Академический Документы
Профессиональный Документы
Культура Документы
1. PROGRAM
3. CONCEPTUAL SPACES
White
Red
Green
Black
Fig. 1. The full color space.
the frequencies of the sounds received by the ear. Thus acoustic frequency
is spatially coded in the nervous system as illustrated in Figure 2. This is
a paradigm example of an intrinsic representation in the sense of Palmer
(1978).
Apart from the basic frequency dimension of tones, we can find some
interesting further structure in the mental representation of tones. Natural
tones are not simple sinusoidal tones of only one frequency, but constituted
of a number of higher harmonics. The timbre of a tone is determined by
the relative strength of the higher harmonics of the fundamental frequency
of the tone. An interesting perceptual phenomenon is 'the case of the
missing fundamental': if the fundamental frequency is removed by artificial
methods from a complex tone, the pitch of the tone is still perceived as that
corresponding to the removed fundamental. 13 Apparently, the fundamental
frequency is not indispensable for pitch perception, but the perceived pitch
is determined by certain of the lower harmonics.
Thus, the harmonics of a tone are essential for how it is perceived. This
entails that tones which share a number of harmonics will be perceived to
be similar. The tone which shares the most harmonics with a given tone
MENTALREPRESENTATION,CONCEPTUALSPACESAND METAPHORS 27
Cochlea . ~ @ ~
Eardrum
0
~z
I
I i 'it
I L I ~ I I
• J It
I ,
Hair
cells
Basilar
membrane
Fig. 2, (a) Position of the cochlea relative to the eardrum. (b) The"unrolled" coclea. Points
along the internal membrane are resonant to specific frequencies so that local hair cells are
stimulated, Acoustic frequency is thus spatially coded in the nervous system. (From R S.
Churchland (1986)).
is its octave, the second most similar is the fifth, the third most similar is
the fourth and so on. This additional 'topological' structure on the pitch
dimension, which can be derived from the wave structure of tones, provides
the foundational explanation for the perception of musical intervals) 4
28 PETER G,~RDENFORS
the angles of my arm joints and thus representing the intended location of
my hand. In the terminology of Palmer (1978), this form of representation
is clearly intrinsic.
The job of the cerebellum is to transform the sensory vectors and the
"intention" vector into the appropriate output motor vector. In the simplest
cases this is mathematically the same as matrix multiplication which is
the standard technique for solving the geometrical problem of going from
one coordinate system to another. In the more general non-linear case,
the transformation is called a tensor, hence the name "tensor network
theory".
I will spare you the neurophysiological details of how the tensor oper-
ation may be implemented on the neuronal level. Let me only note that
it seems to be possible to realize it rather simply in terms of the spiking
frequencies of neuronal fibers. 16 The architecture of the different types of
neurons in the cerebellum seems to be well suited for the mathematical
operations postulated by Pellionisz and Llinas. In computer terminology,
the cerebellum works as a sophisticated parallel processor, operating on a
large number of neuronal 'coordinates' simultaneously. One important fea-
ture of this parallelism is that the tensor transformations can be performed
extremely rapidly, using only a few neuron cycles. (A similar transforma-
tion performed by a traditional sequential computer program would take
several orders of magnitude longer time and is thus not a neurophysiolog-
ically realistic model.)
Here it may objected that I am, after all, relying on the traditional
Turing machine metaphor of brain processes and thus reverting to the
sentential paradigm. My reply is that this form of parallel processing
works with patterns of intensities distributed over a large set of units. These
intensity patterns, which on my account represent positions in a conceptual
space, are non-symbolic and intrinsic representations. Furthermore, they
are transformed by mathematical operations like tensors. In contrast, the
sentential paradigm presumes sequential processing which operates on
symbols from some language. And here the transformations are performed
by applying inference or production rules to sequences of symbols. 17 In
my opinion, the Pellionisz-Llinas theory is an excellent example of how a
cognitive process can be described by non-symbolic computing; thus this
theory is incompatible with the sentential paradigm.
It is an open question to which extent connectionist models are suitable
for representing conceptual spaces (cf. Smolensky (1988) for a compar-
ison between connectionism and the sentential paradigm which he calls
'symbolism'). But at least it seems that the theory of conceptual spaces is
consistent with the connectionist paradigm. One reason is that the topolog-
32 PETER GARDENFORS
ical structures which are fundamental for conceptual spaces are generally
continuous rather than discrete. 18 Conceptual spaces are put forward as
a theory of information representation, while connectionist models are
indeed models of different theories in cognitive science.
The main reason for presenting the tensor network theory here is that
if Pellionisz and Llinas are on the right track (which still remains to be
seen), their concepts of mental representation would fit much better with
the notion of quality dimensions in a conceptual space than with cognition
as sentence-crunching as in the sentential paradigm. The neuronal vectors
of their theory represent points in different cognitive spaces.
Apparently, spatial representation is not unique to the cerebellum. The
cortex also abounds in topographic maps, whereby neighborhood relations
at the sensory periphery are preserved in the arrangement of neurons in
various 'deeper' CNS regions.19 For example, one finds 'retinotopic' maps
in the lateral geniculate nuclei which are arranged in six layers, each
layer arranged in a topographic representation of the retina; there are
'somatotopic' maps representing sensory positions on the body; and there
are 'tonotopic' maps where the orderly mapping of neurons with sound
frequencies is preserved from the cochlea (see the preceeding section) to
diverse areas of the auditory cortex. Another interesting aspect of these
maps are that most of them preserve the modularity of the senses, in the
way that distinct types of receptor neurons are sensitive to different features
of our environment and these features are kept distinct in the maps higher
up in the projection system.
The quality dimensions of these perceptual spaces and motor spaces can
be extracted by empirical neurophysiological and psychophysical research,
as is witnessed, for instance, by the results on color perception. And from
an evolutionary point of view, the simplest hypotheses for the cognitive
scientist is that this topological or geometric mode of representation is
also used for higher forms of mental processes. The research program
associated with the theory of conceptual spaces consists in establishing the
validity of this hypothesis. 2°
According to this picture the linguistic representations postulated by
the sentential paradigm would be rather exceptional and parasitic upon the
topological. In brief, conceptual spaces would provide a more unified way
of accounting for brain processes in general, and mental representations in
particular, than would the sentential paradigm. The idea is congenial to the
general methodology of Pellionisz and Llinas:
This view assumes that the brain is a 'geometrical object', that is to say, (1) activity in the
neuronal network is vectorial, and (2) the networks are organized tensorially: i.e. activity
vectors remain invariant to changes in reference-frames. Understanding brain functions
becomes, then, the establishment of the inherent geometrical properties of the activity
MENTAL REPRESENTATION, CONCEPTUAL SPACES AND METAPHORS 33
vectors and, more fundamentally, the determination of the proper vectorial transformations
occur." (1980, p. 1125)
I finally want to argue that the theory of conceptual spaces is also useful
for more traditional philosophical purposes. As applications of the theory
I will outline two problematic examples of concept formation. The first is
the problem of projectible predicates in inductive inferences which I have
analysed in terms of conceptual spaces in G~irdenfors (1990). The second
is the problem of how metaphors function, which will be the main topic
of the next section. Before this problem is tackled, let me present a brief
outline of the first application.
However, before I turn to these applications I want to make some
remarks on the relation between the theory of conceptual spaces and the
cognitive approach to semantics that has been developed by Jackendoff
(1983), (1987), Langacker (1986), (1987), Lakoff (1987), and others. One
unifying idea of this approach is that semantics is not seen as a map-
ping from language to 'the external world' (traditionally construed as truth
conditions), 21 but rather as a relation between language and a conceptu-
al structure. Langacker (1986) formulates this by the slogan 'meaning is
equated with conceptualization'. He develops an elaborate notation for the
conceptual structures in his (1987). In Lakoff (1987, p. 68) the correspond-
ing structures are called 'idealized cognitive models'.
34 PETERG.~RDENFORS
/
I
o
1 t
- /
67.
t
) 7
\
,r
What occupies the lowest level in conceptual hierarchies? I am neutral in regard to the
possible existence of conceptual primitives. It is however necessary to posit a number
of 'basic domains', that is, cognitively irreducible representational spaces or fields of
conceptual potential. Among these basic domains are the experience of time and our
capacity for dealing with two- and three-dimensional spatial configurations. There are
MENTAL REPRESENTATION, CONCEPTUAL SPACES AND METAPHORS 35
basic domains associated with various senses: color space (an array of possible color
sensations), coordinated with the extension of the visual field; the pitch scale; a range
of possible temperature sensations (coordinated with positions on the body); and so on.
Emotive domains must also be assumed. It is possible that certain linguistic predications
are characterized solely in relation to one or more basic domains, for example time for
(BEFORE), color space for (RED), or time and the pitch scale for (BEEP). However, most
expressions pertain to higher levels of conceptual organization and presuppose nonbasic
domains for their semantic characterization.
6. METAPHORS
OSI CS2
I I
h u m a n 3 - D model
. . . . . fore-.arm 3:___Dmodel
coordinates) and the position of the added cylinder in relation to the dom-
inating one (two coordinates). The details of this representation are not
important in the present context, but it is worth noting that on each level of
the hierarchy an object is described by a finite number of coordinates based
on lengths and angles. Thus the object can be identified as a hierarchical
structured Vector in a (higher order) conceptual space. Figure 5 provides
an illustration of the hierarchically structure of their representations.
A related analysis of shape has been proposed by Pentland (1986). He
uses so called 'superquadrics' instead of cylinders as the basic building
blocks and then adds a fractal analysis of surface structure. Also in this
theory, objects are represented by a small number of coordinates. Thus
both Marr and Nishihara's and Pentland's representations of shapes are
amenable to topological and metric analyses. I believe that this kind of
research will be very helpful for our understanding of how shape metaphors
work.
I will conclude my discussion of metaphors by comparing the present
analysis to some recent theories. One such theory, based on a wealth
of linguistic evidence, is presented by Lakoff and Johnson (1980). They
analyse several networks of metaphors used to talk about special topics.
Among other things, they argue, in line with the description above, that
the introduction of a new metaphor c r e a t e s similarities of a new kind.
These similarities are not 'objective', but once one quality dimension has
been connected to another via a metaphor this connection may serve as a
generator for new metaphors based on the same kind of similarity.
A closely related point is raised by Tourangeau and Sternberg
(1982). Their 'domains-interaction' view is based on the observation that
40 PETER G.,~RDENFORS
"metaphors often involve seeing in a new way not only two particular
things but the domains to which they belong as well . . . . Metaphors
can thus involve whole systems of concepts" (p. 214). In other words,
a m e t a p h o r does n o t c o m e a l o n e - it is not only a comparison between
two single concepts, but involves an identification of the structure of two
quality dimensions. Black (1979, p. 31) makes essentially the same point
by the phrase "Every metaphor is the tip of a submerged model."
Tourangeau and Sternberg's analysis is obviously congenial to the
present one. They even use the notion of 'dimension' when spelling out
their view: "In interpreting a metaphor, we see one concept in terms of
another by construing features or dimensions that apply within in the
domain of the first concept (the subject or tenor of the metaphor) as some-
how parallel to those that apply within the domain of the second concept
(the vehicle); further tenor and vehicle are asserted to have similar values
on these corresponding dimensions" (p. 215). However, Tourangeau and
Steinberg do not interpret dimension as a topological or metric concept;
on the contrary, they assume that "the meaning or structure of concepts
uses semantic networks ... as the chief representational device" (p. 219).
As they point out, this representation means that the two domains that a
metaphor relates "include the same skeleton of semantic relations, repre-
sented as labelled arrows" (p. 420). This mode of representation misses
much of the content of metaphors which, in my opinion, is much better
captured by comparing the structure of the underlying quality dimensions
(their examples of semantic networks on pp. 220-221 do not seem to tell
us much about the meaning of any metaphors). Also Indurkhya (1986)
interprets metaphors in terms of mappings between different domains, but,
again, the topological structure of the domains are not exploited.
Verbrugge (1980, p. 95) notes that "[m]etaphoric processes are not sole-
ly dependent on language as a source of activation or medium of expres-
sion. Perceptual experiences can be metaphoric - for example, recognizing
a familiar object in the guise of a cloud, or recognizing a familiar ocean-
wave undulation in a field of grain." According to the present proposal, this
feature of metaphoric processes is easily explained: Because perception
produces mental representations of objects in a conceptual space, these
representations can be used for metaphoric transfers in precisely the same
way as the mental representations generated by linguistic inputs. 25 As a
wild speculation, I would like to suggest that what happens in s y n e s t h e s i a
(for example seeing colors when listening to music) is a kind of 'automat-
ic' transfer from one quality dimension to another. 26 In contrast, ordinary
metaphors involve a creative 'discovery' of similarity in topological struc-
ture.
MENTAL REPRESENTATION, CONCEPTUAL SPACES AND METAPHORS 41
7. CONCLUSION
w e t w a r e . (In this r e s p e c t it is l i k e l o g i c ; t h e r e is n o e x a c t m a t c h i n g b e t w e e n
logical theory and the way people reason).
O n t h e o t h e r h a n d , I b e l i e v e that a t h e o r y o f c o n c e p t u a l s p a c e s c a n
b e v e r y u s e f u l f o r generating hypotheses a b o u t t h e n e u r o l o g i c a l f u n c t i o n -
i n g o f m e n t a l r e p r e s e n t a t i o n s a n d f o r interpreting s o m e o f t h e r e s u l t s o f
n e u r o s c i e n c e . 27 I b e l i e v e t h a t it d o e s far b e t t e r in t h e s e r e s p e c t s than t h e
s e n t e n t i a l p a r a d i g m . A s an e x a m p l e , I h a v e s h o w n h o w t h e t e n s o r net-
w o r k t h e o r y o f m o t o r c o n t r o l fits v e r y w e l l w i t h t h e t h e o r y o f c o n c e p t u a l
spaces.
A p a r t f r o m its r e l a t i o n s to the n e u r o s c i e n c e s , I a l s o b e l i e v e t h a t t h e
t h e o r y o f c o n c e p t u a l s p a c e s m a y b e a u s e f u l t o o l f o r t a c k l i n g p r o b l e m s in all
c o g n i t i v e s c i e n c e s . In p a r t i c u l a r , it can p r o v i d e the c o n c e p t u a l f r a m e w o r k
f o r a c o g n i t i v e s e m a n t i c s . I h a v e i l l u s t r a t e d this b y a p p l y i n g it to t w o
p r o b l e m areas: p r o j e c t i b i l i t y in i n d u c t i v e r e a s o n i n g in G ~ r d e n f o r s (1990)
a n d t h e f u n c t i o n i n g o f m e t a p h o r s in the p r e s e n t paper. I h o p e that f u t u r e
w o r k w i l l r e v e a l its v i a b i l i t y for f u r t h e r p r o b l e m areas.
NOTES
* I wish to thank Ruth Barcan Marcus, Max Black, Martin Edman, Jaakko Hintikka, David
Pears, and A_saWikforss for helpful comments on a much earlier version of this paper which
was presented at the entretien of the International Institute of Philosophy in Stockholm,
September ! 987. For comments on a later version I am grateful to Paul Churchland, Paul
Hemeren, and Mark Johnson. Research for this work has been supported by the Swedish
Council for Research in the Humanities and Social Sciences.
Here I am referring to the traditional sequential kind of computer programs with 'explicit'
symbol representation and not to parallel distributed processing which may use 'intrinsic'
representations (cf. Palmer 1978). More about this in Section 4.
2 Marcus (1987) names this position "the language centered theory of belief". It is also
closely related to what Palmer (1978) calls 'propositional representation' and what Smolen-
sky (1988, p. 3) calls 'the symbolic paradigm'. Lakoffs presentation of objectivism (1987,
pp. xii-xiii) also encompasses a linguistically oriented version of the sentential paradigm.
3 p. S. Churchland (1986, Chapter 9.6) presents some of the most important theoretical
drawbacks of the sentential paradigm.
4 Cf. Fodor (1981, p. 200): "And, if we have to have internal representations anyhow, why
not take them to be the objects of propositional attitudes ... really we have no choice. For
the account is well evidenced, not demonstrably incoherent, and, again, it's the only one in
the field."
5 This topic is studied at greater length in G~irdenfors (1990).
6 My presentation is inspired by R S. Churchland (1986, Chapter 9.6).
7 Cf. Pylyshyn (1984), p. 29: "... to be in a certain representational state is to have a certain
symbolic expression in some part of memory".
8 Pylyshyn writes (1984, p. 27): "My brain states are not, as we have noted, causally con-
nected in appropriate ways to walking and to mountains. The relationship must be one of
content: a semantic, not a causal relation".
MENTAL REPRESENTATION, CONCEPTUAL SPACES AND METAPHORS 43
9 Cf. Fodor (1981 ): "If mental processes are formal, then they have access only to the formal
properties of such representations of the environment as the senses provide. Hence, they
have no access to the semantic properties of such representations, including the property
of being true, of having referents, or, indeed, the property of being representations of the
environment" (p. 231) and "We must now face what has always been the problem for
representational theories to solve: what relates internal representations to the world? What
is it for a system of internal representations to be semantically interpreted?" (p. 203). These
problems arise for the sentential paradigm because it operates with external representa-
tions.
10 The notion of a quality dimension is closely related to the 'domains' in Langacker's
(1986) semantic theory,
11 The topology of discrete dimensions is further discussed in Gardenfors (1990).
I2 For further connections between conceptual spaces and semantic problems, cf.
GS.rdenfors (1990). See also Sections 5 and 6 of the present paper.
13 See Gabrielsson (1981, pp. 20-21 ).
14 For some further discussion of the structure of musical space cf. G~irdenfors (1988),
Sections 7-9.
i5 See Toulmin and Goodfield (1965).
16 Cf, P, M. Churchland (1986a,b).
iv Pylyshyn (1984, pp. 73-74) has the following remarks on the relation between parallel
processing and the sentential paradigm: "I would not be the least surprised to find that the
mind's architecture is so different from that of a v o n Neumann machine that it might not
be recognized as one of our computers. No doubt, the correct architecture will contain a
great deal of parallel processing. Many people think that a substantial degree of parallel
processing makes a fundamental difference in what is considered computing. The issue,
however, is not whether the mind is a serial computer or a highly parallel one; it is whether
the mind processes symbols, whether if has rules and representations . . . . My point here is,
so long as cognition (human or otherwise) involves such semantic regularities as inferences,
and s~long as we view cognition as computing in any sense, we must view it as computing
over symbols. No connectionist device, however complex, will do, nor will any analog
computer;... ". I agree with everything except that we must view cognition as computing
over symbols.
18 Lakoff (1988, p. 40) discusses the relation between 'cognitive semantics' and connec-
tionism As is clear from the following quotation, cognitive semantics (to be presented in
Section 5) fits very well with the theory of conceptual spaces: "The basic mechanisms of
cognitive semantics include cognitive psychology, mental spaces, metaphor, and metonymy.
Technically, cognitive semantics is consistent with the connectionist paradigm but not with
the symbol manipulation paradigm . . . . At present there is a gap between connectionism
and cognitive semantics: We do not know how cognitive topology can be implemented in
connectionist networks".
i9 Cf. P. M. Churchland (1986b), pp. 281-283, and Cook (1986), Chapter 2. In P. M.
Churchland (1986b, p. 280), we find the following idea: "... the brain represents various
aspects of reality by a position in a suitable state space; and the brain performs computations
on such representations by means of general coordinate transformations from one state to
another", here Churchland uses 'state space' basically in the same way as 'conceptual
space' is used in this article.
20 Again, the relation between connectionism and conceptual spaces is unclear. Lakoff
(1988, p. 39) notes in his criticism of Smolensky (1988) that "Smolensky's discussion
makes what I consider a huge omission: the body. The neural networks in the brain do not
44 PETER GJxRDENFORS
exist in isolation; they are connected to the sensorimotor system. For example, the neurons
in a topographics map of the retina are not just firing in isolation for the hell of it. They are
firing in response to retinal input, which is in turn dependant on what is in front of one's
eyes. An activation pattern in the topographic map of the retina is therefore not merely a
meaningless mathematical object in some dynamical system; it is meaningful".
21 Or a mapping from language to a class of possible worlds, as in intensional types of
semantics, e.g. Montague semantics; or a mapping from a language to a partial world, as
in situation semantics.
22 For those not familiar with the riddle, "grue" is defined as "green and observed before the
year 2000, or blue and observed after 2000" and "bleen" is defined as "blue and observed
before the year 2000, or green and observed after 2000."
23 This analysis is congenial with Lakoff's 'Spatialization of Form hypothesis'. He writes
(1987, p. 283): "Strictly speaking the Spatialization of Form hypothesis requires a metaphor-
ical mapping from physical space into a 'conceptual space'. Under this mapping, spatial
structure is mapped into conceptual structure. More specifically, image schemas (which
structure space) are mapped into the corresponding abstract configurations (which structure
concepts). The Spatialization of Form hypothesis thus maintains that conceptual structure
is understood in terms of image schemas plus a metaphorical mapping.
24 Also cf. Langacker (1986) for some interestnng ideas about spatial structures in seman-
tics.
25 Cf. Verbrugge (1980) " . . . the medium of metaphoric comprehension is closely related
to direct experience" (p. 104) and " . . . semantic descriptions should be cast in terms of
the invariants that structure perception and action, rather than forcing them into a single
mathematical mold, such as a calculus of propositions . . . . Since metaphor traffics in similar
and transformed identities, the assumptions one makes in defining 'identity' are crucial to
a theory of metaphoric structure and process. Realism suggests that a linguistic semantics
rooted in perception and action is the most appropriate basis for explaining the compre-
hension process and the experienced identities of topic and vehicle." (p. 105). Already
Aristotle emphasized the perceptual aspect of metaphor when he wrote that " . . . to make
good metaphors implies an eye for resemblances" (Poetics, 1459a10).
26 Cf. Osgood (1980, p. 204): " . . . the cognitive processes in both color-music synesthesia
and in metaphorical use of language can be described as the parallel alignment of two or
more dimensions of experience, defineable verbally by pairs of polar adjectives and with
'translations' occurring between equivalent regions of the continua".
27 Cf. P. M. Churchland (1986b, p. 41): "There has been a tendency among neuroscientists
to restrict the term 'topographic map' to neural areas that mirror some straightforward
aspect of the physical world or sensory system, such as the retina, or the surface of the skin
This is unfortunate, since there is no reason for the brain to show any such preference in
what it constructs maps of. Abstract state spaces are just as mappable as concrete physical
ones, and the brain surely has no advance knowledge of which is which. We should expect
it, rather to evolve maps of what is functionally significant, and that will frequently be an
abstract state space."
REFERENCES
Black, M.: 1979, 'More about Metaphor', in A. Ortony (ed.), Metaphor and Thought,
Cambridge University Press, Cambridge, pp. 19-45.
Churchland, E M.: 1986a, 'Cognitive Neurobiology: A Computational Hypothesis for
Laminar Cortex', Biology & Philosophy 1, 25-51.
MENTAL REPRESENTATION, CONCEPTUAL SPACES AND METAPHORS 45
Department of Philosophy
University of Lund
Sweden
Note added in proof." This paper was completed in 1989. In the meantime, I have further
developed the theory of conceptual spaces and applied it to new areas. See e.g., the following
articles:
G~denfors, P.: 1990, 'Frameworks for Properties: Possible Worlds vs. Conceptual Spaces',
in L. Haaparanta, M. Kusch, and I. Niiniluoto (eds.), Language, Knowledge, and Inten-
tionality, (Acta Philosophica Fennica, 49, pp. 383-407), Helsinki. Reprinted in From a
Logical Point of View, Vol. 2, (1993), pp. 15-32. Reprinted in S~miotiques, D6cembre
1994, No. 6-7, pp. 99-120.
G~denfors, P.: 1993, 'Inducation and the Evolution of Conceptual Spaces', in E. C. Moore
(ed.), Charles Peirce and the Philosophy of Science: Papers from the Harvard Sesqui-
centennial Conference, University of Alabama Press, Tuscaloosa, pp. 72-88.
MENTAL REPRESENTATION, CONCEPTUAL SPACES AND METAPHORS 47