Intention
(Doing Away with Mental
Representation)
Jay L Garfield
Department of Philosophy,
University of Melbourne
Central Institute of Higher
Tibetan Studies
ROUGH DRAFT—COMMENTS INVITED
DO NOT QUOTE WITHOUT PERMISSION
1. The Metaphor
Wilfrid Sellars used to be
fond of saying two things about metaphors in philosophy: “Philosophy is nothing
but the construction of metaphors. We
pile them higher and higher until, like a house of cards, the entire structure
collapses. Then we search for a new set
of metaphors and start the process all over again.” But also, “all metaphors limp.
A metaphor always needs a commentary, and in the end it is as likely to
mislead as it is to illuminate.”
While the thoughts I express
in this paper are continuous with those I have been expressing for some time,
the immediate impetus derived both from a re-thinking of the importance of
recent work in embodied cognition—particularly that of Clark and Haugeland,
some conversations with Terry Dartnall, and a re-thinking of the importance of
the work and remarks of a few of those figures in the history of philosophy to
whom I regularly return: Berkeley, Hume, Kant and Wittgenstein.
Mental representation is a metaphor. It has perhaps become so entrenched that it
appears to have been frozen, and it is easy to lose sight of its metaphorical
character. Literally, a
representation is a re-presentation, a symbol that stands for something else
because that thing can’t be with us. I
send my parents photos of the grandchildren because e-mail is cheaper than air
tickets. I consult a map of Adelaide to
find the shortest route to the philosophy department because wandering through
the streets would take too much time. Perhaps
in a similar sense I use words in this discussion because the very ideas in my
head have no way of being transferred directly to yours. It is tempting to think that when you hear
them thoughts in your head substitute in further processing for the sounds I
produce or for the words they encode.
But if I am right in what follows, we should resist this temptation.
When we talk as though beyond
photographic, stylized pictorial representations, there are also linguistic
representations we make an interesting, and perhaps not entirely literal
extension of the concept; when we go one step further and take it that thought
demands mental representations we continue to borrow the idea of
“standing for” and suggest that the vehicles of our thought stand for their
referents in much the same way that pictures stand for theirs. This way of talking has, as I have noted,
become so terribly entrenched that it appears to encode an obvious truth. Many of us in the philosophy of mind and in
cognitive science have become so accustomed to thinking through this metaphor
that to question it smacks of eliminativism about the mental, behaviorism, or
some equally dark pathology. The only real questions, at least the only
questions that most of us debate seriously, then turn out to be those about the
precise mechanism, structure, or other features of our representations. (In
fact, even self-avowed eliminativists such as Paul Churchland (2002) use the
term without any hesitation, and talk both about representation in the mind and
in the brain, reserving their eliminativism for specific forms of
representations, such as propositional attitudes. I remain a friend of belief, though I find myself becoming a foe
of representation, and so position myself at the same time closer to the mainstream
and further on the lunatic fringe. Such
is the conceptual geography of this field.)
So now I want to disturb this
venerable house of cards, and to suggest that this metaphor has had its day;
that it no longer even limps, and that it is time to put it out of its misery.
First, a bit of loosening
up: As I noted above, we use
representations when the original just isn’t available. It would be great to see the grandkids, but
a picture will have to do. On the other
hand, while it would be nice for the grandkids to be among us, would it really
be so great to get the grandkids into our skulls? And what would it even mean to get
them into our minds? (Note here that on a properly Humean
understanding according to which the mind and its perceptions do not contrast
with, but “spread over” the world, this last idea might not be so bad, but in that
sense we are offering an alternative to a representational view of
perception and thought.) Do our thoughts about them serve as second-best
substitutes in the same sense that our photographs do? That is, if my mom asks how my son looks,
could I tell her just to think about him and find out for herself? Hardly.
Things are even worse in the case
of perception: If I see the children in
front of me, they are already present.
Why re-present them? What good would it do me to see them by
seeing an image of them? To what to which the children are not already
present would that re-presentation be present? And why is that second presence of
any special value? Finally, must it be
re-presented once again in order to be of any use? Surely the regress has to be
cut off somewhere and the obvious place to cut if off is at the root. This is not, however, to say that it is
obvious just how to cut that regress without loss of explanatory power. That is the burden of this essay.
In a short while I will start
talking about how to think about thought without thinking about representation,
and how to do so in a robustly realist, intentionalist spirit. But first, having emphasised the
metaphorical nature of the idea of representation, I want to prosecute the
claim that it is no harmless metaphor.
It really does mislead.
2.
The
Prevalence and Costs of the Metaphor
If I were to ask one hundred contemporary philosophers of mind for the clearest
example of an eliminativist with respect to the mental, I’ll bet that ninety
would answer, within about one hundred milliseconds on average, “Paul
Churchland.” So it is instructive to
turn to one of Churchland’s recent discussions of the topic of mental
representation, his APA Presidential address:
Suppose also the internal; character of each of the representational
spaces is not fixed by some prior decree, either divine or genetic, but
is slowly shaped or sculpted by the extended experience of the developing
animal, to reflect the peculiar learning procedures embodied in the brain’s
ongoing business of synaptic modification.....
...
So we begin by expanding the number of representational spaces, into
the hundreds and thousands, far beyond the Kantians pair. We locate them in
discrete anatomical parts of the brain.
We make each one of them plastic and multipotent in its semantic content
and its conceptual organization. And we
reach out to include motor cognition and practical skills, along with
perceptual apprehensions and theoretical judgment, as equal partners in our
account of human knowledge. [Churchland 2003, p 206]
Whatever it is that Churchland
proposes we eliminate in cognitive science, it is clear that it is not representations. In fact he supposes here quite blithely that
the neurocognitive processes that mediate not only conceptual thought and
perception but also motor control and skill acquisition are
representational. His brief in an
account of a plastic, distributed and superpositional model of representation
is, if anything, not the elimination of the category of representation
from cognitive science, but the extension of the category to virtually
everything that the nervous system does.
This push for homogeneity in theoretical approach may have something to
it, though I think that Churchland goes too far. I will, however, take sharp issue with the idea that the relevant
homogenizing category ought to be that of representation.
Andy Clark is another who might
be cited as an opponent of much explicitly representational accounts of
cognition. I will pay closer attention
to his work below. But for now note
that he shares this presupposition that
even the most distributed neural processing, so long as it is teleologically
connected to an object, is representational:
A distributed representation is an inner encoding in which the target
content is not carried by an individual recourse (e.g. a single neuron) and is
not necessarily carried by a spatially localized group of units or
neutrons. Instead, the context
(concerning e.g., the notion of an individual digit) is carried by a pattern of
activationwhich is spread across a population of neurons or units. Distributed encoding present a number of
advantages and opportunities, For example, the pattern itself can encode
significant structural information in such a way that minor variations in the
pattern reflect small but sometimes important differences in what is currently
represented. ... [T]he point for now is
simply that the brain may be using quite complex, overlapping, spatially
distributed representational schemes even in cases where we might intuitively
have expected a simple, spatially localized encoding strategy, as in the case
with there M1 motor area. Nature’s way,
it seems, is to use spatially overlapping distributed encoding to govern
related (but nonidentical) types of finger movement. The final picture is thus one in which specific cortical neurons
play a role in controlling several finger muscles and do so by
participating in widely spatially extended patterns of activity which
correspond to different types and directions of finger movement. [Clark 1997,
pp 132-133]
It is worth noting two features
of this brief discussion in getting a fix on just how embedded the
representational metaphor is in our thinking about the mind: First, there is the assumption that any kind of information-bearing states
or processes are representational, however much they might not look to be—that
representation is the default assumption regarding the nature of cognitive
activity; second, even the processes that subserve motor control are to be
thought of as representational—the drive to treat all cognition homogeneously
is not taken as a drive to treat conceptual activity as similar to motor
control in virtue of being non-representatnional, but the reverse.
We could go on multiplying
examples, but I trust the point is made.
Even the most radical opponents of “classical,” linguistic, or digital
models of thought presuppose a
representational model. But once we
begin talking, thinking and theorising in terms of representation, we find
ourselves drawn inexorably into a labyrinth of dialectical dead-ends. They constitute a depressing proportion of
recent literature in the foundations of cognitive science, and include such
debates as those concerning whether our representations are digital or analog;
whether they are imagistic or propositional; whether we share representations
with infralingual children and animals; and whether, in pursuing cognitive
science, we should be realistic or instrumentalist about these mental stand-ins
for reality. In each case we seem to
fall into a state of collective antinomy.
In each case, I suspect the
reason for the antinomy is that the debate in question proceeds by means of
metonymic equivocation: we are led by the representational metaphor to shift
conceptual focus from the contents of our thoughts (about which many of
these questions can be intelligibly and even fruitfully raised) to the nature
of the thoughts themselves, where the questions and the debates lose their
grip.
Debates about whether human
representation is digital or analogue have been notably sterile. Reasons for each alternative are matched by
rebuttals and equally plausible but equally refuted reasons for the opposing
conclusion. There are four possible
reasons for this state of affairs: It
might just be too hard a question. OR,
it might be that some human cognition is analogue (image rotation, say) and
some is digital (syntactic analysis).
OR it might be, that like the painting, it is digital with respect to
some level of analysis or explanatory interest, but analog with respect to
another. OR finally, it might be, that
like taking hold of the pipe, it doesn’t make any sense to ask of thought
whether it is one or the other, and that we are really asking questions about
the objects of thought and confusing them for questions about the vehicle.
Detailed argument against each of
the first three alternatives would be a long and difficult matter. Here are a few motivators for rejecting each
of them, however: questions about
whether a particular form of representation is analog or digital simply aren’t
in general too hard. Now, to be sure, questions about the mind are often
very hard, but in cognitive science we have become used to real progress on
empirical questions of all kinds, and this seems an empirical question. There simply is no independent reason to
suspect its enormous difficulty unless one accepts antecedently the
representational metaphor of the mental and thinks that those
representations are especially occult.
The some-some response gets its
motivation from intuitions about the phenomenology of mental imagery,
apparently corroborated by such data as those of Shepard and Metzler (xxx) on
the one hand and about the conviction that the computational model of mind must
be at least generally right, on the other.
But each of these is deeply
problematic: To the extent that our
intuitions about what goes on in our mind are any guide to what actually occurs
(and they are notably fallible in this regard across the board, as we all know)
they are useful with regard to the contents of our mental states and
processes—that towards which they are directed, and not with regard to the
nature of those states and processes themselves. It might well be that we can discover by introspection that we
think about images, or about sentences, but to argue that we have
some kind of privileged access to the vehicles by means of which we
think about those objects .
Nobody doubts this. The rotation studies are designed precisely
to get behind mere intuition, and the fact that response time for matching
block patterns to targets is directly proportional to the angle of rotation of
the pattern with respect to its target suggests that an inner image—an analogue
representation—is being manipulated by a process of mental rotation—an analog
process defined over analog representations.
But this interpretation of those data already presupposes a
representational model of thought. Let me explain: If I rotate a physical arrangement of blocks in my hand, the
blocks are rotated, not represented. If
I watch those blocks being rotated, I must visually detect those blocks, but
there is no need to represent them; I can simply see them. There may well be something analog about the
seeing: the states in my visual system that respond to the light reflected off
the blocks may vary continuously with their rotation. There may also be some digital features of the seeing. But none of these are features of
representations, only of states that mediate detection and interaction. Now, when I imagine the blocks being
rotated, my cognitive activity is similarly directed on the block, and in fact
recruits many of the same brain processes that subserve actual vision. What is going on is a kind of ersatz vision. To the extent that anything is being
re-presented it is not the blocks, but the visual or visuo-motor processes
involved in seeing them being rotated or in rotating them. Imagining on this account is exactly as
representational or as non-representational as vision. Analog processing does not demand analog
representation.
The view that thought is in
general a set of digital operations defined over a set of representations in a
language of thought was once central to our field, but I take it that by now it
has had its day. I will not flog that
particular dead horse, but only recall some of the blows responsible for its
demise: the failure to discover anything in the brain plausible like such
symbols; the apparent causal necessity of natural language acquisition for so
much human cognition; the success of distributed models of processing as models
of cognitive activities, etc... So
much for the some-some response. I
conclude that it might be a good idea to recuse from the debate about the nature of mental representation and ask
instead whether we can do better than representation as a metaphor for the
mental.
3.
Clarke
and Haugeland on Representational Realism
As I said at the outset, a
lot of my thinking about this problem was occasioned by another look at recent
work on embodied cognition. So now I
want to pay attention to an extended interchange between Andy Clark and John
Haugeland. Here is Clarke on
Haugeland:
Cognitive scientists often talk of both brains and computer models as
housing internal representation.” This
basic idea provided common ground even between the otherwise opposing camps of
connectionism and classical artificial intelligence. The differences between connectionists and the classicists
concerned only the precise nature of the internal representational system, not
its very existence. Classicists
believed in a “chunky symbolic,” inner economy in which mental contents were
tokened as strings of symbols that could be read, copied, and moved by some
kind of inner central processing unit.
For all that, explicit, chunky symbolic representations and
distributed vectorial connectionist representations were both seen as species
of internal representation, properly so called. This overarching species is present, it has been argued, whenever
a system meets certain intuitive requirements.
Haugeland (1991) unpacks these by describing a system as representation
using just in case:
4.
It must coordinate
its behaviours with environmental features that are not always “reliably
present to the system.”
5.
It copes with such
cases by having something else (in place of A signal directly received from the
environment) “stand in” and guide behaviour in its stead.
6.
That “something
else” is part of a more general representational scheme that allows the
standing in to occur systematically and allows for a variety of related
representational states. (62)
[Clarke 1997, pp 143-144]
Again, we have the
presupposition that in talking about inner cognitive states and processes we
are talking about representations and operations on them. Clarke notes that Haugeland analyses representations derivatively, beginning
with a characterization of the organisms or systems that use representations. For Haugeland, as Clark notes, the crucial
characteristic of such systems or organisms is that they coordinate their
activity with the (at least sometimes) absent or non-existent, and that they do
so by means of a “stand in” that substitutes for that which is not
present. That stand in, if it is part
of an articulated system of such stand-ins, and processes defined over them, is
a representation. Clarke quarrels with
the requirement that representations represent the non-present. He considers the neural states that track
rats’ head positions. These are
activation vectors that co-vary with the orientation of the head in the
environment :
It
seems reasonably clear that by glossing states of the neuronal populations as
codings for specific head positions we gain useful explanatory leverage. Such
glosses help us understand the flow of information within the system when, for
example, we find other neuronal groups (such as motor control populations) that
consume the information encoded in the target population. The strict application of Haugeland’s
criteria would, however rule out the description of any such inner systems of
non –decouplable inner states as genuinely representational. This seems unappealing in virtue of the very
real explanatory leverage that the representational gloss provides…. (Clarke
1997, p 145)
A couple of things deserve
mention here before we take stock and move to some of Haugeland’s more recent
remarks: Clarke argues that states such as those tracking and controlling rats’
head positions are representational despite the fact that they cannot be
decoupled from their putative representational content because they encode
information that is used elsewhere in the system. This may seem a small terminological quibble, but it in fact
raises an important issue. Clarke is
again drawing attention to a certain kind of homogeneity: The processes that track that which is
present are not all that different from those that track that which is
absent. So to the extent that we take
representational status seriously, one set of processes and states is
representational exactly to the extent that the other is. I think that intuition is right on target,
as is the observation about the
character of the cognitive processes in question, both with respect to their
intrinsic character and with respect to their relation to their objects. But
Haugeland suggests that we only have
representation when we have something that can “stand in for” or “re-present” that
which is absent. That is absolutely right as well. But these two premises give us, instead of the conclusion that
Clarke’s rat neuron populations are representational the conclusion that most
of the cases that he, Haugeland and others in this field regard as
representational are not.
This is not the conclusion
to which Haugeland comes, of course.
Consider this discussion in the context of a discussion of Cummins’
account of representation in Hugh Clapin’s excellent volume on the philosophy
of mental representation: Discussing a
dog’s cognitive engagement with a ball, Haugeland says:
...Shelia
doesn’t have any intenders the function of which is to target specifically
tennis-balls. Nor, likewise, can any of her representations have
specifically them as their contents. Of
course, if she has representations, then they must have contents; and then we
face the familiar muddle about how to specify them, given that all our
words are way too laden and sophisticated.
But that doesn’t strike me as a killer objection so much as an annoying
limitation. (Indeed, I think that we
ourselves have and use representation that we couldn’t put into words). So, taking the scare quotes suitably seriously,
I think we can say things like this. If
Rob were to throw a wad of paper across the room, such that Shelia caught it
out of the corner of her eye and lunged for it, maybe she mistook it for
a ‘tennis-ball.’ That is, maybe she targeted
‘that projectile’ and applied her ‘tennis-ball’ representation to
it. And when she caught it, she might
even realize her mistake (and glare at Rob). Why not? [Haugeland
in Clapin 2002, p.
141 ]
Dogs, too, according to
Haugeland, have representations, and even if their contents are hard for us to
specify, these are representational in the same sense that ours are. Indeed, on this view, the only way to
explain Sheila’s behaviour, whether accurately directed upon a tennis ball or mistakenly
directed upon a paper wad is to say that she was guided by an internal
representation of a tennis ball, which representation happens to match tennis
balls and fails to match paper wads.
Her mistake is one of comparison.
Now, there are mistakes of
comparison: When I see a photograph of
Brett and take it to be of Shane, I mistakenly apply a representation to that
which it does not represent. I have
compared it and mistakenly taken it to match.
But not all mistakes are mistakes of comparison or of
misapplication. When I add five and
seven and get thirteen I do not misapply a representation of thirteen to
twelve. I do not compare one with the
other and decide that they match. And,
closer to Sheila, when I swing wildly and miss because the ball was faster than
I expected, I do not first represent the ball as here, and then swing; I see
the ball and act. The relevant
cognitive processes are tightly coupled to the perception-action cycle. The error in question is more like the
failure of an arrow to hit a target than the misidentification of one target
apple as another. So, while it might be
that both error and success, and so all cognition, presupposes norms of
correctness, it in no way follows that cognition presupposes representation
whose success or failure to match that to which it is applied constitutes
success or error.
What do we learn from this interchange? The unquestioned
assumption that representation is the
foundation of cognition leads to quandary:
Representation should be decoupled from represented. Similar cognitive processes should be treated similarly by theory and
metaphysics. Immediate sesnsorimotor
engagement with the environment is a paradigm of non-representational activity,
but yet seems continuous in important respects with a great deal of more
reflective activity. So where is the
line to be drawn?
7.
Historical
Excursion
Berkeley famously attacks the
idea of mental representation because of its entailment of a doctrine of double
existence, a doctrine he finds unintelligible due to our inability to conceive
of any representeds for our putative representations to represent. I don’t want to worry about that
argument, and I don’t think that many today have much sympathy with the idea
that the external world is really inconceivable.
But I do want to focus on one
saner insight that Berkeley develops in the Dialogues en route to that
conclusion: an insight he develops in the context of his discussion of the bust
of Caesar. He grants that the
bust is a representation, precisely because it presents to us what is
absent, and more importantly, because it mediates our knowledge of that
which it represents. To perceive
Caesar, it would be best to have the Roman emperor himself in plain view. But alas, he is long dead, and the best we can
do is to view his bust. That is a way
of perceiving Caesar, but only a mediate way. Moreover, Berkeley notes, that mediation is made possible
precisely by a great deal of background knowledge (importantly, for instance,
the knowledge that it is in fact a bust of Caesar, and not, say, Cicero—a kind
of knowledge he correctly, if for the wrong reasons, notes must be absent with
regard to our inner states) and by a set of representational conventions—seeing
the bust we do not take Caesar to be represented as a cold, white, legless
monster, but as a flesh and blood Roman.
Without the relevant knowledge and the conventions, representation of
Caesar by the bust would be impossible; were we face to face with
Julius, the representation would have been unnecessary. Representation is a mediate way of knowing
the absent, in which mediation is achieved through the use of external props
scaffolded by a network of relevant conventions and auxiliary knowledge. Berkeley’s central and important insight
then is that perception (which he might well have understood in part in
that older sense of the term, given his ecclesiastical calling) is immediate
communion with its object, and so is the very antithesis of
representation.
Hume is another early
perceptive critic of the representational metaphor. He takes the doctrine of double existence to be a monstrosity,
while pointing out both that the only things immediately present to the
mind are its perceptions and that “‘tis vain to ask whether body exists;
this is something that must be taken for granted in all of our reasonings.” The
depth of Hume’s position only emerges when we take seriously the task of
rendering these claims—claims that to our ears, accustomed as they are to the
trope of representation, are prima facie inconsistent—consistent.
While Hume is sometimes read
as holding an imagistic model of thought, this reading is hard to sustain on a
careful reading of the Treatise, where literal talk about the mechanisms
of thought is almost always causal talk. Given Hume’s disparagement of the idea of double existence
and his insistence on taking for granted the existence of external
objects, it is tempting to think that we can only understand him as arguing
that there is no second internal existence of external objects as
representations. In fact, things are
more complex and more interesting.
Hume, at least as much as
Berkeley, takes seriously the etymology of “perception” as originally denoting
the act of receiving the Eucharist.
Hume is not so much a foe of the outer, as naïve readings of the Treatise
would have it, nor a foe of the inner, as a proto-behaviourist
reading might have it, as he is a foe of the very distinction between
inner and outer. In this regard he is
very much more a post-modern figure before his time—a fellow traveler of
Nietzsche and Heidgger—than he is a modernist to be lumped in intellectual
history with Descartes, Malebranche, Locke and Berkeley. Hume, after all, is writing a treatise of human
nature, and is very much concerned, as he puts it “to introduce the
experimental method into the moral sciences.”
So transcendental ontology is far from his mind. For Hume, perception is not in the
mind, nor is the object outside of the mind. Rather person, culture (this will become important later) and
environment are continuous aspects of nature, and perception is but a rapport
of the perceiver or thinker with its object—a communion. Perception then names the process or
relation, not an inner state re-presenting an outer object.
It is Kant, in the first Critique,
who really introduces the vocabulary of representation—vorstellung—into
philosophy in the form we know it even if Descartes had already represented the
“idea idea” as Rorty has called it. And
it is Kant who transforms modern thought about thought by turning away from the
idea as the unit of thought in favor of the judgment. These developments are, of course,
closely intertwined.
Let us begin with the second of
these contributions. Kant’s insight that the fundamental epistemological unit
is the judgment, the assertion that something is the case, as opposed to an
idea, an unstructured something of which the mind is the subject, opened the
way to seeing the contents of mind as having propositional structure, and so as
having the logical structure of linguistic representations. This insight opens
the way to seeing thought as involving deductive reasoning, and to seeing
cognitive success as akin to the truth of utterances. Now utterances, at least the assertions with which Kant was
primarily concerned as models of judgment, are plausibly at least often
representational. As Wittgenstein was to put it much later, they re-present states
of affairs. So Kant, by
internalizing these linguistic representations with their logical structure,
provided a clear analysis of just how representation might go, as well as a
mechanism for representing facts, not things, and so of
representational thought as active thinking, not the passive having
of ideas.
This contribution becomes
important in the present context for two reasons: First, in seeing thoughts as judgments, and so as akin to inner sentences
Kant provides a compelling reply to the critique of the double-existence model
of representation adumbrated by Berkeley and Hume. For on this account there is no second existence in the mind of
the objects of thought—trees, tables, numbers, golden mountains or unicorns—any
more than there is a second existence of these things in sentences of German or
English. Kant hence, by focusing on
truth as cognitive success, as inferential relations as constitutive of content
and sensitivity to norms as central to the cognitive, develops the first
sentential model of thought. This
proves to be an enduring model in the history of philosophy of mind in the
West, and preserves the representational content of the idea idea down to the
present. Secondly, by using languageand
the relations between sentences as the model for thoughts and cognitive
processes, Kant directs our attention to what, as I will argue below, is the
important germ of truth in the representational theory of mind: when we learn
to think in sentences, we learn to think about representations, and hence, in a
derivative sense, to represent.
Now back to the first
contribution: Kant uses the term “representation” in three ways. In one sense, a representation is a mental
episode, in the case of judgments, one that represents reality as being in a
particular way; in the case of concepts or intuitions one that represents an
abstract entity or a singular entity.
In another, a representation is an object. Space is a representation, and so is a house or a boat. In short, all phenomena, as opposed to
noumena, are representations in this second sense. In the third sense, representation is the process of
representing a representation (in the second sense) via a representation (in
the first). Representation is thus an
activity of mind as well as a mental episode, as well as the object of
representation.
It is important to see that Kant
hence provides the foundation for the contemporary orthodoxy regarding the
ubiquity of representation in cognition and regarding the linguistic structure
of the posited mental representations. On the other hand, Kant’s equivocation
on just what a representation is should give us pause. Do we need all three
senses in order to get all of these cognitive benefits? While it is impossible
to make sense of human cognition without taking the objects of thought
sometimes to be representations (Kant’s second sense, but in a restricted
domain) and while that requires us to take seriously the activity of
representation as one kind of thought, it requires us neither to see the objects
of thought as re-presented, nor to see all of thought, and in
particular, the vehicle of thought, as representational.
Wittgenstein in the Tractatus noticed
that Kant’s metaphor of representation included not only thoughts, but also the
overt utterances that were their Kantian model. This insight is, of course, momentous for the subsequent course
of Twentieth Century philosophy of language and philosophy of mind. The picture theory of language has, of
course, been subject to trenchant criticism, as has the view that all language
is representational; and indeed the view that representation is the
foundation of linguistic meaning—and indeed the latter two critiques originate
famously with Wittgenstein himself.
Despite the fact that these theses bid fair to be the core of the Tractatus
model of language, there is a significant baby to be strained from the
bathwater, and indeed the baby destined to grow into the later
Wittgenstein: it is not thought—something
interior to the mind—that is the primary locus of representation, but language—a
public activity, whose representational power is vouchsafed by rules of
interpretation. The Kantian insight
that representation is crucially dependant upon responsiveness to norms is
retained; but the norms in question are norms of interpretation rather than
exclusively norms of inference; while their essential publicity has not yet
come to the fore, it is certainly in the offing.
In the Investigations, Wittgenstein
explores the limits of the picture metaphor, an exploration sufficiently familiar
that we can cut to the chase. The chase
scene emerges at section 308:
308. How does the
philosophical problem about mental processes and states and about behaviorism
arise? The first step is the one that altogether escapes notice. We talk of processes
and states and leave their nature undecided. Sometime perhaps we shall know
more about them -- we think. But that is just what commits us to a particular
way of looking at the matter. For we have a definite concept of what it means
to learn to know a process better. (The decisive movement in the conjuring
trick has been made, and it was the very one that we thought quite innocent.)
-- And now the analogy which was to make us understand our thoughts falls to
pieces. So we have to deny the yet uncomprehended process in the yet unexplored
medium. And now it looks as if we had denied mental processes. And naturally we
don't want to deny them.
We don’t want to deny them. But precisely what do we not want to
deny, and why? And what is the
difference between accepting mental processes and accepting the “analogy which
was to make us understand out thoughts?”
Answering these questions takes us to the heart of the problem caused by
the representation metaphor and allows us to glimpse its solution. It would be mad to deny that we think; that
we have beliefs, and so forth. And it
would be mad to deny that when we think and when we believe, we think and
believe about things. Neither
cognition nor its intentionality is up for elimination. But it is all too easy to move from thinking
to thoughts, and then from talk about thoughts to talk about particulars that
are thoughts. Having hypostasised these
particulars, and having characterized them intentionally, we move right along,
via the picture metaphor (or its linguistic cousin—there is no real difference
here), to thinking of them as inner representations and hypostasising the
processes in which they figure as computations over representations.
But now, if we have followed the
argument of the Philosophical Investigations to this point, everything does
fall to pieces: For we find ourselves
trying to understand thought both as primitively representational—as
meaningful—and committed to understanding meaning as essentially
rule-determined, and rules as primitively social. The “decisive move in the conjuring trick” was the
reification—the hypostasy of specifically inner states and processes as
the embodiments of thoughts and of thought.
Accepting the mental, however, no more requires positing these inner
entities standing for that about which we are thinking than accepting the fact
that the arrow points left requires positing an entity within the arrow
that stands for leftness.
The problem with the
representational model of mind—independent of the particular model of
representations—is that it takes thought about things to consist precisely in
the deployment of inner tokens or processes whose purpose is to stand in for,
rather than to enable rapport with, the objects of thought. The question that is left open—that of just
how standing in for something enables thinking about it—is exactly the same
question that representation was posited in order to answer in the first place:
how can thought be about something.
Pointing to an isomorphism is plainly inadequate and even
irrelevant. The same questions arise
immediately. And if we shift ground and
talk about selection history or ability to guide behaviour it turns out that
the representation relation has not been explained but rather abandoned. For nothing in these stories requires the
doctrine of double existence.
5. Intending without representing
As I noted above, much of my
discontent with profligate talk about mental representation arose from reading
Andy Clark’s recent work even though Clark, at least in some moods, does not
endorse my position. Consider the
following remarks:
The
status of an inner state as a representation thus depends not so much on its
detailed nature... as on the role that it plays within the system. It may be a static structure or a temporally
extended process. It may be local or
highly distributed. It may be very
accurate or woefully inaccurate. What
counts is that it is supposed to carry a certain type of information and
that its role relative to other inner systems and relative to the production of
behaviour is precisely to bear such information.
…[L]et us call a
processing story representationalist if it describes whole systems of
identifiable inner states (local or distributed) or processes (temporal sequences
of such sates) as having the function of bearing specific types of information
of external or bodily states of affairs.
Representationalist theorizing thus falls towards the upper reaches of a
continuum of possibilities whose nonrepresentationalist lower bounds include
mere casual correlations and very simple cases of what might be termed
“adaptive hookup.” Adaptive hookup goes
beyond mere casual correlation insofar as it requires that the inner states of
the systems are supposed (by evolution, design, or learning) to coordinate its
behaviors with specific environmental contingencies. But when the hookup is very simple (as in a sunflower, or a
light-seeking robot), we gain little by treating the inner state as a
representation. Representation talk gets
its foothold, I suggest, when we confront inner states that, in addition,
exhibit a systematic kind of coordination with a whole space of environmental
contingencies. {Clark 1997, p 147]
The
most potent challenge to a representation-based understanding comes, we saw,
from cases in which the web of casual influence grows so wide and complex that
it becomes practically impossible to isolate any “privileged elements” on which
to pin specific information-carrying adaptive roles. Such cases typically involve the continuous, reciprocal evolution
of multiply tightly linked systems, whose cumulative (‘emergent’) effect is to
promote some kind of useful behaviour or response. [Clark 1997, p166-167]
The
notion of internal representation thus gets a grip only when we can make
relatively fine-grained assignments of inner vehicles to information-carrying
adaptive roles. … At the very least, we
can now see more clearly what it would take to undermine a representation-based
approach: it would require a demonstration that, even in the
representation-hungry cases, it remains practically impossible to isolate any
system of fine-grained vehicles playing specific information-carrying adaptive
roles. [Clark 1997, pp 168-p.169]
Clark certainly recognizes that a
lot of cognition may involve simply what he calls “adaptive hookup,” and that
this in no way implicates representation.
But he does argue that there is nonetheless a class of
“representation-hungry” cognitive processes that require us to posit a system
of mental representations at least for complex mammals such as ourselves, and
he asserts that the necessary conditions for a state’s being a representation
are that its function is to bear information about the environment and that it
is a member of a family of “fine-grained vehicles playing specific
information-carrying adaptive roles.”
Now I will shortly deploy other
remarks of Andy Clark against the position he defends here. But first I want to note some points with
which I do not wish to quarrel: There are certain activities in which human
beings engage for which representations are indispensable. And indeed for something to count as a
representation requires that it carry information, or at least that its function
is to do so and that it be a member of an articulated family of such
information bearers.
But these conditions are plainly
not sufficient for representation, as opposed to what Clark felicitously calls
“adaptive hookup” and with which he contrasts representation. His examples of adaptive hookup include the
processes that enable sunflowers to track the sun. Note that these states satisfy Clark’s set of putatively
necessary and sufficient conditions for representations. Without belabouring the point, or getting
further into Chisholming away at sets of necessary and sufficient conditions
for representation, we can get directly to the heart of the issue in this
way: What is missing from Clark’s
account of representation is the very notion at issue—that of standing
in. Carrying information and coordinating
behaviour is indeed a crucial function of many of our cognitive states and
processes, and indeed we can rank cognitive processes on some kind of scale of
sophistication reflecting the fineness of grain required of the
information-bearing states in question; the quantity of information required;
the complexity of the processing involved, etc. And it might make sense to
distinguish in this way between information-hungry and information-declining
processes in this way. But this does
not get at the distinction between hookup and representation. For all the
information in the world might not involve the standing in function.
In fact, Clark seems to be
investigating not the distinction between hookup and representation, but rather
grades of intentionality, and it is the distinction between intending
in this sense and representing that is at issue here. The states to which Clark adverts, right
from the merely hooking-up states of the sunflower to the complicated
articulated states of a tennis-ball seeking border collie, and on into those of
a human strolling through the art museum, intend their objects—the
position of the sun; the ball that was just thrown, Monet and his role in world
history. But none need stand in for
them. None need re-present them.
They need to coordinate increasingly sophisticated interaction, and do so in
concert with increasingly complex families of states and processes. But that only gets us to higher degrees of
intentionality. None of it constitutes
representation, though once we hit the most sophisticated of these embodiments
of intentionality, that of the art aficionado, for instance, it is clear that
this intentionality presupposes a background of representation, a backgrounds
to the conditions of the possibility of which we now turn.
6. Real representations: Institutions, Innovations and
Language
The truth is almost
embarrassingly obvious, that it is human natural language that sets us apart
from other species, because it gives us, and only us, for the first time, a
genuinely open-ended compositional, manipulable, medium of representation
suitable for any topic. [Dennett in Clapin 2002, p. 191 ]
Does this mean that we are
approaching a nihilistic conclusion regarding representation? Not at all.
We simply have been looking in the wrong place. Let us recall the lesson I suggested we
learn from Wittgenstein: Representation is essentially a public,
convention-governed phenomenon. For
something to stand in for another—in any sense—requires that we have a set of
rules for instituting and using the proxy.
Just as I can stand in for the dean if I am appropriately deputized, and
a cheque can stand in for a pile of money given appropriate banking
institutions, a photograph can stand in for the grandchildren if we have a
convention for reading visual information from a flat surface, and words can
stand in for the picture if we speak the same language. (It might be thought that one of the central
morals of the first half of Philosophical Investigations is precisely that,
contra the Tractatus, the language is not representational. This would be a serious exegetical error—and
one that would impute a serious philosophical error to Wittgenstein. In fact the point one might so misrepresent
is that representation is one of the many functions of language, and
presupposes, but is not presupposed by the more general practice of
rule-following.)
The alternative approach is
that we have seen taken by Clark, to take representation as a matter of bearing
information in virtue of reliable covariation, and doing so in virtue of having
that function (however that function is determined). So the neurons that bear information regarding the orientation of
the rat’s head, that provide that information to motor control systems, and
that have precisely that function qualify in that sense as
representations. Here is what is wrong
with that approach: First, such systems incorporate no stable,
context-independent recurrent states that would naturally be identified as
tokens of representations. That’s not a
big problem, as one might simply
begin a story about representation in dynamic systems. But it does begin to strain the
applicability of the metaphor. Second,
and more significantly, we can note that defining representations this generously gives representational status to
the states of sunflowers, thermostats, and so forth. We should worry that we are about to lose the ability to draw
distinctions that need to be drawn if we are to isolate the natural kinds of
use to cognitive science.
Third, and now most
importantly, we can point to the nature of that relevant distinction. We are after the distinction between intentionality
and representation. It is certainly the case that the rat neurons to which
Clark adverts, as well as the relevant structures of sunflowers are about that
about which the carry information; that is to say, they are intentional. But
to be intentional is not sufficient to represent. Representation is important: it is what makes puts the sapiens
in Homo sapiens. Intention
is indeed a necessary condition of getting one ting to stand for another, but
to get from intention to representation we need to go through convention, and
thus through the enormous social intelligence for which we, alone among
terrestrial species, are specialized.
Clark simply sets the bar too low, and misses the importance of this
distinction.
Representation is symbolic, and
the use of symbols is an innovation requiring norm-imposing and norm-enforcing
institutions that make it possible for rules to be followed and hence for
proxies to be employed. The pre-eminent
and indeed, for humans, omnipresent, system of such representation is, of
course, natural language—an innovation so critical to our species that we have
become specialized for it, breeding out those who cannot cope easily with it
and selecting for those with linguistic facility. The ubiquity of linguistic
representations and the ambient bath of norm-governed practices enabling and
enabled by it is so great that like water for the fish it is invisible to
us. We use linguistic representations
when we speak, when we read, when we listen, and when we think. And when we use language, just as when we
use pictures or cheques, we can allow symbols not only to intend, but to stand
in for that which is absent (though, to be sure, that is not all we
do with language). Here is Andy Clark
again:
What does public
language do? There is a common, easy answer, which though not incorrect, is
subtly misleading. The easy answer is
that language helps us to communicate ideas.
It lets other human beings profit from it from what we know, and it
enables us to profit from what they know.
This is surely true, and it locates one, major wellspring of our rather
unique kind of cognitive success.
However, the emphasis on language as a medium of communication tends to
blind us to a subtler but equally potent role: the role of language as a tool
that alters the nature of the computational tasks involved in various kinds of
problem solving.
…
Public language is in many ways
the ultimate artifact. Not only does it
confer on us added powers of communication; it also enables us to reshape a
variety of difficult but important tasks into formats better suited to the
basic computational capacities of the human brain. Just as scissors enable us to exploit our basic manipulative
capacities to fulfill new ends, language enables us to exploit our basic
cognitive capacities of pattern recognition and transformation in ways that reach
out to a new behavioral and intellectual horizons. … Finally, the sheer
intimacy of the relations between human thought and the tools of public
language bequeaths an interesting puzzle.
For in this case, especially, it is a delicate matter to determine where
the user ends and the tool begins! [Clark 1997, p 194]
Public language is indeed the
ultimate artifact. And indeed its great
importance is not is power to convey that which we could already
think. There is, I am afraid, not much
of that worth conveying—though there is some, to be sure. The great value of the ultimate artifact—as
Dennett puts it, that which sets us off from all other species—is that it makes
possible representation in the full-blooded and fully versatile sense. And this makes possible the human thought we
know, love, and really do wish to convey.
We do not simply coordinate our behaviour with our fellows or with our
non-human environment; we acquire the ability to manipulate in thought, that
is, to reason about, to imagine, to plan regarding, things that are not present
to us; to categorize and re-categorize using labels, that is, to conceptualize,
both the concrete and the abstract in our environment. This is the true dawn of representation. And here we see the intuition that there is
something very special about representation vindicated. The ability to stand in for is more than the
ability to mediate purposive behavior.
Ignoring that function in favour of mere intentionality is perilous to
the philosophy of mind.
But what does this mean regarding
the idea of mental representation?
Here is what I think it means:
There is none, really. We do
represent, and we use our minds to do it, but only in a derivative sense. The representational burden is carried by
the tokens of our public language. They
stand in for their denotata when we represent. What our mental processes are
good for is intending. But with the
development of the institution of language, we acquire the ability to intend
any of a vast number of articulate representations. Mental intentionality plus linguistic representation equals human
thought.
This picture is buttressed by
important evidence from developmental psycholinguistics, particularly when
those data are viewed from a broadly Vygotskyan perspective. The principal data I have in mind come from
research into the acquisition of theory of mind, and in particular regarding
the role of pretence in that process.
But before I turn briefly to those data, let me note that Clark’s
intuitions converge. Reflecting on recent
work by Berk and his colleagues, he writes:
The
Vygotskian image is supported by more recent bodies of development
research. Berk and Gavin (1984)
observed and recorded the ongoing speech of a group of children between the
ages of 5 and 10 years. They found that
most of the children’s private speech (speech not addresses to some other
listener) seems keyed to the direction and control of the child’s own actions,
and that the incidence of such speech increased when the child was alone and
trying to perform some difficult task.
In subsequent studies (Bivens and Berk 1990; Berk 1994) it was found
that the children who made the greatest numbers of self-directed comments were
the ones who subsequently mastered the tasks best. Berk concluded, from these and other studies, that self-directed
speech (be it vocal or silent inner rehearsal) is a crucial cognitive tool that
allows us to highlight the most puzzling features of new situations and to
better direct and control our own problem-solving actions.
[Clark
1997, p 195]
Data from our own laboratories,
those of Thomasello and his colleagues concerning social development and the
findings of the de Villiers and Senghas and Pyers corroborates this intuition
in a dramatic way: The de Villiers
demonstrated dramatically that the development of competence in reasoning about
mentalistic states immediately follows and requires mastery of the syntax and
semantics of the sentential complement constructions and the verbs that govern
them. Without this linguistic competence,
it is impossible for children to reason about psychological states. Senghas and Pyers have shown that this
inability to pass theory of mind tasks without mastering the language necessary
for representing those phenomena. Our
own research has shown that pretence plays a special role in coming to be able
to reason about mental states: children master verbs of pretence and the
ability to reason about pretence prior to attaining corresponding mastery of
the language of the mental and of the ability to reason about the mental.
This pattern—the learning the
relevant language preceding the ability to perform the reasoning; the
precedence of mastery of reasoning regarding joint pretence over that regarding
inner episodes—confirm the Vygotskyan intuition that representation is
initially a public activity and that as the public representational medium of
language is learned, it can scaffold autonomous thought and reasoning. The most plausible mechanism of this scaffolding
is that our ability to intend is turned upon this representational medium,
allowing us to make autonomous use of language. The representational weight is born by language; the ability to
make use of that power in thought requires our complex cognitive ability to
intend—to engage cognitively with our environment. When that environment includes symbols, the magic of human
thought occurs. Clark puts the point
this way:
Experience
with external tags and labels thus enables the brain itself, by representing
these tags and labels, to solve problems whose level of complexity and
abstraction would otherwise leave us baffles-an intuitive result whose
widespread applicability to human reason is increasingly evident.
Learning a set of tags and labels (which we all do when we learn a language) is,
we may thus speculate, rather closely akin to acquiring a new perceptual
modality. … And of course the whole process is deeply iterative—we coin new words
and labels to concretize regularities that we could only originally
conceptualize thanks to a backdrop of other words and labels. [Clark in Clapin 2002, p. 42 ]
7. The abstract, the distant and the nonexistent, but also
the general, the motley and the meta
Discussing with approval
Dennett’s view of the role of language in representation in Hugh Clapin’
excellent conference proceedings on mental representation—Clark edges closer to
this position:
Florid representing occurs, recall, when there is a knowing use of representations, where representations are (at least) some kind of maipulanda: objects that bear contents and that can be somehow shuffled, reorganized, and recombined in ways sensitive to, and exploitative of, those contents). … The objects (the manipulanda) involved in florid representing bear the contents they do only in virtue of a bedrock of skills and capacities, rooted in multiple non-propositional mind-tools. But florid representing depends on making those skill-based contents into objects suitable for the exercise of other (non-propositional) skills—skills of combining, shuffling, and so on. And it is this ‘objectification’ of certain aspects of content that supports the highly versatile and open-ended range of thought characteristics of (and perhaps uniquely characteristic of) human understanding. Finally, … it is our experiences with public symbols that are said to teach us to make more manipulable objects of our thoughts and ideas [Clark in Clapin 2002, pp 81-82]
Here we see the recognition
that (1) representations are manipulanda—the objects of cognitive
operations, and not their constituents; (2) that their content depend upon
non-propositional skills; and (3) that public language plays a crucial role
both in generating human representational power and in generating the
possibility of thought about thought.
This is all to the good. All that is missing is the recognition of the
deep connection between (2) and (3)—that the basis on which representational
content is constructed comprises both the non-propositional intentional
capacities and skills of the organism and the representational capacity
of public language, joined in the ability to think about language, and
derivatively about that which it represents.
Does this claim about
derivative mental representation threaten to collapse the distinction between
intention and representation? One might
think so, since on this view when I think, for instance, that if I were now in
the northern hemisphere I would be in summer, I think about the northern
hemisphere and do so via a representation, and hence the relevant thought is
representational. And since there is
nothing of ontological significance to distinguish between the nature of this
thought and the states of my perceptual or motor control systems that I want to
point to as paradigms of the sub-representational intentional, the distinction
collapses. But this is not the end of
the story. The representation in this case is, crucially, not the vehicle of
thought, not a cognitive state, but rather the immediate object of
thought, a linguistic item.
Cognition intends that object, which in turn represents the
abstract fact in question. This
derivative representational character also allows us to understand how the
continuous processes of cognition can have discrete semantic content. Their continuity is resolved by the discrete
objects they intend. The kernel of
truth in the language of thought hypothesis is the intuition that representation
must have determinate, and indeed, compositional, content, and that only
language can provide that. It does not
follow, however, that thought is in language, only that it is of language.
It is a commonplace in
contemporary foundations of cognitive science that language is a necessary
condition of the representation of the non-existent, the abstract and the
distant. We can now reformulate this
point: representation is a necessary
condition of the intention of the non-existent, the abstract and the
distant. Without the capacity to
represent, one simply can’t think about these things at all. But once one can represent, the sky isn’t
even the limit.
To this trio, we can add
another trio not often noticed: the general, the motley and the meta. Dennett notes the point about generality:
Animals
are not only capable of believing general propositions in sensu composito.
That is, consider that case of learning that all orange mushrooms are
toxic. What the dog can learn, perhaps learning from training from its mother, is the disposition, whenever
you see an orange mushroom, to shun it. That is, the disposition to
acquire a particular 'this is toxic' belief, whenever encountering a particular
orange mushroom. [Dennett
in Clapin 2002, p.
50]
An infralingual creature can
surely reliably intend each member of some class on each occasion, and even be
said counterfactually to do so. In one
sense this might count as having a general attitude. But, as Dennett points out, this is a kind of amphiboly: it is instead generally true of such an
animal that it intends each member in the same way. Generality emerges with quantification, and quantification is a
linguistic device.
We can believe things of
motley collections, for instance that each item either on my desk or once owned
by GE Moore is beautiful. This power
derives specifically from the representational power of language, and outstrips
any power of mere intending. It is
central to the full generality of thought.
Finally, we can think about thought, both in virtue of intending our
thoughts and in virtue of representing them.
Thinking about thought requires the kind of iteration of intentionality that is made possible only by
representation, precisely because it is only representation that gives us a
description of thought as intentional, and so makes it possible for thoughts qua
thoughts to be objects of thought.
To be a human thinker is
precisely to be able to think in full generality, about just anything, and to
reflect on that capacity. The gap
between regular adaptive coupling with the local environment and this capacity
is precisely the capacity between intentionality and representation that is
made possible by language.
8. Not just a notational variant!
It may be tempting at this point to reply to all of this that I have only rehashed the difference between linguistic and non-linguistic representation—that my term “intending” is nothing but a notational variant of “non-linguistic representation.” This would be a grave error. I am neither after the point that language makes mental representation more complex and powerful than non-lin