CREATIVITY, THOUGHT AND REPRESENTATIONAL

REDESCRIPTION

TERRY DARTNALL •
Griffith University

1. Introduction
We talk about 'flashes of insight', 'inspirational moments' and 'sudden ideas'. We
say that ideas and solutions 'come to us' and 'pop into our heads'. We even talk
about divine inspiration and 'being spoken to by the Muse' , and on a more mundane
level say that we don't know what we believe until we have written it down.
We talk, too, about 'expressing our feelings' and being 'compelled' or 'driven'
to write or compose something. Verdi was so profoundly affected by the death of
Allesandro Manzoni that he felt a "heartfelt impulse, or rather necessity" to write
the Requiem in his honour. He was motivated by a passionate desire to honour
Manzoni, and expressed his feelings rather than abstract thoughts.
At least two things characterise these accounts. One is that we have very little
control of our creative processes on these occasions: the ideas or products just
'come to us'. (D. H. Lawrence said "Wait for the creative moment and trust it when
it comes.") I distinguish here between generating a product and evaluating it. The
product, once generated, has to be evaluated. It may be accepted or rejected or
changed. The overall goals may be adjusted and the cycle repeated. The generative
process itself may be coaxed and nurtured. But the generative process itself is not
under our conscious control.
The other thing that characterises these accounts is that the idea or product
is usually new in a very important way: there is no significant sense in which it
is a recombination of old elements or old ideas. There is no significant sense,
for instance, in which Verdi's Requiem was created out of previously existing
elements. Consequently there really is a sense in which creativity 'gets something
out of nothing'. If we dismiss this notion out of hand we will deny much of the
phenomena that, in studying creativity, we set out to explain.
Of course, we cannot literally get something out of nothing. But I wish to
resist the move which says that because creation ex nihilo is impossible, creativity
involves putting together things which already existed, so that all creativity is
combination or recombination. In their paper in this section, Wales and Thornton
quote Johnson-Laird as saying:
• I wish to thank Margaret Boden, Andy Clark and Aaron Sloman for helpful comments on an
earlier version of this paper.

43
T. Dartnall (ed.), Artificial Intelligence and Creativity, 43-62.
© 1994 Kluwer Academic Publishers.
44 TERRY DAKfNALL

First of all, like all mental processes, [creativity] starts from some given building
blocks. One cannot create out of nothing.
In his Introduction to the section on Creativity and Connectionism, Chris Thornton
says:
The starting point for any study of creativity is generally the observation that
creativity involves the creation of something new. This observation is qualified
with the observation that creation ex Ilihilo is a logical impossibility. (You can't
create something out of nothing.) Therefore creativity must be the creation
of something new out of things that already exist. But if this fact is logically
necessary then it seems to imply that combination theories of creativity (Le.
theories which propose that creativity involves combining old stuff to make
new stuff) must be essentially correct.
This idea (that creativity ex nihilo is impossible, so that all creativity is combi-
nation) is seductive, and is a principal reason why most theories of creativity are
combination theories. I think it involves two mistakes.
The first mistake is that no clear distinction is drawn between combination theo-
ries and recombination theories. Combination theories say that creativity involves
the combination of elements into wholes. Thus Poincare said that the creative pro-
cess is the subconscious combination of atomic ideas into molecular wholes: the
atoms have hooks on them, and hook together to form structures. Recombination
theories say that we begin with structures and reconfigure them. The most common
way in which we do this is to modify one structure in the light of another, which
is usually most successful when the structures are from distant domains. Koestler
called this the "bisociation of matrices"-today we call it 'analogical thinking'.
The second factor is that there is a trivial sense in which combination theories
are true. It is true (but trivially so) that Verdi's Requiem consists of notes, and it
is true (but trivially so) that Hamlet consists of letters. It is also trivially true that
these elements were put together by concatenative rules.
These two factors allow combination theory (now broadly construed to include
recombination theory) to vacillate between the true but trivial and the non-trivial
but false. It is trivially true, for instance, that Verdi's Requiem consists of notes, but
it is not true that Verdi started with a structure or structures-unlike Kekule when
he was trying to determine the structure of the benzene molecule, or Rutherford
when he compared the structure of the atom to the structure of the solar system.
I do not wish to buy into the very broad picture here, which involves the
Language of Thought hypothesis and the confusion between vehicle and content. I
limit myself to saying that we are able to produce manipulable structures that are
not adequately characterised as mere combinations of notes or letters. This is a vital
issue in our understanding of creativity, and in our understanding of intelligence
and mind, for it is related to the emergence of structured thought out of skills and
abilities. These things (structured thought on the one hand, skills and abilities on
the other) lie on opposite sides of what we might call "The Great Epistemological
 CREATIVITY. lHOUGHT AND REPRESENTATIONAL REDESCRIPTION 45
Divide". AI has been held up for a long time by the problem of commonsense
knowledge, and as Hubert Dreyfus is always telling US, l such knowledge is rooted
in skills and abilities. What is the child doing when she endlessly fills and empties
cups of water? The answer seems to be that she is accumulating thousands of
related skills, which will eventually blossom forth as structured knowledge, as
explicit beliefs that she can reflect upon and change.
The principal thesis of this paper, then, is that, at least sometimes, creativity
gets something out of nothing. It does this by transforming states that are in us, that
are responsible for our skills and abilities, that drive our behaviour, into structures
that are available to us, and that we can access and manipulate. 2 In this sense
creativity brings thoughts. ideas and other structures into existence. This is not
creating something out of nothing, but it is the next best thing.
I present this thesis as part of a general theory of creati vity that tries to account
for the phenomenon within a mechanistic framework. Within that framework the
question becomes 'How (if at all) can a machine go from something like 'connec-
tionist competence to structured symbolic thought?' I paint a broad picture that
locates the concept of creativity with respect to the concepts of thought, knowledge
and intentionality. The picture is outrageously simplified, but, to steal a phrase from
Dennett, 'such idealisation is the price we must sometimes pay for synoptic insight'
(Dennett, 1992).
The paper is structured as follows.
Section 2 glosses Margaret Boden's account of creativity, outlined in the previous
paper and worked out in detail in Boden (1990/92). My gloss is that a central and
significant type of creativity involves 'breaking out' of a ruleset by modifying it,
or tweaking it, using higher-level (heuristic) rules.
Section 3 suggests that this picture is incomplete in at least three ways. First, the
rulesets we break out of are sometimes part of our procedural knowledge, so that
we cannot access them as declarative structures. Second, there are problems about
the origins of declarative structures, and how they are grounded in the environment.
Finally, we have to explain why some cases of 'breaking out' are creative whilst
others are not.
Section 4 does some phylogenetic story-telling, which suggests some answers:
simple organisms have no knowledge in them; more complex creatures have knowl-
edge in them, but are unable to access it; people have knowledge in them, which
they can articulate as declarative structures. This suggests that one type of 'break-
ing out' involves articulating procedural knowledge as declarative knowledge, and
modifying it in appropriate ways. If this is correct, there is an intimate relationship
between being creative and having a mental life, for people would acquire mental
lives through discovering, and modifying, what they think and believe.

1 See especially Dreyfus, 1979.

2 The distinction between knowledge that is in us but not available to us, and knowledge that is
available to us is due to the developmental psychologist Annette Karmiloff·Smith. More on this later.
46 TERRY DARfNAIL

Section 5 turns to the empirical evidence. Annette Karmiloff-Smith's Repre-

sentational Redescription Hypothesis (RRH)3 is a theory of cognitive development
that tries to explain how the human mind goes beyond domain-specific constraints.
It maintains that we are endogeneously driven to redescribe our implicit procedural
knowledge as explicit declarative knowledge, and to continue to redescribe our
knowledge in increasingly abstract terms. Thus we are natively endowed with the
kind of creative ability suggested in Section 4. This empirical evidence is inde-
pendent of the earlier arguments and suggestions. Creativity therefore takes centre
stage in our efforts to understand human cognition, because we have grounds for
believing that it is in our natures (and may be unique to our natures) to be creative.
This package of creative ability, mental life and native endowment appears to be
more than a coincidence, and we might wonder whether there is a unifying factor
that we have missed. This hunch is spurred by the fact that we would be surprised
to find that a creative system could not, even in principle, have thoughts, or that a
system with a mental life could not, even in principle, be creative. If this is correct
then we have the following: (1) a system is creative if and only if it can redescribe,
(2) a system is creative if and only if it has thoughts; therefore (3) a system has
thoughts if and only if it can redescribe. Thought, creativity and redescription begin
to emerge as facets of the same thing. Can we throw light on this?
In Sections 6, 7 and 8 I suggest that the missing link is that it is only through
redescription that we can have (genuinely intentional) mental states. That is, it is
only through redescription that states and representations can be about anything. I
suggest that we initially acquire intentional states by being causally situated in the
world. Initially we do not have access to these states, which constitute our implicit,
procedural knowledge. Later we redescribe this knowledge in terms of explicit,
accessible structures-that express our knowledge, thoughts and beliefs.
Section 9 returns to creativity. This can now be seen as the struggle to articulate
what it is to be (causally situated) in the world. We first acquire intentional states by
being situated in the world. These constitute our implicit, procedural knowledge.
But rather than staying on this plateau of behavioural mastery (like the beaver,
with its fixed, dam-building procedure), we redescribe our knowledge in terms
of explicit, accessible structures-as thoughts and theories that we can reflect
upon and change. This gives us a general theory of creativity that consists of
a phylogenetic component, distinguishing people from slugs and Sun4s, and an
ontogenetic component that accounts for individual differences in terms of our
abilities to deploy our redescriptive powers.
Section 10 flies a kite by suggesting that rare individuals, such as Mozart and
Picasso, may have direct access to their redescripti ve powers. If machine learning
theorists can implement this ability then machines may become more creative than
people.

3 Karmiloff-Smith 1986,1990,1991, 1992a, 1992b, 1993, etc.; Clark: and Karmiloff-Smith, 1993.
 CREATIVITY, TIfOUGHT AND REPRESENTATIONAL REDESCRIPfION 47
2. Boden on 'breaking out'

In the preceding paper in this volume, and in more detail in her book (Boden,
1990/92; see also Boden, 1993, 1994), Margaret Boden argues that creativity is
the mapping, exploration and transformation of conceptual spaces (shortened to
'METCS' in Boden, 1994). A conceptual space is a space of structures generable,
that is, defined, by the rules of a generative system. We map it and explore it by
combining its elements according to the rules-thereby discovering structures in
the space. We transform it by mapping and exploring it, and then modifying the
rules to produce different types of structures. More recently, Boden has termed the
first kind of creativity 'improbabiIist' and the second kind 'impossibiIist' (Boden,
1994).
Boden argues for her thesis by first considering the notion that creativity involves
the novel combinations of old ideas. She accepts that creativity involves novelty,
but she observes that such accounts do not tell us which combinations are novel,
nor how novel combinations can come about. Her main, and related, criticism is
that many creative ideas not only did not occur before, but could not have occured
before. Previous thinking was trapped in a framework, relative to which new
ideas were impossible. Consequently, creativity sometimes requires us to 'think the
impossible', to have ideas that are impossible in the present framework: we need
to 'break out' of a conceptual space by changing the rules that define it. Kekule,
for instance, broke out of the space defined by the rules of nineteenth-century
chemistry, and opened up the new space of aromatic chemistry. We 'break out',
says Boden, by using heuristics (that is, higher-level rules) to modify the rules that
define the space.

3. More about 'breaking out'

It is undoubtedly true that we perceive a relationship between creativity and 'break-
ing free', whether from frameworks of ideas, traditions, chains (think of Houdini),
or prison. In the Portrait Joyce says that nets are thrown at the soul at birth-the
nets of nationality, language and religion-and that the soul must try to fly free of
them. The notion of being trapped in a picture is a reCUITent one in philosophy:
Hobbes said the philosopher is like a bird that has flown into a room down the
chimney, and cannot escape. Wittgenstein said it is like the person who cannot get
out of a room, even though the door is open behind him.
Consider, then, two cases. First, imagine, you are a Newtonian physicist who
believes that light travels in straight lines. There is now a sense in which (as Boden
says) it is unthinkable for you that light waves should bend, since it is logically
impossible for Newtonian physics to be true and for light waves to bend: necessarily,
ifthe laws of Newtonian physics are true then light does not bend. Whilst you remain
within the Newtonian framework you are trapped by this necessity. Next, imagine
you are following the rules of Euclidean geometry. Within the framework defined
48 TERRY DAKrNALL

by these rules it is logically impossible for the sum of the angles of a triangle to be
other than 180 degrees. But if we can represent the rules to ourselves, we will see
that the necessity lies in the hypothetical 'lfthese rules then X'-not in X itself.
This shows us how we are limited by the assumptions we make, but it does
not explain why we find it so hard to articulate the assumptions-why we 'get
trapped in pictures', why we suffer from 'tunnel vision'. And there is a stronger
case. Sometimes we are not rule users at all: we are rule followers. Rule users use
rules, but rule followers are subject to them. We are rule-followers when we blindly
follow the rules of grammars. We become rule-users when we articulate the rules
to ourselves, and we 'break out' when we change them in appropriate ways. How,
then, do we articulate knowledge that wefollow (that is, procedural knowledge) as
knowledge we can lise (that is, declarative knowledge)? Douglas Hofstadter asks
the same question in his paper in this volume:
how about the fact that not all tacit assumptions are available, not even to
a brute-force search, since they may be encoded as procedural rather than
declarative knowledge?
A related question is: Where do the explicit rule-structures come from? If we
created them (i.e. located them in a space of rule-sets), it is difficult to see how
we can avoid a regress. If on the other hand we did not create them, it seems that
we are stuck with a hierarchy of rules, floating, as it were, in mid-air. Rowe and
Partridge (1992), and Partridge and Rowe (1994), make a similar point. They quote
Hofstadter (1986: 544) as saying:
we should side-step the topless tower of bureaucracies and meta-bureaucracies
above by making rule-like behaviour emerge out of a multi-level bubbling broth
of activity below. 4
Hofstadter is talking about flexibility and control, but the point can be restated
in terms of aetiology. We can make sense of rules emerging out of low-level
interactions with the world: we can imagine how they might have developed and
how they are grounded in the environment-but the same cannot be said for rules
and meta-rules that go up into the ether, rather than down into the earth.
Finally, Boden's account is in danger of begging the question by saying that
improbabilist (or routine) creativity is the creative search of conceptual space (or
the search of a limited 'creative space'). After all, Mozart was consistently more
successful than Salieri in searching the musical space of his day, so did he search
his space more creatively, or was his space already limited to a subset of interesting
cases (to a 'creative space')? Generative rules only determine wellformedness, so
we need to know why some citizens of conceptual space are creative whilst others
are not.
The same applies to impossibilist (or radical) creativity. Modifying a ruleset at
random will give us a new conceptual space, but it will not necessarily give us an
4 In his paper in this volume, Hofstadter explains how new concepts 'bubble up' in the COPYCAT
program.
 CREATIVITY, TIiOUGHT AND REPRESENTATIONAL REDESCRIPTION 49

interesting one: it might give us a space that is different and arbitrary, or different
and dull. We need to know what modifications to a ruleset, and what new spaces,
are creatively interesting, and which are not.
To summarise. The creative process does not always start with declarative
knowledge. It sometimes starts with implicit knowledge, which we need to articu-
late as a declarative structure. And we need to know, in general, where declarative
knowledge comes from. And we need to know why some of the citizens of con-
ceptual space are creatively interesting whilst others are not.

4. Phylogeny
Imagine a 'cognitive ladder'. At the bottom of the ladder are marshflies, hoverflies
and ants. The marshfly is exasperatingly persistent and follows us about despite our
attempts to swat it away. Hoverflies meet and mate in midair. Ants exhibit apparently
complex behaviour. These creatures, however, have little or no knowledge, either
procedural or declarative. Their behaviour is principally driven by information that
is in the environment, not in the organism (Dreyfus 1979). The marshfly follows
the carbon dioxide that we emit. Hoverflies are hardwired to transform a specific
signal into a specific muscular response (Boden, 1990). Ants can detect contours,
and have routines for selecting the most level path. 5
When we discover these things we should abandon any suspicions we might
have had about these creatures having a mental life. They are not merely mindless:
they have no knowledge worth speaking about. They are not even rulefollowers.
Midway up the ladder is the beaver. It has structured, procedural knowledge
about how to build a dam. However, it has no access to this knowledge. It is trapped
in a procedure, doing what its rules tell it to: first you put in stones, then you put
in big logs, then you put in small ones. If a storm sweeps away the small logs, the
beaver starts again from scratch. (This may not be empirically true of beavers-if
it is not, run the argument with a more stupid creature. 6 ) In this respect the beaver
is like a computer program: it is a rule follower rather than a rule user. It does not
represent the rules to itself as an accessible structure. If it did it would be able to
directly access the rule for putting small logs on top of big ones.
Nor does the beaver have any thoughts. We see it scurrying about trying to put
the log there. The log keeps slipping out, and the beaver keeps putting it back.
We want to say that the beaver thinks that the log should go there-but we do not
want say that it has the thought 'the log should go there'. This is a distinction that
we commonly draw in the case of people: the tennis player is 'thinking what she's

5 Simon, 1969. See discussions in Pylyshyn (1979/1981), Winograd (1981), and Dreyfus and
Dreyfus (1987).
6 There is a species of wasp that drags its egg to the entrance of the nest, enters the nest to make
sure that it is safe, comes out, and drags its egg into the nest. If the egg is moved slightly away from
the entrance of the nest whilst the wasp is inside, the wasp moves it back to the entrance, re-enters
the nest, comes out... and repeats the process as many times as the egg is moved.
50 TERRY DARfNALL

doing' when she plays intelligently, but this does not mean that she has explicit
thoughts about what she is doing.
(Adrian Cussins (1990) draws a similar distinction between thoughts that have,
or do not have, conceptual content. If Jo believes that Bill is a bachelor, then Jo has
grasped the concept 'bachelor'. But when we say that Fido believes that the noise
(or the scent) comes from the south, we do not believe that Fido has grasped the
concept 'south'! Andy Clark discusses this in the next paper in this volume.)
We tie these facts together when we say that the beaver 'doesn't know what it
thinks'. It thinks that the log should go there, but it doesn't know that it thinks this.
It does not have access to its procedural knowledge. .
Now consider Le Penseur (Rodin's sculpture of someone engaged in deep
thought-head bent, brow furrowed). Le Penseur's human equivalent has thoughts
by virtue of being able to express his knowledge as declarative structures that he
can reflect upon and change. This is an ability that all of us have.
Enter the empirical evidence-that we are l1atively el1dowed with an ability to
redescribe our procedural knowledge as declarative knowledge (and to continue to
redescribe it at increasingly abstract levels). We are endogenously driven (driven
from within) to go from being rule followers to rule users.

5. Representational redescription

The Representational Redescription Hypothesis (RRH) is a theory of cognitive de-

velopment proposed by Annette Karmiloff-Smith (see references earlier) that tries
to account for the way in which the human mind goes beyond domain-specific con-
straints. It maintains that the mind is endogenously driven to go beyond behavioural
mastery and to represent its knowledge to itself in increasingly abstract forms. It
does this, as it were, under its own steam - without need of exogenous pressure.
Initially, the system's knowledge is embedded in procedures. It is implicit in the
system's abilities. It is not available to other procedures, nor to the system as a
whole. As Karmiloff-Smith puts it, it is il1 the system, but not available to the sys-
tem. Later the system redescribes it as declarative knowledge, which is available to
other procedures. The system continues to redescribe its knowledge on increasingly
abstract levels, all of which are retained and may be accessed when necessary.
The most commonly cited evidence for this is the 'funny men' pictures, origi-
nally presented in Karmiloff-Smith (1990). See Figures 1 to 6.
We start with a group that has achieved mastery of a task. Children between
the ages of four and six, for instance, are able to draw houses and stick figures of
people. But when they are asked to draw a funny man or a funny house, they can do
very little. They are locked into their house-drawing or man-drawing procedures:
first you draw the head, then you draw the body, then you draw the legs, etc .. To
begin with they can only modify parts of the procedure, such as drawing a square
head or a round arm or squiggly windows. But slowly they acquire the ability to
access the procedures as declarative structures, whereupon they can apply the rules
 CREATIVITY. THOUGHT AND REPRESENTATIONAL REDESCRIPTION 51

H-Jessie 4.11 M-Jade 5,7

M- Philippa 5.11 M -Leo 8,6

Fig. 1. Shape and/or size of elements changed (ages are in years, months)

in any order they wish. Children between eight and ten are quite fluent at this, and
can draw hands on the end of legs, and feet on the end of arms. They can even
merge the elements of different procedures together and draw winged houses (cf.
winged horses!) and creatures that look like centaurs.
'But,' you might say, 'the younger children could have done things like that-it
just didn't occur to them to do so'. To meet this objection, Karmiloff-Smith asked
the 4-6 year-olds to draw a man with two heads. Some of the children drew a man,
drew a second head, and then went on to draw a second man: they were locked into
a man-drawing procedure.

The plot so far, then, is this. Sometimes we are creative in the sense that
we redescribe implicit procedural knowledge as explicit declarative knowledge
(commonly we go on to modify this knowledge).7 With this transcendence corne
thoughts and the mental life. And there is empirical evidence that we are natively
endowed with this ability.
7 We might call the process of redescribing procedural knowledge as declarative knowledge, and
then modifying it, 'revelation'--a term lowe to my Ph.D. student, Corey Venour.
52 TERRY DARI'NALL

Ii - Nalaaha 4.11 Ii - Rossini 5. 1

Ii - Nicola 9. 4

Fig. 2. Shape of whole changed (ages are in years, months)

We are bound to ask more. We are bound to ask whether it is only redescribers
who are creative and have mental lives. In the next section I argue that this is
indeed the case, since our declarative structures can only be about anything if they
are grounded in implicit knowledge brought about by our being causally situated
in the world.

6. Intentionality and redescription

The problem of intentionality is the problem of how mental states, symbol struc-
tures or artifacts can be about anything. (In AI circles it is known as the 'symbol
grounding problem' (Hamad, 1990).) I have argued that a system can be creative
only if it can access its knowledge structures. But what is it for a structure to
represent knowledge for a system? The standard computational account is that
the structure is in the mind's 'knowledge bin' or 'thought bin' and that the mind
manipulates it according to formal rules. Dennett (1986) calls this "High Church
Computationalism".8 The Knowledge Representation Hypothesis (see especially
Brian Smith, 1985) says that a system contains knowledge if and only if it contains
a syntactic structure that means something to us, and that causes the system to
8 See also Richardson, 1981 and Fodor, 1987. Clark (1993) provides a detailed analysis of what
he calls this 'text-based' approach.
 CREATIVITY, lHOUGHT AND REPRESENTATIONAL REDESCRIPTION 53

hi-PETER 5, 3 H- MARY 5,3

hi-ANNA 5, 6 hi-BETTIAH 10,2

.t5 At
H- JOSHUA 8, 8 hi-VALERIE 9,0

Fig. 3. Deletion of elements (ages are in years, months)

behave in an appropriate way. E.g. the system knows that tigers bite if and only
if it contains a structure such as 'Tigers bite' that causes it to get up trees in the
presence of tigers.
It is not clear how locating a structure within a system can enable it to know
what the structure means. And understanding a sentence surely involves more than
being propelled by its morphology. Rather than pursuing these points I shall outline
an account of intentionality that gives a major role to representational redescription,
and that throws light on creativity.

7. The infonnation theoretic account of intentionality

My account is based on the Information Theoretic account of intentionality. This is
most commonly associated with Fred Dretske, though John Heil and K. M. Sayre
54 TERRY DARTNALL

M- Viki 8, 7 M- C;uy 9, 6

M - Amir 9, 1 A- Koji 10,2

Fig. 4. Insertion of new elements (ages are in years, months)

have developed their own versions (Dretske 1980, 1981; Heil, 1983; Sayre, 1986).
The Information Theoretic account exploits the mathematical theory of information
advanced by Shannon and Weaver, which says that one state carries information
about another just to the degree that it is lawfully dependent on that other state.
Dretske realised that this sheds light on intentionality, since" Any physical system,
then, whose internal states are lawfully dependent, in some statistically significant
way, on the value of an external magnitude ... qualifies as an intentional system."
(1980: 286). Thus abolttness is not a unique feature of mental states but is found in
all causal relationships.

Now, however, Dretske faces the problem of cognitive error (of how we can
have false beliefs, etc.). He maintains that information is distorted by the cognitive
system. His critics complain that this commits him to saying that we can gain
knowledge by removing the errors from our cognitive systems. This implies that
cognitive systems merely distort information-and this runs counter to our belief
that more complex systems can gather more and better information. I shall return
to this.
 CREATIVITY, THOUGHT AND REPRESENTATIONAL REDESCRIPTION 55

H.-J.ssi. 9, 8 Ii - Hanuka 8, 10

Ii-Guy9,6 Ii - Justin 10, l '

Fig. 5. Position/orientation changed (ages are in years, months)

8, Combining accounts

Let us run the story with a redescriber. The redescriber is placed in the world. Its
peripheries are bombarded. It develops internal states that constitute procedural
knowledge: it can build dams, draw pictures, pronounce words, etc. This is knowl-
edge in the system that is not available to the system. It consists in the system
being in a certain state. Because this state was causally determined by the world in
a lawlike way, it is knowledge about the world.
Consider NETtalk (Sejnowski & Rosenberg, 1986). When it is bombarded with
phonemes, NETtalk develops internal states that are about phonemes. It settles
into internal states that (to use Dretske's words) are "lawfully dependent, in [a]
statistically significant way, on the value of an external magnitude". These states
account for its ability to pronounce words. Of course, it requires something like
cluster-analysis to identify the states that it has settled into, so that, as Clark and
Karmiloff-Smith (1993) and Clark (this volume) say, the states are not available to
NETtalk. They are only just available to us!
We can provide a straightforwardly causal account of the intentionality of such a
system. Moreover, its knowledge is relatively undistorted, since its internal state is
56 TERRY DAlUNALL

M - Dominic 9, 5

ii-Viki 8, 7 ii- Fina 8, 3

Ii-Sonya 10,9 Ii - Koji 10,2

Fig. 6. Insertion of cross-category elements (ages are in years, months)

a direct reflection of its environment. On the other hand, this knowledge is limited,
inflexible and inaccessible.
Back to our redescriber. The redescriber now redescribes its implicit knowledge
as explicit knowledge-as knowledge available to the system. This is about the
world because it is a redescription of knowledge that came about by being situated
in the world. Its intentionality lies in its pedigree.
With redescription comes risk and an increased likelihood of cognitive error.
This resolves Dretske's problem of why complex cognitive systems go wrong.
Cognitive error is the price that we pay for increased abstraction, flexibility and
 CREATIVITY, THOUGHT AND REPRESENTtrrlONAL REDESCRIPTION 57
an improved ability to gather information: there is a trade-off between flexibility
and risk. (Margaret Boden and Aaron Sloman have pointed out to me that skills
and abilities are fallible as well-see Boden's chapter on Hoverflies and Humans
in Boden (1990/92). Hoverflies can miss one another in mid-air. Beavers can
build faulty dams by putting in the small logs first. Of course. The point I am
making is that at least some of the fallibility that is associated with higher cognitive
systems-e.g. false belief-is a natural consequence of redescribing our knowledge
in increasingly abstract terms.)

Thus redescribers can acquire intentional states and gain knowledge through be-
ing situated in the world. They then redescribe this knowledge in terms of accessible,
modifiable structures that constitute genuine representations, such as thoughts and
beliefs. They are genuine representations, moreover, because they are redescrip-
tions. They have intentionality because they are redescriptions of states that, we
might say, have 'prime' intentionality-intentionality that is due to direct causal
influence. Genuine representations are grounded by virtue of the system 'being in
the world'.

Are all genuine representers redescribers? Consider the beaver on the one hand,
and a Sun4 on the other. The Sun4 contains datastructures but has no mental states.
Its datastructures do not have any intentionality for it, do not express knowledge
for it, do not 'mean anything to it'. Let us suppose that this is because they are
not redescriptions of prime intentionality. The beaver (certainly the marsh fly) is at
the other extreme. It has prime intentionality, but doesn't know what it thinks. Let
us suppose that this is because it hasn't redescribed its knowledge into accessible
structures. Given these assumptions, the Sun4 and the beaver have no mental life
because they are not redescribers. If having no mental life is always due to one
of these two impediments, then only redescribers have a mental life. (Kant (1787,
B75) said that "concepts without percepts are empty, percepts without concepts are
blind". We might say that 'ungrounded representations are empty, prime states are
blind' .)

This, however, only restates the problem. Consequently I shall resort to an onus
of proof argument-sometimes referred to as 'the best bad argument in philosophy'.
Let us assume that (despite some contemporary philosophical opinions!) we really
do have mental states. These are intentional states. Now, what grounds have we got
for thinking that intentional states can arise in any way other than through causal
stimulation from the environment? If we have no grounds then we must believe that
this is how intentional states come about in us. The RRH then tells its redescriptive
story. Surely the point about being a physical symbol system, rather than a virtual
one, is that it can causally acquire intentional states. Then, if the architecture is
right, it can redescribe the knowledge implicit in these states in terms of explicit
symbolic representations. Cognitive scientists and machine learning theorists must
show us how this is done.
58 TERRY DAKfNAlL

9. A general theory of creativity

We now have the rudiments of a general theory of creativity. Such a theory con-
tains a phylogenetic component and on ontogenetic component. The phylogenetic
component concerns the difference between creative systems (such as people) and
non-creative systems (such as slugs and Sun4s). The ontogenetic component is
concerned with individual differences in the ability to redescribe.
The phylogenetic component locates creativity at what Clark and Karmiloff-
Smith (1993) call "a genuine joint in the natural order". One of the great evolu-
tionary divides appears to be between systems that can redescribe their procedural
knowledge as accessible structures, and that have thoughts and a mental life, and
systems that cannot. Clark (1993) says:
What some adult humans can do and fish can't, it seems, is to map a highly
abstract picture of the structure of a particular problem solution on to a new
domain.
Now look at ontogeny. The child redescribes its procedural knowledge as declar-
ative knowledge. Here ontogeny recapitulates phylogeny.9 We then continue to
update our framework of beliefs, struggling to give expression to what it is to be
causally situated in the world.
Of course, such frameworks are not everywhere grounded in redescription.
They are only grounded as a whole. Quine and Ullian talk about a loosely grounded
web of belief (Quine and Ullian, 1970). The RRH says that there are layers of
(re)description, grounded (if my account is correct) in the bottom layer of prime
intentionality. No layers are lost or destroyed, and all can be accessed. And our
articulated beliefs, as well as a fluctuating environment, cause us to continually
restructure and update the framework.
This is most obvious in the Arts. Art tries to express what we know or believe
by articulating it as a declarative structure: Verdi's Requiem, Bach's Passions, EI
Greco's Revelation of St John, Owen's war poetry, Munch's Scream-all articulate
by extemalising. We sometimes say that we do not know what we believe until we
have written it down. In writing Sons and Lovers, D. H. Lawrence articulated his
complex attitude to his parents, that had driven him to regard his father as a despot
and tyrant. Having voiced his attitude he was able to evaluate it and change it. "We
shed our sicknesses in books," he said.
The theory of redescription explains what is going on here, and shows why it is
creative. Redescription does not give expression to beliefs that were already there: it
actllally brings them into existence. Hitherto we believed that ... such and such Gust

9 Clark (1993) quotes Halford (1989: 134):

A major difference between young and old learners is their ability to access and flexibly use
competencies they possess. The findings of phylogenetic and developmental research are ... con·
sistent. It appears to be the ability to acquire an abstract principle, demonstrated by interproblem
transfer, that best discriminates between different groups of participants.
 CREATIVITY, THOUGHT AND REPRESBNTATIONAL REDESCRIPflON 59
as the beaver believes that the log should go there, but does not have the thought
'the log should go there '). We were in states that drove our behaviour. Redescription
expresses the dispositions inherent in these states as accessible representations.

10. Are there two types of creative ability?

I will finish by flying a kite, and suggest that there may be two types of creative
ability.
For most of us the creative process is a slow and painful one. We engage
in a cycle of externalisation and evaluation, in which we externalise something,
evaluate it, readjust our goals, and repeat the process (see the papers by Edmonds
and Sharples in this volume). Attempts to produce computerised art and computer-
assisted art have found precisely this. In developing his figure-drawing program
AARON, Harold Cohen found it necessary to continually externalise and evaluate
AARON's ability in order to find out what its procedural knowledge enabled it to
do. Edmonds (this volume) quotes Cohen (1983; see also McCorduck, 1991) as
saying
we externalise in order to find out what it is that we have in our heads ... It is
through this externalising process that we are able to know what we believe
about the world.
A few rare individuals, however, seem to create with consumate ease. Mozart
said that he would experience a composition all at once-in a moment-both
before and after he wrote it down: .. Aah, what a feast is there," he said. We may be
sceptical, but the manuscripts contain no errors. There is no apperceptive agonising
here, no evidence of an externalisation/evaluation cycle. Picasso, too, was unable
to put a foot wrong: "I do not seek-I find," he said. Mozart and Picasso seem to
have directly and unproblemmatically achieved what the rest of us blindly struggle
for.
Mozart's talk about instantaneousness suggests that he had direct and sponta-
neous access to a declarative structure. In 'A Conversation with Einstein's Brain',
the tortoise!Douglas Hofstadter (1981: 433) invites us to imagine what it would
be like to experience a piece of music instantaneously by looking at the side of a
long-playing record:
So, since all of the music is on the face of the record, why don't you take it in
at a glance, or at most a cursory once-over? It would certainly provide a much
more intense pleasure.
And
Why don't you paste all the pages of the written score of some selection upon
your wall and regard its beauties from time to time, as you would a painting? ...
instead of wasting a full hour listening to a Beethoven symphony, on waking
up some morning you could simply open your eyes and take it all in, hanging
60 TERRY DARfNALL

there on the wall, in ten seconds or less, and be refreshed and ready for a fine,
fulfilling day.
Was it like this for Mozart? (He claims to have experienced it all at once before he
wrote it down.') The RRH says that we have the ability to spontaneouslyredescribe,
so did Mozart have full control, or nearly full control, of this ability?
This bizarre suggestion is rendered less implausible by •Shostakovich 's Secret'.
In The Man Who Mistook His Wife For a Hat, Oliver Sacks discusses the suggestion
that Shostakovich had a metallic splinter, a mobile shell-fragment, in his brain, so
that when he leaned his head to one side he could hear music. If this is true, there is
a sense in which he had direct access to the music. Apparently he was very reluctant
to have the fragment removed!
It may be, then, that there are two types of creative ability. The first involves
the externalisation/evaluation cycle: we construct theories about our abilities by
observing them in action. The second rests on the fact that abilities are grounded in
states. Balls are able to bounce because of their molecular structure. NETtalk is able
to pronounce words because it has settled into a subsymbolic state. Our abilities are
similarly grounded in states. Now suppose that we can redescribe the knowledge
that is implicit in these states. Then we would no longer need to construct theories
about what we know and believe by observing our actions. The RRH tells us that
we have exactly this ability. Mozart, perhaps, was especially good at exercising it.

11. Conclusion
I have argued that creativity is rooted in our ability to redescribe implicit procedural
knowledge as explicit declarative knowledge, and I have couched this in a causal,
mechanistic framework. We now have to ask: Is this the right framework? Within
the framework, the question becomes "How (if at all) can a machine go from
something like connectionist competence to structured symbolic thought?". This
question lies at the heart of AI and cognitive science, and it is significant that an
analysis of creativity, when couched in mechanistic terms, leads us straight to it.
An optimistic gloss is that we are asking the right questions. If so, we are either
close to discovering more about intelligence and creativity, or to discovering that
we have got the wrong framework, and that we are not machines.

References
Boden, M. A.: 1990, The Creative Mind: Myths and Mechanisms, Weidenfeld & Nicolson, London.
Revised edition, 1992, cardinal, London.
Boden, M. A.: 1993, Mapping, exploring, transforming, in Dartnall, T. H. (1993).
Boden, M. A.: 1994, summary of Boden (1990), with peer reviews, Behavioural and Brain Sciences.
Clark, A. C.: 1993, Associative Engines: Connectionism, Concepts and Representational Change,
MITlBradford Books, Cambridge, Mass.
Clark, A. C. and Karmiloff-Smith, A.: 1993, The cognizer's innards: a psychological and philosophical
perspective on the development of thought, Mind and LangULJge, 8: 3.
 CREATIVITY. TIlOUOHT AND REPRESENT.4JIONAL REDESCRIPTION 61
Cohen, H.: 1983, Catalogue, Tate Gallery, London.
Cussins, A.: 1990, The connectionist construction of concepts, in Boden, M. A. (ed.) The Philosophy
ofArtijiciallnteUigence, Oxford University Press.
Dartnall, T. H. (ed.): 1993a, ArtijiciallnteUigence Ii Simulation of Beha\liour Qunrterly, Special
Issue on AI and Creativity, No. 85, Autumn.
Dartnall, T. H: 1993b, AI, Creativity, Representational Redescription, Intentionality, Mental Life: An
Emerging Picture", in Dartnall, T. H., et. al. (1993).
Dartnall, T. H: 19948, Creativity, combination and cognition, peer review of Boden (1990/1992),
Behavioural and Brain Sciences, 17: 3.
Dartnall, T. H: 1994b, Redescribing Redescription, peer review of Karmiloff-Smith (1992), Be-
havioural and Brain Sciences, 17: 4.
Dartnall, T. H., Levinson, R., Kim, S., Subramaniam., D., Sudweeks, F. (eds): 1993, AAAlTechnical
Report on ArtijiciallnteUigence and Creativity, AAAI Press.
Dennett, D.: 1986, The logical geography of computational approaches: a view from the east pole, in
Brand, M. and Hamish, R. M. (eds), The Representation of Knowledge and Belief, University of
Arizona Press, Tucson.
Dennett, D.: 1992, Commandos of the word, 1imes Higher Education Supplement, 27/3/92.
Dretske, F.: 1980, The intentionality of cognitive states, Midwest Studies in Philosophy, 5.
Dretske, F.: 1981, Knowledge and Ihe Flow of Information, MIT/Bradford, Cambridge, Mass.
Dreyfus, H. L.: 1979, What Computers Can ~ Do: The Limits ofArtiJiciallnteUigence, 2nd ed., Harper
and Row, New York.
Dreyfus, H. L. and Dreyfus, S. E.: 1987, Mind Over Machine. The Power of Human Intuition and
Expertise in the Era of the Computer, The Free Press, New York.
Fodor, J.: 1987, Psychosemantics: The Problem of Meaning in the Philosophy of Mind, MIT Press,
Cambridge, Mass.
Halford, G. S.: 1989, Cognitive processing capacity and learning ability: an integration of two areas,
Learning and Individual Differences,l: 1.
Hamad, S.: 1990, The symbol grounding problem, Physica D, 42, pp. 335-346.
Heil, J.: 1983, Perception and Cognition, University of California Press, Berkeley.
Hofstadter, D. R.: 1981, A conversation with Einstein's brain, in Hofstadter, D. R. and Dennett, D. C.
(eds), The Mind's I, Basic Books, New York.
Hofstadter, D. R.: 1986, On the seeming paradox of mechanising creativity, in Hofstadter, D. R.
Metamagical Themas, Penguin Books, Harmondsworth.
Kant, I.: 1787, Kritik der Reinen Vernunft. There are several good English translations, incl. N. Kemp
Smith (1964),lmmanual Kant's Critique of Pure Reason, Macmillan.
Karmiloff-Smith, A.: 1986, From metaprocess to conscious access: evidence from children's met-
alinguistic and repair data, Cognition, 2: 3.
Karmiloff-Smith, A.: 1990, Constraints on representational change: evidence from children's drawing,
Cognition, 34.
Karmiloff-Smith, A.: 1991, Beyond modularity: innate constraints and developmental change, in
Carey, S. and Gelman, R. (eds), Epigenesis of the Mind: Essays in Biology and Knowledge,
Laurence Erlbaum Associates.
Karmiloff-Smith, A.: 1992a, Beyond Modularity: a developmental perspective on cognitive science,
MITlBradford, Cambridge Mass.
Karmiloff-Smith, A.: 1992b, PDP: preposterous developmental postulates? Connection Science.
Karmiloff-Smith, A.: 1993, Is creativity domain-specific or domain-general? Clues from normal and
abnormal development, in Dartnall, T. H. (1993).
McCorduck, P.: 1991,Aaron's Code, W.H. Freeman and Company, New York:.
Partridge, D. and Rowe, J.: 1994, Computers and Creativity, Intellect Books, Oxford.
Popper, K. R.: 1972, Objecti\le Knowledge: An Evolutionary Approach, Clarendon Press, Oxford.
62 TERRY DAKfNAlL

Pylyshyn, Z. W.: 1979/81, Complexity and the study of artificial and human intelligence, in Ringle, M.
(ed.) Philosophical Perspectives in Artificial InteUigence, Humanities Press, Atlantic Highlands,
Nl.
Quine, W. V. O. and U1lian, 1. S.: 1970, The Web of BeUef, Random House, New York:.
Richardson, R. C.: 1981, Internal representation: prologue to a theory of intentionality, Philosophical
Topics, 12.
Rowe, 1., and Partridge, D.,: 1992, Creativity: Learning flexible representations through emergent
memory, Research Report no. 215, Department of Computer Science, University of Exeter.
Sacks,O.: 1985, The Man Who Mistook His Wife For A Hat, Duckworth, London.
Sayre, K. M.: 1986, Intentionality and information processing: an alternative model for cognitive
science, Behavioural and Brain Sciences, 9.
Sejnowski, T. & Rosenberg, C.: 1986, NETtalk: a parallel network that learns to read aloud, John
Hopkins Electrical Engineering and Computer Science Technical Report, IHUIEEC-86/01.
Simon, H. A.: 1969, The Sciences of the Artificial, MIT Press, Cambridge.
Smith, B.: 1985, Prologue to 'reflection and semantics in a procedural language' , in Brachman, R.
and Levesque, H. (eds), R£adings in Knowledge R£presentDtion, Morgan Kaufmann, Los Altos,
California.
Winograd, T.: 1981, What does it mean to understand language?, in Norman, D. (ed.), Perspectives
on Cognitive Science, Ablex, NJ.