Re: Imagery Debate

From: Harnad, Stevan (
Date: Tue Jan 30 1996 - 21:36:10 GMT

> From: "Smith, Wendy" <>
> Date: Tue, 23 Jan 1996 09:22:22 GMT
> One of the characteristics which distinguishes cognitive psychology
> from behaviourism is the acceptance of a mental life - experiences.
> Part of experiences, thinking, cognising, is to form some sort of
> mental representation. The nature of these representations is the
> subject of fierce debate.

Partly because "representation" is itself controversial, even
homuncular: Representation of WHAT to WHOM? If it's a little picture in
your head, or even a description, that's bad news, because someone has
to look at the picture or read the description, and that just starts the
problem from the beginning again. On the other hand, if the internal
"thing" is just some sort of structure and/or process that gets
something DONE (solves a problem, performs a task), then it's not clear
why it should be called a "representation" at all. -- Of course, we know
why we want to call, say, images, representations: Because they're
mental; because there's something it's LIKE to be "seeing" an image. The
question then is: Is that kind of introspective experience in useful or
relevant to cognitive explanation, or is it misleading?

> Imagery has long been a contender for the medium of thought. Aristotle
> considered imagery of prime importance, and the introspectionists (e.g.
> Galton) wallowed in it. The behaviourists were suitably unimpressed,

Say why: What made behaviourists reject imagery?

> and imagery lay dormant until cognitive psychologists took a fresh
> look. However, the vague terms and imprecise definitions of the
> previous era were no longer sufficient. Cognitive psychologists wanted
> to know exactly what imagery was, what mechanisms operated the process,
> and what use imagery was to cognition. Two basic sides formed - one
> side considering images to be picture-like representations, and the
> other side considering them to be forms of propositional
> representation.
> Paivio suggested there were two distinct systems for representing and
> processing information. One system dealt with verbal material and
> stored it in a verbal form; the other dealt with image based
> representation and processing. Interference studies and
> neuropsychological evidence appeared to provide support for this.
> However, it still failed to explain exactly what images were, and how
> they were represented. Shepard performed experiments in which subjects
> had to mentally rotate an object in order to answer a question. It was
> found that the further the object had to be rotated, the longer it took
> the subject to answer.

Good, but how was this interpreted, and how did it favour imagery (and
what KIND of imagery)?

> Kosslyn asked subjects to mentally scan a map
> in order to answer questions, and again, the further the distance
> "scanned" on the image of the map, the longer it took to answer. These
> findings were taken as evidence that images were a form of analogical
> mental representation.

Why? What was the basis for that conclusion?

> In contrast, propositional representations
> consist of an abstract symbol system which can be used to express the
> contents of the mind via predicate calculus.

And what did the propositionalists (symbolists) make of the imagery
evidence, e.g., Shepard's, and Kosslyn's? It can be explained both ways,
imagistically and propositionally: Say how, and why/how one way is better
than the other, if it is.

> The imagery debate began in earnest with Pylyshyn (1973) when he
> claimed that picture-like entities were an inappropriate way of
> considering images. He made two basic criticisms. First, the concept
> of imagery was too vague to be accepted in its present form. Second,
> and more importantly, imagery was merely epiphenomenal to the process
> of thinking. Rather than images being the means by which thinking is
> carried out, they are an effect of the thinking process. Basically,
> Pylyshyn suggested that there was only one form of representation, and
> this was propositional. Furthermore, it followed that if there was a
> verbal and a nonverbal form of processing, then there must be some
> superordinate code into which both could translate, which would act for
> both carrying out the process and representing the abstract, underlying
> concept. The superordinate code was propositional. All the evidence
> produced to date could be re-interpreted in line with a propositional
> model. If images were like pictures, then a "mind's eye" was needed to
> see them; which would need another eye; and so on infinitely. This is
> the homunculus problem - who looks at the image?

Good. Now that it has been established that it COULD all be done by
propositions, need it be? And does an image really always demand a
homunculus to look at it? But then why doesn't a proposition demand a
homunculus to read it?

> Several points were answered by "imagists". They countered that images
> were not pictures as such, but were generated via perceptual memories,
> and contained spatial information. Also, although propositional
> representations could explain the empirical findings after the fact,
> imagery would actually predict the findings. The homunculus problem
> cannot be easily dismissed. However, there is another side to the
> issue. In Pylyshyn's account, if a propositional code is necessary to
> translate between nonverbal and verbal material, what translates
> between verbal and propositional? Or between the new code and
> propositional? Where do the initial propositions come from? Are they
> innate? If so, how can they be interpreted? Or do they have to be
> learned? If so, how are they learned, if thinking is via
> propositions? This is also a homunculus problem, and therefore poses a
> challenge to both sides of the debate. The concepts were not well
> specified, and because of the lack of empirical evidence, it made this
> account little better than introspection.

Good. But does "innateness" resolve this? How does innateness give
meaning to a proposition? There can be innate rules, and innate feature
detectors; but how does that help in the interpretation of propositions?
And if images are "perceptual memories," how does that (whatever it
means) solve the problem of the interpretation of the images (or the
memories)? How to substitute a process instead homunculus that does the

> One of the most difficult points to answer was that imagery was
> unnecessary, because, indeed, everything that can be explained by
> images can be explained by propositions, even if longwindedly.

This is otherwise known as the Church-Turing Thesis, according to which
anything and everything can be simulated by digital (discrete) serial

> Kosslyn
> approached the problem by attempting to specify the distinctive
> properties of images, and also remove the homuncular quality. He
> proposed a theory which acknowledged an important role for
> propositions, but also had an important role for images. Images were
> represented in a medium which had specific properties, including the
> capacity to portray spatial information. This theory was tested in
> several experiments. The subject had to image various objects, and it
> was found that the difficulty or time taken for the tasks was related
> to the "visual" properties of the image. However, introspective
> evidence of this sort can be misleading. Apart from experimenter
> expectation effects, the presence of accurate spatial data is not
> necessarily limited to non-propositional images. Neuropsychological
> evidence suggested that selective impairments of imagery existed.
> Although individually each piece of evidence was not conclusive, they
> did converge together.

They did converge, but wouldn't the most decisive evidence be a machine
that could do what we can do, and did it using internal analogue
processes that preserved the spatial properties of input, and did it
(obviously) without a homunculus? And if the digital computer version
that did the same thing had to be much bigger and more complicated, and
had to have more and more propositions built in in order to do what the
analogue processor could do easily without any special additions?

> Meanwhile, Pylyshyn (1984) had developed his earlier ideas. He now
> claimed that cognition was computation, computation was symbol
> manipulation, and this was propositional. He criticised Kosslyn's
> findings on several counts. The subjects could have been imagining
> that they were visually scanning an object, rather than performing
> operations on an image. By giving different instructions, the results
> could be changed. This is a good point, but unfortunately, Pylyshyn
> made it very complicated. He claimed that the mind had a functional
> architecture. This was not influenced by beliefs and goals etc. The
> medium for images would have to be in this architecture, and so the
> images would exist there, too. As part of the "hardware" they would
> not be influenced by beliefs etc. Pylyshyn termed this as "cognitively
> impenetrable". As beliefs etc are propositional ("software"), then if
> images are cognitively penetrable, they cannot be in the special
> medium, and will be epiphenomena based on propositions. As changing
> the instructions could affect the images, then the images were
> cognitively penetrable.

A bit of a complicated way of saying that maybe images are just a result
of a story we tell ourselves. But what about the (hypothetical) machine
that does, say, Shepard tasks (mental rotation) by "rotating" internal
analogues of its input? Would that be "architecture"? Would it be
"penetrable"? Does it matter?

> However, because one construct can influence another, does it make the
> second an epiphenomenon? It could equally be argued that beliefs are
> epiphenomenal because they can be influenced by images. His view of
> images versus propositions may also be an inappropriate way of
> resolving the issue. Kosslyn certainly thought so (1995). He
> suggested that imagery could be both propositional and visual. He
> further developed this idea, and proposed that imagery was related to
> visual perception. Imagery did not ride on the back of perception, as
> had been suggested, but was an essential component of it. When an
> object presents in a novel way, information from the input can be
> matched to the representation, and the object can be recognised. If
> there isn't enough information to get a proper match, then an image of
> the next best is generated for comparison for comparison to the
> perceptual image. This is not a template, but rather is generated in
> an effort to provide the missing information. As soon as a stored
> representation is generated, it begins to send an image for use as
> feedback within the system. Kosslyn did provide some empirical
> evidence for this proposal, but not all of it was convincing.

Imagery is not epiphenomenal because it can be influenced by
propositions. But EXPERIENCING imagery may be epiphenomenal: It's fine
for the hypothetical analogue processor to generate and use internal
analogues that are rotated and matched. But what possible causal role
can it serve that all of that is EXPERIENCED? It's the homunculus that
is viewing the imagery (us) that doesn't seem to need to be there to get
the job done; the problem is not with the form of processing, because
both internal analogues and internal symbols can be processed perfectly

> However,
> it did share similarities with the way Jeannerod addressed motor
> imagery. He suggested there was a link between motor imagery and
> preparation. The image is a representation of the intention to act,
> and will echo the movement itself, although not exactly, because it has
> no biomechanical constraints, and may be influenced by the memories of
> previous movements. Motor processes are not automatic in the way
> sensory processes are. However, if the nature of the image depends
> upon the representational mechanisms which process the sensory input,
> it may well reflect this sensory input.

That last part went by too fast: What is the difference (if any) between
the problem of motor imagery and the problem of sensory imagery? How
does each relate to the homunculus problem and to the possibility of
doing everything computationally?

> The debate is far from concluded. The main problems are: first,
> anything which can be described by imagery can also be described by
> propositions; second, accessing imagery tends to have been
> introspectional, and performance needs to be more prominent; finally,
> any account of imagery has to dispense with a homunculus to be
> convincing. This appears, not surprisingly, difficult, but a computer
> model may be called for at this stage.

Good. But explain how it might settle the matter; what directions could
the true facts end up going, depending on the outcome of the computer
modeling? This can be related to the computation issue too, as well as
the consciousness issue (and, via the symbol grounding problem) to the
categorisation issue as well!

This archive was generated by hypermail 2b30 : Tue Feb 13 2001 - 16:23:57 GMT