show that it is both extremely plausible and widely accepted that we can and sometimes do acquire sensory concepts from experiences. And what is actual is possible. We can therefore conclude that it is not both the case that (a) this process, abstraction*, requires a certain kind of attention, and that (b) this kind of attention presupposes application and therefore prior possession of the concept that abstraction* results in.

One may of course wish to know, if it not the case that both (a) and (b), which one is not the case? I do not think we need a fully worked-out answer to this question to be assured of our present conclusions, but I offer some brief thoughts.

If we focus on the human situation—concept acquisition or learning as we in fact go through it—I am inclined to think that abstraction* requires some attention; minimally some attention—either stimulus-driven (bottom-up) or goal-driven (top-down)—to some region of experienced space. Unless I at least attend to a region of experienced space where a specific sensory quality appears I tend not to acquire from experience the ability to think such things as, “ripe tomatoes do not have that shade of colour”. However, I see little reason to suppose that this kind of attention should require possession of the relevant sensory quality concept. And there is strong reason to suppose that this is not the case since it is (to repeat) (1) extremely plausible that we can acquire a concept of blue from experiences of blue (Sect. 4), and (b) plausible that such concept acquisition requires that kind of attention (this paragraph).²⁵

8 Concluding Remarks

I have set aside some large questions about whether sensory concepts provide the basis for all our thought. But I have noted that it is both plausible and widely accepted that we can and sometimes do acquire sensory concepts from experiences of that to which they apply (Sect. 4). I have also (in effect) observed that some received characterisations of “learning” and “nonlearning” are nonequivalent, and that, drawing on some of these characterisations—in particular, the characterisations of learning processes as “computational” and “in the domain of cognitive psychology”—there is reason to place the acquisition of a concept of blue from experiences of blue among learning process (Sect. 5). Some widely circulated concerns about concept acquisition by “abstraction” from experience do not provide reasons to revise that conclusion (Sects. 6 and 7). Taken together, this calls into question a currently wide-spread consensus that primitive sensory concepts cannot be learned (recall Sect. 1).

I have rather uncritically assumed Fodor’s characterisations as guides to what learning is. Let me close with some with some reflections, not constrained by that assumption, on how one might, in view of the above, think about “the learning-nonlearning distinction”. I will mention four possibilities.

First, one could propose a kind of eliminativism about learning. For purposes of education, research heuristics, or understanding cognition and cognitive architecture, it matters—so this thought goes—whether a given process of cognitive development is rational, cognitive, computational, active or passive at the personal level, and in the domain of cognitive psychology. However, the above shows that our concept of “learning” is a confused mix of dissociable elements. This concept is therefore best left behind by attempts to understand cognitive development in developmental psychology, education sciences, and philosophy of mind.

Alternatively, one could propose that, while the above shows that our present concept of “learning” is confused, we should conclude nothing stronger than that this concept needs some—possibly minor—refinement.

A third thought is that our concept of “learning” is perfectly fine. Specifically, it usefully collects a number of different features that almost never dissociate. Admittedly, the preceding shows that the concept does not clearly apply, or fail to apply, to abstraction*. But as Sect. 1 noted, this case might be exceptional; it might be a hard case that makes bad law, and we should not on the basis of this one case abandon a concept that is otherwise useful.

A fourth thought is that the preceding illustrates limitations of a reductive approach to learning. We should not expect learning to be well captured in other terms, like “cognitive”, “rational” or “computational”. Learning is what it is and not another thing, and it should be approached in an appropriately nonreductive manner.

I leave it for further research to determine whether either of these conclusions is right.