Wittgenstein, Semantics and Connectionism: Laurence Goldstein and Hartley Slater, University of Hong Kong
Wittgenstein, Semantics and Connectionism: Laurence Goldstein and Hartley Slater, University of Hong Kong
Wittgenstein, Semantics and Connectionism: Laurence Goldstein and Hartley Slater, University of Hong Kong
ISSN 0190-0536
294
Philosophical Investigations
conceiving the mind and its operations is what we shall call the
Symbolic Paradigm in Cognitive Science.3
It is clear, even from this brief sketch, that the Symbolic Paradigm
carries in its train certain commitments in the philosophy of language. The key notion is that of a mental representation. There are
stored meanings, and there are mental sentences which represent
states of affairs. Thus Ray Jackendoff, in defending a computational
theory, writes: . . . meanings are mentally represented . . . meanings
must be finitely representable and stored in a brain,4 and Jerry
Fodor, a prominent champion of the Symbolic Paradigm, contends
that we could not account for the fact that we think new thoughts
(productivity), and that there are systematic relations between the
thoughts we are able to think (systematicity) unless we treat
thoughts as composed of sentential representations which are
variously manipulated in the process of thinking.5
One can easily see what makes the traditional view and the modern computer implementation of it seem plausible. For clearly there
is a world of difference between a human being asserting Tinas
father is a bus driver and a parrots uttering exactly the same sentence. Since externally things may be much the same, the difference,
so it must seem, is internal, something going on inside the human
being, but not inside the parrot internally we humans perform an
act of meaning. How is that to be construed? Well, it seems there
must be a mental counterpart to the overt sentence, and this
connects to the world by representing it in some sort of way. By
295
contrast, the parrot does not have any mental representation of any
part of the world in mind when it makes its utterance.
If meaning involves, in some way, mental sentences related, somehow, to the world, then it is rather natural to posit a tight
connection between meaning and truth. In a widely accepted view,
the meaning of a sentence is identified with its truth conditions.
Here, the notion of truth plays a pivotal role in the explanation of
meaning. The origin of such an account is a formal theory of truth,
due to Tarski, in which one structured entity, a sentence, is said to
be true when it stands in a certain relation (that of being satisfied by)
to another structured entity, an ordering of objects and sets of
ordered n-tuples of objects. Tarski regarded this truth theory as capturing (and making more precise) the traditional conception of truth
in which a sentence is said to correspond to an ordered arrangement
of objects in the world sometimes called a fact, sometimes a state
of affairs.6
We have, then, within the Symbolic Paradigm, a conception of
the mind as a processor of representations, and associated views on
meaning and truth. This whole web of ideas becomes fragile if any
part of it is dislodged. Anyone with even a passing acquaintance with
Wittgensteins later writings will know that the constellation of
views sketched above was anathema to him. He vigorously
inveighed against the idea that meanings are objects (or any mental
state, process or relation) and ferociously rejected the notion that
thinking is the activity of doing calculations upon mental sentences.7
Clearly, Wittgensteins arguments, if they are sound, sweep away the
foundations on which the Symbolic Paradigm rests.8
Connectionism may be roughly characterized as a computational
theory, but one which, by contrast with the Symbolic Paradigm,
regards the mind not as a calculating device which manipulates data
structures but as more like a sensory-motor processor which is taught
to behave in certain specific ways only through engagement with the
6. For a criticism of Tarski on this score, see Donald Davidson, The Folly of
Trying to Define Truth, unpublished manuscript.
7. See, for example, L. Wittgenstein, Zettel (Oxford: Blackwell, 1967), 605612.
The ferocity of Wittgensteins attack is quite likely attributable to a vehement antiscientism that we (the authors) do not share.
8. The point has been noted, briefly, by other writers, see, for instance, Philosophy
and Cognitive Science by James H. Fetzer (New York: Paragon House, 1991), 75.
Fetzer traces his own ideas back to Peirce, but the relevant ones clearly bear a striking likeness to those of Wittgenstein, as Fetzer acknowledges.
Blackwell Publishers Ltd. 1998
296
Philosophical Investigations
9
II. Truth
There are two points to be made straight away about Wittgensteins
notion of truth: first that it is not Tarskian, and second that it is not a
correspondence theory as traditionally conceived. Indeed, in as
much as anything like correspondence is involved, this is a matter
of the resemblance between one situation and another in the world
not a relation between words and some situation in the world.
What one must remember is what Wittgenstein had to say about
standard objects. The standard meter in Paris was one of his examples. By an act of decision, this object was determined to serve as the
paradigm of meterhood, and we could imagine samples of colour
being kept, hermetically sealed, in Paris the standard sepia, for
example. This sample is an instrument of the language used in
ascriptions of colour.10 The use of other predicative terms likewise
depends on there being samples which serve as standards.11
But if the application of predicates rests on the prior acceptance of
standard samples, then recognising of some other object that it is a
meter long or that it is true that it is a meter long is recognising
that there is a relation between this object and the relevant paradigm
297
12
a relation akin to resemblance. What is true is then the objective situation, in the sense that the situation is true to type, i.e. to the typical
situation, and the continuity of the sense of the term true with that
in true hyacinth, true conservative, for example, becomes evident.
Moreover, it is not the sentence used, It is a meter long which is
then, primarily, what is assessed as true, although no doubt by
metonymy it may be called true. The sentence is the medium by the
use of which truth, i.e. the closeness to the paradigm, is expressed.
Wittgensteins point about paradigms is even provable, using the
refined logic of predicates as set out in Hilberts epsilon calculus.13
For even in the predicate calculus it can be proven that:
(Ex)((Ey)Py . Px)
and contrariwise
If Aristedes is corruptible, everyone is.
It immediately follows that such paradigm objects are what hold the
concept fast in all possible worlds, and all human minds.14 Thus if
someone does not recognise that paradigms of justice are just, then
they do not have the concept of justice and can make no proper
judgements which employ it. And this goes for all concepts, since
the predicate calculus thesis is a quite general thesis.15 Having the
12. This is not to say that recognising the appropriate resemblances is always easy.
For a discussion of some of the difficulties about extending the use of colour words
from the initial examples see M. Dummett, Origins of Analytical Philosophy (London:
Duckworth, 1993), 9091.
13. B. H. Slater, The Epsilon Calculus and its Applications, Grazer Philosophische
Studien 41 (1992), 175205.
14. B. H. Slater, Descriptive Opacity, Philosophical Studies 66 (1992), 167181.
15. Further arguments for this general thesis, within cognitive science and developmental psychology, are to be found in discussions of what are sometimes called
prototypes. See, for instance, Hubert Dreyfus, From Micro-Worlds to Knowledge
Representation in J. Haugeland (ed.) Mind Design (Cambridge MA, M.I.T. Press,
1988), p.185, and Eleanor Rosch, Human Categorization in N. Warren (ed.),
Advances in Cross Cultural Psychology Vol 1 (London, Academic Press, 1977), p.30.
Blackwell Publishers Ltd. 1998
298
Philosophical Investigations
e.g. It is true that snow is white, and this shift to a non-metalinguistic expression is the prime move which takes us away from
the symbolic account, involving the presumed uniform language of
thought Mentalese.17 Prior writes:
16. The importance of the cultural dimension is stressed by R. McDonough,
Linguistic Creativity, in Linguistics and Philosophy: The Controversial Interface, ed. R.
Harr and R. Harris (Oxford: Pergamon Press, 1993), 125162.
17. Although it is the Tarskian account which has come to be regarded as traditional, it is interesting to note that Ramsey, in his classic essay Facts and
Propositions published five years before Tarski treats It is true that as an operator expression (see, for example, Foundations, ed. D. H. Mellor (London: Routledge
and Kegan Paul, 1978) p.44).
Blackwell Publishers Ltd. 1998
299
The truth and falsehood with which Tarski is concerned are genuine properties of genuine objects, namely sentences. The truth
and falsehood with which we have been concerned here might be
described as properties not of sentences but of propositions; but
this means that they are only quasi-properties of quasi-objects, and
it might be less misleading to say that we have not been concerned
with the adjectives true and false at all but rather with the
adverbs truly and falsely. The basic form which Tarski defines is
The sentence S is a true one; the form which we define is not
this, but rather X says truly (thinks correctly, fears with justification) that p.
The truth is that snow is white; the justification for saying this is that
snow is or resembles paradigmatically white things. But the justification of a claim must be distinguished from its meaning. It is true
that snow is white means no more than that snow is white,19 and
snows being white entails nothing about language.
18. A. Prior, Objects of Thought (Oxford: Oxford University Press, 1971), pp.9899.
19. Wittgensteins espousal of this deflationary view predates Ramseys. For example, in the Notebooks 191416, p.9, entry for 6.10.14., he writes p is true, says
nothing else but p and it is clear that he is here not using the quotation marks as a
device for quoting sentences. In the Philosophical Grammar, where he repeats the
deflationary formula, he insists that the sign p is a propositional sign, not the name
of the shape of a particular ink mark. He says In the end one can say that the quotation marks in the sentence p is true are simply superfluous (Philosophical
Grammar, p.124; see also PI 136). Paul Horwich relates this view of truth to
Wittgensteins views on meaning in Meaning, Use and Truth, Mind 104 (1995),
pp.355368.
Blackwell Publishers Ltd. 1998
300
Philosophical Investigations
III. Meaning
Paul Horwich (see note 19) has argued that minimalism with respect
to truth is just what is required for a defence of the claim (made by
Wittgenstein at PI 43 and elsewhere) that the meaning of a word is
its use in the language. An understanding of Priors non-metalinguistic operator expression allows us to see the link. The metalinguistic form p is true is syntactic a meaningful sentence is
returned if we substitute for p a sentence of any language, although
the meaning of that operand is not given. But a legitimate substitution instance of the non-meta-linguistic form It is true that p
incorporates the meaning of whichever sentence substitutes for p,
since it is not that sentence but what it means which is being spoken
about. However, the sense in which it is spoken about must be
clarified, since the operator It is true that is not a predicate which is
completed with a name to yield a whole sentence. Instead it is completed by a sentence to yield a further sentence. And sentences are not
names, for they do not designate their meanings, but express them.
Hence truth as characterized by the operator expression is not a
property, with the result that there is nothing that it is a property of;
it has no bearers.20 What is going on in Priors expression, in contrast with Tarskis meta-linguistic, predicative expression, therefore is
just the process of expressing the meaning of p simply by using
p appropriately. So we come to see that it is purely the appropriate
using of words which not only gives them their meaning, but is, i.e.
entirely constitutes their meaning.
In fact it was J. L. Austin who formulated most clearly this point
about meaning.21 Meaning ones words is a speech act which is rudimentary in the sense that meaning what one says which is just
performing a locutionary act is a prerequisite for giving ones words
an illocutionary force or producing a perlocutionary effect. Meaning
20. See Dorothy Grover, A Prosentential Theory of Truth (Princeton: Princeton
University Press, 1992), 2224, 132133, 175178; Susan Haack, Philosophy of Logics
(Cambridge: Cambridge University Press, 1978), 127134.
21. J. L. Austin, How to do things with Words (Oxford: Oxford University Press,
1952), 92f. Fodor thinks of the meaning of mentalese symbols as being given by their
causal origins, but this is to ignore the social choices involved in making cows mean
cows, etc. We are trained, in English speaking countries, to make the conventional
associations which give English words their meanings. But no such training could be
involved with symbols of Mentalese, since, as is made plain in the text, there is nothing we can do with such words.
Blackwell Publishers Ltd. 1998
301
some words is using them with the appropriate sense and reference,
which is indeed something someone is automatically taken to be
doing if they are uttered in some communal arena. There are further
dimensions of use involved in illocutionary force, and even perlocutionary effect: but it is only the locutionary act which properly
constitutes meaning. Notice that, since meaning ones words is an
act, it is thereby immediately not an object. Indeed the noun meaning is a gerund. Ones inscriptions and coups de bouche are objects,
but it is the appropriate use we make of those objects that constitutes
the meaning of them.
What, therefore, distinguishes the aforementioned parrot from a
human speaker is not its lacking internal words, but rather its
inability to complete the full locutionary act, i.e. to make proper use
of its words as opposed to merely uttering sounds. It does not use its
words in customary ways, does not rely on public instruments such
as dictionaries to justify its usages, does not use its words relevantly
in connection with public objects, particularly paradigms. And this is
where the key argument against Mentalese arises. For, if there were
some objects installed in the parrots head which accompanied all its
words or phrases, then, as Wittgenstein observed, these
Doppelgngers would need to be used appropriately if they are to
have any semantic significance. And the point is not just that, unlike
with overt inscriptions and coups de bouche, any use we make of them
cannot be publicly checked, so the standard private language argument eliminates them. For, in a proper sense, such private objects
cannot be deployed or put to use, since they are not instruments or
tools: we cannot do things with such words.
One can imagine someone might say:
But symbols in a language of thought are used, in exactly the way
that symbols in a computer are used. They are used by the computational processes and programs which manipulate those
symbols. Symbols in the head are not instruments or tools used
by whole human agents, but then we dont want them to be.
Words, that is public ones, are syntactic objects in need of an
interpretation, but, as Wittgenstein said, interpretation has to stop
somewhere, otherwise we fall foul of an infinite regress.
Mentalese is the point where interpretation stops. Thats why it
explains how public words are used meaningfully by whole agents.
There are just causal processes inside the agents brain which are
organised, as in a computer, to preserve semantic relations. That is
just another version of the standard demand that all homunculi be
discharged.
Blackwell Publishers Ltd. 1998
302
Philosophical Investigations
But symbols in a computer are not used, any more than a gate uses
a latch to hold itself shut if it blows shut. The symbols in a computer
are maybe used by a programmer, via some software which calculates with them. But that is derived intentionality, and so use of the
word calculates with the software is metonymic.
More importantly, there isnt a pre-existing semantics for a is F,
for instance, which is there prior to the human, public use of such a
sentence, and which would therefore cause the behaviour. For the
meaning of a word is a matter of convention: if one sees how is F is
used that might determine a set of objects, but there isnt that set of
objects, delimited somehow by proxy in our brains, before the public
use. For, for one thing, that use is a communal use, and so not a
matter of one brain and there may be arguments. That there is no
pre-set meaning is just the force of Wittgensteins anti-Platonism in
this area. Maybe something in the environment, at each application
of is a game, has prompted the decision one way or another, but
there need be nothing in anyones brain which does this, since every
application of every word is arbitrary. Semantics is not compositional: there is no given mapping from a is a game to a and some
property of being a game or a and some set of all games since
there is nothing before our agreed use which settles how we are to
use is a game, and so what its meaning is. You might as well ask
what the name of a baby is before you make your choice of a name
for it.
IV. Thought
We can see now how the parrots impotence relates to the Cartesian
inner theatre. That ghostly stage is supposedly peopled by mental
words, perhaps redolent with images. But there is nothing which
could put these items in touch with the real world of social linguistic
intercourse. This is just what makes the theatre seemingly inner.
The mental words Tinas father is a bus driver may seem to dance
in that order on the Cartesian stage; some images may join in the
play. But there is a great difference between this and an agents actually thinking that Tinas father is a bus driver. The latter kind of
thought is not at all Cartesian, since it requires the thinker to be
located in certain ways to Tinas father and also have the concept
of a bus driver, i.e. be able to identify correctly standard and central
Blackwell Publishers Ltd. 1998
303
304
Philosophical Investigations
305
easy to see. For a belief must have a believer that accounts for one
of the relata and to individuate beliefs, for example to distinguish
between As belief that Nixon was foolish and Bs belief that Clinton
is wise, it seems that we must advert to the differences between the
sentence in which beliefs are expressed. Hence, so it seems, the other
relatum has a linguistic structure. But let us, as before, look at the
beliefs themselves rather than at the means people use to express their
beliefs. Believing is a state of mind. It has duration; and that independently of the duration of its expression in a sentence, for example (PI
p.191). To believe that p is to have the settled disposition to act as if it
is the case that p, i.e. to have the disposition to engage in acts that
conform to the agents supposition that p is the case. Now the activities associated with a belief will be physically diverse, even if they are
functionally similar in that, with all other factors held constant, the
belief that p will lead different agents to much the same behaviour, or
dispositions to behaviour. But there is nothing linguistic in the disposition to engage in such physically diverse but functionally similar
bouts of behaviour.27 So why should one suppose that there is linguistic structure in the having of a belief? Indeed, while there is believing
that p, there is no believing p.28
Perhaps the most decisive way to undermine the Symbolic
Paradigm is to look at some attitudes classically conceived as propositional and prove that they can have no linguistic propositional
content. We take the following example from a paper by Michael
McKinsey. Consider
[1] Oscar assumes that just one fish got away, and Oscar wishes it
had been the case that he caught it (that very fish).
McKinsey writes:
Now suppose Oscar is right that just one fish got away, at time t,
say. Let us call this fish Bubbles. Then it is clear what proposition
Oscar would be wishing true. It is the proposition that is true at a
possible world w just in case in w, Oscar catches Bubbles at t. But
now suppose that Oscar is wrong. Perhaps what Oscar thought
was a fish on the end of his line was really just an old boot or an
underwater branch. Then it seems quite impossible to specify the
27. For a suggestion as to how our neural pathways are shaped by learning so that
our beliefs are generally reliable indicators of what is the case, see D. Papineau,
Teleology and Mental States, Proceedings of the Aristotelian Society 65 (1991): 3354;
P. Godfrey-Smith, Signal, Decision, Action, Journal of Philosophy 88 (1991):
709722.
28. P. Churchland, Neurophilosophy (Cambridge MA: MIT Press, 1986), 386399.
Blackwell Publishers Ltd. 1998
306
Philosophical Investigations
proposition that Oscar wishes had been the case . . . Thus it is possible to have a particular wish, without there being any
proposition that one is wishing true.29
While endorsing a modified conclusion that there may be a particular wish without its specific content being linguistically expressible,
we cannot agree with McKinsey that the content of Oscars wish is
completely inexpressible when it is an old boot, at the end of his
fishing line. A more realistic theory would be that, in that case the
content is expressible, but not just linguistically.30 The pragmatic
context must also be given. Indeed, use of epsilon terms supports
this result very firmly, since whether the descriptive assumption
There is a fish that got away is true or false, the epsilon term ex(x
is a fish that got away) still functions referringly (see notes 12, 13).
[1] has the form:
Ao(E!x)Fx.WoCxFx
where xFx is the fish which got away. But the fish that got away
need not be a fish, any more than Dartmouth need be at the mouth
of the Dart, or The Morning Star be a star. It therefore brings no
definite concept of the object into what is wished for, or assumed
about. The fish that got away, used referringly, functions in a pragmatic context merely to call the pure referent itself to the mind, and
so renders otiose all senses (i.e. representations) of it. The wish is
about that object, whatever it is that we call it, i.e. independent of any
linguistic description. For the epsilon term symbolises de re attitudes
and direct reference.31
29. M. McKinsey, The Internal Basis of Meaning, Pacific Philosophical Quarterly 72
(1991): 143169, see 161. A similar position is R. J. Nelsons connectionist theory of
reference, in Naming and Reference: The Link of Word to Object (London: Routledge,
1992).
30. This example, in other words, supports a recent theory of attitude ascriptions
which proposes that a belief is about an individual if and only if there is an individual which is the source of that belief. See R. Holton, Attitude Ascriptions and
Intermediate Scope, Mind 103 (1994): 123130. An unseen old boot may be the
source of ones belief that one nearly caught a fish, and that belief is then entirely
about that boot. The point relies on Donnellans distinction between reference and
attribution (see Reference and Definite Descriptions, The Philosophical Review 75
(1966), 281304). See also, for instance, Ruth Barcan Marcus, Modalities (Oxford:
Oxford University Press, 1993) Ch. 14.
31. Some of the consequences of there being de re beliefs are examined in Andrew
Woodfield, Thought and Object (Oxford: Oxford University Press, 1982), viii. In particular he shows that the mind-brain identity thesis is false, and that the mind is the
brain-environmental complex.
Blackwell Publishers Ltd. 1998
307
So strong is the grip of the relational account of propositional attitudes that even a reader sympathetic to our objections to that
account may earlier have felt that there was no remotely plausible
alternative to the relational view. But here, with formalised de re
beliefs, we come to recognize the reasonableness of saying that there
may be no difference in brain condition between a person who
believes that p and someone who doesnt. Suppose, for example,
two, more or less identical, twins enter (blindfold) into two otherwise identical white rooms: twin one subsequently believes the first
room is white, twin two that the second room is white. But the difference between them resides entirely in their spatial locations, not
in their brain states, hence there can be no representations in those
brains which represent the different beliefs. Receiving an impression
of ones surroundings is thus not too much like getting a blob of ink
smeared onto one,32 since the surroundings are not absorbed in any
appropriate way. Yet it is more like relating to an ink blot than to a
piece of language,33 because of its non-semantic, linguistically
unstructured nature. And the same goes for connectionist systems.34
Mark Richard writes: For Maggie to think that Odile is tired, she
must have some representations of Odile and of being tired put
together in an appropriate way. In some broad sense of sentence,
she must employ a mental sentence saying that Odile is tired.35 This
is exactly the position we have attempted to repudiate. In order to
think Odile is tired, Maggie must, of course, have the appropriate
words, but her thinking that Odile is tired cannot be split into wordlike parts. Her thinking that Odile is tired is related to an external
object in a certain way. It consists in her taking there to be close
32. This idea is suggested in David Bohm and Basil Hiley, The Undivided Universe:
An Ontological Interpretation of Quantum Theory (London: Routledge, 1993).
33. Two good accounts of connectionist alternatives to sentential mental representation are Tim van Gelder, What is the D in PDP? A Survey of the Concept of
Distribution, in William Ramsey, Stephen Stich and David Rumelhart (eds.),
Philosophy and Connectionist Theory (Hillsdale: Lawrence Erlbaum, 1991), 3359 and
Terence Horgan and John Tienson, Structured Representations in Connectionist
Systems? in Steven Davis (ed.), Connectionism: Theory and Practice (Oxford: Oxford
University Press, 1992), 195228.
34. Further objections to the sentence-processing view are marshalled in William
S. Robinson, Computers, Minds and Robots (Philadelphia, Temple University Press,
1994), p.154 ff. The last three chapters of this book constitute an excellent, straightforward sustained but restrained defence of the Connectionist approach.
35. Mark Richard, Propositional Attitudes: An Essay on Thoughts and how we Ascribe
them (Cambridge: Cambridge University Press, 1990), 2.
Blackwell Publishers Ltd. 1998
308
Philosophical Investigations
309
310
Philosophical Investigations
where Bap says that a believes that p. But the latter entailment is
defeasible, as is obvious from the fact that two people with a common pair of beliefs may not arrive at identical conclusions. Good
reasoners are logical and their inferential practices conform to rules
which can serve as norms for humanity in general (another example
of the paradigms previously mentioned). But bad reasoners, as a
group, deviate unsystematically from such norms, and their wayward
inferential moves therefore entirely resist neat classification. The
bulk of the population lies between these two extremes, of course,
fumbling along in the main, with occasional falls into pits of madness, but rising up, at times, to heights of clarity. This is not the kind
of behaviour that any self-respecting Turing machine would wish to
emulate.
311
312
Philosophical Investigations
of an apple (and may well not have one) so long as he has learned to
respond to apple by checking the labels on his drawers. His
response to red is quite different. Here there is a procedure: look
up that word in a table which matches such words with colour samples; then find the relevant objects (apples) whose colour matches
the sample. In order to comply with five, the shopkeeper says the
series of numbers (which he knows by heart) up to the word five
and, for each number, he takes an apple out of a drawer. The
numerals, that is to say, are the paradigms against which other sets of
things are judged to have this number or that. This ability to number off objects by reciting a series of sounds and accumulating a
collection by adding one new object for each succeeding sound
requires a considerable amount of training. Thus the uses of the
words five, red and apples are quite different from each other:
each word invokes (or is woven into, in the sense of PI 7) a different suite of actions. It would be an insult to the complexity of
human understanding of language, as well as being misleading,
therefore, to say that one understands each word by matching it to
its meaning this general notion of the meaning of a word surrounds the workings of language with a haze which makes clear
vision impossible (PI 5). To disperse this fog, one needs to avoid
talking about the meanings of words and instead look carefully at
how different words are used (PI 1).
Connectionist machines hold out the promise of being able to
develop such varied dispositions, whereas a symbolic system can
involve no such know-how. Take, for one last example, the indexical word mama. In its use this does not refer to a general concept,
but to the specific mother of the user. Its place in the pragmatic content is therefore paradigmatically apparent. Now after some time, the
baby who is learning to speak reaches a relatively stable state in its
use of the word, and the associated neural organization readies the
child for significant new cognitive achievements.41 But there is no
41. See A. Clark and A. Karmiloff-Smith, The Cognizers Innards: A Psychological and Philosophical Perspective on the Development of Thought, Mind and
Language 8 (1993): 487519. William Bechtel, in The Case for Connectionism,
Philosophical Studies 71 (1993): 119154, shows that these phenomena are to be
explained by a systems adapting to the use of external symbols (natural language),
not by its manipulating internal symbols. Indeed, he shows that learning a language
itself promotes productivity and systematicity in our thinking. On this point too, see
Neil Tennant and Florian von Schilcher, Philosophy, Evolution and Human Nature
(London: Routledge, 1984), 204250.
Blackwell Publishers Ltd. 1998
313
VII. Conclusion
If our case for a social-anthropological conception of the nature of
language has been made out, what are the consequences for
Cognitive Science, and specifically for the debate regarding
Connectionism? It is commonly said, in defence of the older
Symbolic Paradigm, that, given Churchs Thesis, neural networks
cannot be superior to Turing Machines as models of the mind
because they can perform no more calculations than can be achieved
by using a programmed, symbolic architecture. What this comparison presumes, we now see, is that the mind is essentially a calculating
device.42 But it is not: it is a sensory-motor processor which has
been taught to behave in certain specific ways only through engagement with the contingencies of this changing world. As a result, one
would expect it to be somewhat errant, erratic and error-prone, and
neural networks thus provide a more sympathetic model of the difficulties humans experience with being rational and logical. The
processors which have been constructed within the neural network
paradigm, neither have nor acquire a program, and they acquire
their talents entirely by external correction of their responses to contingent stimuli. Certainly such machines can be trained to do
calculations and will naturally extend their abilities beyond the
core of exercises from which they acquired their competence. But
42. Further arguments against the conception of mind as a calculating device are set
out in Ruth Garrett Millikan, White Queen Psychology and other Essays for Alice
(Cambridge MA: MIT Press, 1993), esp. chap. 8. In a fascinating study of Turings
little-known work on unorganized machines, Diane Proudfoot and B. Jack
Copeland show that one of Turings own most important but least appreciated
achievements was to provide cognitive science with the conceptual resources for
understanding how the mindbrain could fail to be equivalent to a Turing Machine.
See their Turing, Wittgenstein and the Science of Mind, Australasian Journal of
Philosophy 72 (1994), pp. 497517.
Blackwell Publishers Ltd. 1998
314
Philosophical Investigations
calculation is just one of their skills, just as it is just one of ours. The
Symbolic Paradigm ignores the symbiotic relationship between
minds, words and the social word; Connectionism makes the connection.
Dept. of Philosophy
University of Hong Kong
Pokfulam Road
Hong Kong
Dept. of Philosophy
University of Western Australia
Nedlands
Western Australia 6907