Posted  7 Jul 1996
Newsgroups: comp.ai.philosophy
References: <4rcir5$ioh@usenet.srv.cis.pitt.edu>
Subject: Re: Ryle on imagery and cognitive science (VERY LONG)

Apologies for great length. This message is stored as

    http://www.cs.bham.ac.uk/~axs/misc/ryle.and.cogsci

I've combined my answers to two postings by Anders Weinstein
(andersw+@pitt.edu) one on 3rd and one on 6th July. The second was in
the "deification of logic and rationality" thread.

> Date: 3 Jul 1996 01:36:37 GMT
> Organization: University of Pittsburgh

In the first message he makes some interesting comments which I think
clarify his position better than the previous messages I've read from
him.

(There's a double time lag at work here: first, the time it takes his
messages to reach my machine, and second the delay until I have time to
pay attention and respond. So we probably have messages crossing paths).

I think Anders is concerned to delineate what I would describe as

    "the Normal Adult Western Culture Common Sense Personal Realm",

or NAWCCSPR which I'll abbreviate as NPR -- i.e. the kinds of mental
states and processes normal untutored adults ascribe to themselves and
one another in our culture.

I think he has taken a coherent position, but a very narrow one, which
essentially says: I know about NPR (the normal personal realm) and I
just don't want to know how it relates to any other kinds of knowledge,
including any theory that might explain what makes NPR, and other
"personal realms", possible.

In addition to clarifying this internally coherent position, Anders
produces arguments against people like me who wish to relate NPR to a
wider reality, especially to theoretical constructs of a nascent
cognitive science.

I think this is because he has come across flawed attempts to make such
links and does not notice that rather different approaches are possible,
which avoid those particular flaws. Perhaps cognitive science at Pitt is
stranger than I believe it is.

I think I wrongly attributed to him a kind of anti-scientism that I have
found in many philosophers trained in the linguistic analytic tradition.

However, I do think he believes (like P.F.Strawson) both that there is
some timeless set of truths embodied in NPR, and that somehow by doing
philosophy one can discover these truths, which are immune from any
correction or revision in the light of scientific advance, of any sort.

By contrast, I see NPR as just an episode in the history of mind.

Moreover, it's not even unique now, for there are minds of many kinds,
including infants, toddlers, sufferers from senile dementia and brain
damage, etc. to say nothing of other animals, and the strange sorts of
mental states to be found in some other cultures.

I would expect anyone wanting a deep understanding of NPR to wish to see
how it fits into this broader framework.  We can do this if we take the
design stance and see how designs for minds can vary and environments
(niches) in which they function can vary, and what the implications of
the variations are: i.e. which features of the NPR would be changed.

Moreover we might then have explanations of the features we find in
ourselves, just as the atomic theory of matter explained features of the
common sense physical world.

This deeper and broader investigation is not possible if we restrict
ourselves only to concepts applicable to instances of a very specific
design and environment (i.e. the NPR) and steadfastly refuse to build or
pay attention to hypotheses about explanatory architectures.

I think part of the problem is that Anders thinks that all scientific
explanations (a) are *causal* and (b) must be couched in terms of
underlying *physical* structures, whereas I don't these are assumptions
that have to be taken by cognitive science.

When I explain how a word processor works by explaining how it is
implemented in datastructures and algorithms at a lower level I am not
making claims about causation, at least not as "causation" is normally
construed. I am talking about *implementation*.

Moreover, high level functionality can be explained in terms of a type
of virtual machine which is not physical, but still has explanatory
power. My impression is that Anders is unaware of the possibility of a
cognitive science that talks about implementation in terms of high level
virtual machines (rather than data structures in the brain, such as
neural maps). This is surprising, seeing that he works with computers
and there's a sophisticated group of cognitive scientists at Pitt, with
whom he associates.

Even in the physical sciences, the concept of implementation can be
useful as an alternative to causal explanation. The structure of H2O
molecules EXPLAINS the properties of water but does not CAUSE them
(except in a very general sense of causation which I would prefer not to
discuss now, and which does not involve temporal precedence).

Many properties of water such as taste, visual appearance, chemical
properties, electrical properties, etc. are implemented in and explained
by the architecture of the molecules. When A is implemented as B the
explanatory links between B and A are much stronger than the causal link
between, for example, smoking and lung cancer, but they work only one
way: for there may be more than one implementation of the same
properties. (E.g. sweetness can be implemented in different molecular
structures.)

Although Anders' position is coherent

(a) It ignores or rejects as somehow nonsensical questions about the
    following:
    - how the NPR relates to unconscious mental processes about
      which common sense knows nothing,
    - how the possibility of the phenomena that occur in NPR might
      be explained by a science of mind that goes beyond common sense,
    - how a science of mind could explain the more general possibility
      of a variety of different types of "personal realms", which might
      emerge in different cultures, different animals, different human
      individuals (infants, brain damaged, etc.), and machines of the
      future.
    - how the NPR relates to mechanisms in the brain (I am not sure
      about this: he does seem to follow Ryle here.)

All sorts of "personal realms" may differ from NPR in all sorts of
different ways including forms of perception, forms of language, types
of concepts used in thinking about the world, types of motivation, etc.
A good theory of mind should accommodate them all. You can't do that by
sticking to NPR.

I am not going to try to show here how all these differences might be
accommodated within an enriched science of mind which explains the
possibility of NPR and many other "personal realms". Instead I'll turn
to some details in Ander's rejection of this possibility, and his claims
about the "adverbial" analysis of mental contents.

Anders wrote as follows responding to some comments of mine about Ryle,
starting off by agreeing with me that Ryle was not a behaviourist.

> ...
[AW]
> I *don't* believe what people say about Ryle. For example, they say he
> was a "logical behaviorist" who wanted to "reduce mental states to
> dispositions to behave"; since we know this can't be done, no point
> reading Ryle, they say. On to brain science!
>
> This attitude is doubly wrong. First, as Sloman points out, Ryle (like
> Wittgenstein) did not deny the reality of mental processes.  And
> second, Ryle (like Wittgenstein) evinced no interest in materialism or
> in reducing the mental to anything else. Rather, in highlighting the
> largely dispositional, non-occurent bases of our mental conduct
> characterizations.

So far we agree.

> Still: Ryle did take as his target something he called "Descartes'
> Myth", the idea of the two parallel streams, one inner, one outer.

On this point, Ryle was hampered by the fact that he was writing
in the 40s long before ideas that we now take for granted about
information structures in virtual machines were commonplace. Thus he
thought that the ONLY possible sort of inner stream of (non-physical)
activity must be some sort of ghostly replica of external streams of
events and processes perceived internally in something like the way we
perceive external events and processes.

He argued that the idea of this sort of inner stream was incoherent
because it led to an infinite regress: there would have to be a ghostly
inner perceiver of these picture-like or object-like ghostly events, and
that ghostly inner perceiver would also have an inner stream of ghostly
events with a ghostly inner perceiver, etc. This sort of argument
against a Cartesian inner theatre was commonplace among analytical
philosophers at the time Ryle was writing (at least in Oxford).

So he was trying to propose an alternative conceptual framework for
accommodating mental states and processes, and, as Anders says (in more
than one of his postings), Ryle did this in terms of global dispositions
of the whole person, including dispositions to produce dispositions to
produce dispositions, etc. I.e. dispositions could be hierarchical.
Internal events and processes could then be externally unobservable
changes in such dispositions, some of which might never manifest
themselves externally because the conditions for manifestation of those
particular dispositions in observable behaviour were never realised. But
the dispositions continued to exist, such as the fragility of a vase
that is never manifested in shattering continues to exist as a property
of the vase.

Thus Ryle allowed a different sort of "inner stream" a stream of changes
of dispositional states, including high order dispositions.

I think Ryle came very close to inventing the concept of what a modern
computer scientist would describe as "an abstract machine". An abstract
machine is defined in terms of a (possibly extendable) collection of
abstract data types and a collection of (possibly extendable) abstract
operations on those data types, together with a set of control
constructs that can be used to combine "commands" to create and
manipulate instances of those ADTs.

An abstract machine is capable of being implemented in many different
ways, and the details of implementation are irrelevant to the
specification of the properties of the machine. The specification of
what the abstract machine is is totally independent of its
implementation (just as NPR can be specified independently of its
implementation). However if you ever have a working instance of the
abstract machine (or NPR) then the question arises: how is it
implemented? More generally we can ask what sorts of implementations are
possible in brains, or some other class of machines.

If Ryle had gone on to talk not just about dispositions of the whole
system, i.e. what I've called "global" dispositions, but also about
various clusters of related coexisting and interacting dispositions, he
would essentially have invented the same notion as what we might now
describe as a multi-processor virtual machine architecture.

That's what I think contemporary cognitive science is about (though not
all cognitive scientists appreciate this, especially those steeped in
either behaviourist or neuroscience traditions).

[AW]
> He
> did take as his target the "intellectualist legend" that has it that
> all overt intelligent conduct derived its character as clever, witty,
> deliberate, voluntary, etc. from para-mechanical inner intellectual
> operations which precede and cause it.

Note the phrase "para-mechanical". Ryle was arguing against an inner
stream of events which was essentially LIKE physical events, e.g.
ghostly pictures, ghostly sentence utterings, etc. These would then need
ghostly eyes and ears to be used to see or hear them (in a ghostly
fashion).

The "inner intellectual operations" Ryle admitted were of a different
kind: they involved dispositional states which, when their conditions
were satisfied, triggered changes in other dispositional states, which
could themselves change yet more dispositional states, etc.

So he introduced another sort of inner realm: but it wasn't
"para-mechanical" -- i.e. it did not have homunculi observing ghostly
pictures in something like the way we observe real scenes.

[I may be merging memories of talks I heard Ryle give in 1957 and the
next five to ten years with what's in his 1949 book. No doubt his ideas
continued to evolve after completing The Concept of Mind, though I think
everything I say about his ideas is either explicit in the chapter on
imagery or is a natural development of it.]

Ryle rightly thought that the dispositional states and processes he was
talking about could not be *defined* in terms of their implementation in
brains or whatever (just as abstract computational machines cannot be
*defined* in terms of their implementation on any particular hardware
platform), but I don't think he ever denied that they depended on brain
function. If he did he was just *wrong* !

Ryle did argue (as Anders seems to argue) that when everything was
working correctly brain processes could not explain mental processes,
whereas when things went wrong, brain processes might explain why. I
think Ryle was just wrong here. (It was a fashionable claim in the 50s
in Oxford, which several philosophers found "obvious" but for which I
never heard any argument, except the fallacious argument, which I think
I also saw in one of Anders' messages, that there are other explanations
of normal human behaviour, in terms of what people think and want,
therefore there cannot be explanations in terms of brain functions. This
assumes the false premise that any kind of phenomenon has at most one
type of explanation.)

Ryle (and Anders) must be wrong about this. For example *normal* new
born humans appear to have learning capabilities that *normal* new born
gorillas do not have. These differences can be explained in terms of
differences in the brains humans and gorillas have at birth. These brain
differences will include differences in the *implementation* of higher
level capabilities which also differ and which can provide a different
higher level sort of explanation of the observable differences. (An
example might be something like: "brains of type X cannot handle the
stack mechanisms needed to use a language with a certain sort of
grammar".)

I suspect Ryle had no conception of the notion of explanation by
implementation. One aspect of this is the non-uniqueness of most
implementations: the very same word processor can have different
implementations in different languages and on different machine
architectures. Thus explanation by implementation does not fit some of
the older canons of science: namely the requirement for generality.
The implementation of my grasp of English may be enormously different
from yours, even if our knowledge of English is the same.

It also does not fit the requirement for prediction: it may be
impossible to use an explanation of how something works to predict how
it will behave because of interactions with other things in the context,
or because some parts of the system use random mechanisms.

Note that the divergence in implementation details will typically be far
less at higher levels in an impelementation hierarchy than at lower
levels, just as divergence between human languages will generally be far
less at high levels of ontological commitment and pragmatic
functionality than in the low level details of vocabulary, syntax and
phonology. That's what makes translation possible. In the other case,
that is what makes comparisons of mental states possible, despite
differences in low level implementation.

It may be that Anders takes these divergences from "standard" notions of
scientific explanation to imply that there cannot be any scientific
explanation of the high level features, whereas my claim is that the
models of scientific explanation have to be extended for complex
information processing systems, such as people and other animals.

He sometimes writes as if he believes that any scientific explanation of
mental capabilities must refer to physical structures in the brain, and
ridicules this as if that demolished cognitive science. But that ignores
the possibility of implementation hierarchies with non-physical
intermediate levels, with which he should be familiar in computing
systems.

I think we can move beyond Ryle's position to something more
sophisticated:

(a) We now know a lot more about how to design and implement a variety
of sophisticated information processing engines whose information states
involve various high level dispositions which don't necessarily map in
any simple way on to the physical details. One common example is a
sparse array which can exist in a computer even though it has more
components than there are bits in the computer, or even more than there
are particles in the universe. Other examples are logical databases
whose contents are implicit in the combination of rules and stored facts
(deductive closure need not be assumed for these contents, e.g. if all
accesses to the databases have to be time-constrained). Other examples
are holographic memories and neural nets using superimposed distributed
representations.

(b) Although Anders wants to talk only about GLOBAL dispositional states
of the WHOLE system, we now know that that is a very restrictive way of
thinking about a complex information processing architecture (in fact,
if Ryle thought like that it is probably the closest he came to being a
behaviourist as many behaviourists make the same mistake, as do some
dynamical system theorists).

In particular we know that many complex systems (including operating
systems, distributed file management systems, graphical word processors,
factory control systems, and many AI systems), are composed of a lot of
different coexisting information processing modules with different
capabilities, different links to other modules, and different functions
within the whole.

E.g. in a robot's control system there may be low level sensory analysis
modules finding structure in sensory input, various higher level
perceptual modules looking for good interpretations of that structure in
terms of an external environment, modules concerned with motivation,
modules that store huge amounts of information in a content-addressable
form, modules that create and manipulate temporary structures for
reasoning, planning and other purposes, modules that look for
generalisations from examples, modules that can create and reason about
possible future events and actions, modules that create and evaluate
possible plans, and many more. (My examples are not all at the same
level of functionality: some of these modules may be used to implement
others.)

The mistake Anders seems to make repeatedly is to think that anyone who
talks about an architecture composed of these modules is either
postulating Ryle's "para-mechanical" inner reality (leading to an
infinite regress of homunculi) or else trying to identify these internal
components with portions of the brain, and the specific information
contents of such modules with specific bits of neural or other physical
matter (neural maps).

But there's no requirement for any simple relationship between
information states and physical states, and certainly no requirement
that functionally or structurally identical states have identical
implementations.

(I don't know why he makes this sort of mistake: he works with computers
and should know about virtual machines, like the Lisp, Prolog and Pascal
virtual machines, and the Sparc, Pentium and Alpha virtual machines, at
a lower level, unlike many philosophers who discuss these topics and get
into deep muddles about supervenience.)

HISTORICAL PRECEDENTS
This notion of an architecture consisting of interacting information
processing modules is not new and has nothing to do with computational
models.

Plato thought of a mind as a sort of political system, with law makers,
unruly masses and some sort of police force. Freud distinguished ego, id
and superego as different information processing subsystems. Kant talked
about different faculties as processing different sorts of information.
Faculty psychology used similar notions.

Even common parlance often refers to different aspects of a person that
can interact and manifest themselves at different times: e.g. "the
showoff in him led him to make claims he would later regret", "there's a
part of me that really wants to punish him and another part that accepts
that it was not his fault", and it's even an old idea from Christianity
"It is not I but sin that dwelleth in me" (St Paul???) Multiple
personality disorder is related phenomenon. I suspect it differs only in
degree from the "normal" case.

SUB-PERSON MODULES
It is easy to go wrong in thinking about sub-personal modules in a
larger architecture. One mistake is to assume that these modules must
have all the same characteristics as the whole system (e.g. having the
full range of beliefs, desires, intentions, moods, emotions, etc.). This
leads to the sort of infinite regress Ryle rejected.

Another mistake is to assume that the modules must all be directly
implementable in separate physical subsystems.

Another is to assume that if they exist we must somehow be aware of
them.

SUB-MODULES AND EVOLUTION
Since we already know that there are many animals that do not have the
full range of human mental capabilities it should not surprise us that
we include subsystems that are more like very simple animals which take
in information, analyse it, store it, transform it, transmit it,
react to it, etc. Evolution often produces mechanisms, which it copies
then varies them and combines the variants into a new integrated whole
with "emergent" capabilities. (Spinal vertebrae are an example).

Our modules won't be exactly like such animals since they are not self
sufficient in the way animals are. (This is a possible flaw in the
Society of Mind theory, if interpreted wrongly.)

SUB-MODULES AND COMMON SENSE
What we know about ourselves will depend on which self-monitoring
capabilities we have, and since any such set of capabilities will be
limited, it should not surprise us that there are sub-systems of which
we are unaware, but which are nevertheless deeply implicated in many of
the mental processes of which we are aware. (Visual and linguistic
parsing subsystems, and many learning mechanisms, could be important
cases.)

Rylean and Wittgensteinian arguments cannot refute this. Nor can any
other armchair philosophising, for it is ultimately an empirical
question what sort of information processing architecture a typical
human being has, and how the abnormal ones differ.

[AW]
> Now Dennett and others have argued that computational models are immune
> to Ryle's regress arguments.

Notice that I am not assuming that all the modules have to be
"computational" in the sense of being implementable on Turing machines
or mathematically equivalent machines. Whether such implementations
suffice is a *separate* and secondary question. That's why I talk about
"information processing" modules rather than computational modules.

[AW ... continuing on Dennett]
> To take one example, the operations that
> they posit to precede voluntary actions are not themselves voluntary
> actions, so the explanatory regress comes to an end.  Similar
> observations apply to the sense in which, at the sub-personal level,
> inner checklists might be looked up, inner images might be scanned,
> inner rule-books might be consulted.  Dennett said that the digital
> computer showed how to replace the para-mechanical hypotheses Ryle
> mocked with a set of *literally* mechanical ones.

Notice that I am making a DIFFERENT point.

I am talking about inner structures that do NOT have to be *literally*
mechanical as suggested above, or even metaphorically mechanical. Rather
they can be information structures in virtual machines, and they may be
implemented in all sorts of different ways, e.g. as abstract data
structures like sparse arrays, or as theorems implicit in a logical
database, or as patterns stored holographically in a neural net, or in
new ways yet to be discovered.

Some of them may include control functions, e.g. programs or
condition-action rules. Some may store facts or generalisations about
the system itself or about the environment. Some may store goals to be
achieved, plans currently being executed, hypothetical states of affairs
being considered, e.g. as possible futures.

Of course, if they are actually to work in the world they will
ultimately need to be implemented in some physical machine.

(What justifies semantically rich descriptions of such modules is
another question. I've addressed that in other places, e.g. ICJAI85
paper available in the Birmingham Cognition and Affect project ftp
directory: ftp://ftp.cs.bham.ac.uk/pub/groups/cog_affect/ )

EXPLORING DESIGN SPACE
We already know how to build some modules of the kind in question, and
we know how to put them together to form complex functioning systems,
but we don't yet know how to do so in such a way as to produce anything
remotely like a complete set of human (or gorilla, or squirrel, or
magpie) capabilities. (Animal vision for example is still largely a
mystery.)

When we build such a complex architecture we can show how some of the
subsystems explain the global capabilities of the whole system. Also by
analysing alternative architectures or alternative contents or
developments in the same architecture, we can explain the similarities
and differences between different systems using those architectures. We
are not stuck with a unique architecture required for NPR.

[AW now attacks what I regard as a straw man]
> This claim is completely correct, but I think it misses the larger
> point. Dennett himself seemed to be pretty clear that the subject is
> changed (pun) when one moves from what he called the personal to the
> sub-personal level. Those literally mechanical hypotheses --

I would be very surprised if Dennett thinks they are all or mostly
"literally mechanical". Even if he did once, he has certainly learnt
better in the last 10 years or so. (As shown by his recent books, which
I think have rightly moved from the intentional stance to the design
stance)

[AW] ....
> hypotheses about what inner mechanisms enable us to have the capacities
> for intelligent conduct we do, what Ryle might have called "wires and
> pulleys" questions, are simply not the same conceptual questions Ryle
> was addressing in his critique.

Ryle was not infallible!! He talked about the ability to hum a tune to
yourself in terms of a succession of dispositional states. But he was
not visionary enough to see the possibility of such capabilities being
implemented in collections of information processing modules that we are
totally unaware of but which cooperate to produce the effect.

Talk about wires and pulleys is totally irrelevant in this context. As a
programmer, you should really know better!

(Similarly Gibson was right in saying that perceptual mechanisms get
information about affordances from the optic array, but wrong in saying
that it happened in some mysterious "direct" fashion, without the need
for cooperating information processing modules. Both Ryle and Gibson
come close to advocating a belief in inexplicable magic, like
Wittenstein's comment that the head could be full of sawdust.)

If you've read Consciousness Explained (a misleading title, I agree), I
don't see how you can think Dennett was talking about "wires and
pulleys" questions. By the time he wrote that book he had learnt a lot
about computing and AI.

[AW]
> Moreover, it is still common, though becoming far less so, I think, for
> cognitivists to simply assume the sub-personal mechanisms involve
> explicit representation-governed processes where none are needed;

I think you are just assuming too narrow a classes of mechanisms and
therefore attributing this assumption to others. There are many degrees
and kinds of non-explicitness in information stores, including several
mentioned above (sparse arrays, logical databases, neural nets, etc.) If
there are cognitive scientists who are unaware of these possibilities,
so much the worse for their theories.

[AW]...
> ...but to assume that the sub-personal processes
> involve inner theorizing is an instance of a wholly unmotivated
> intellectualizing on the Rylean view.

Watch your quantifiers!!

Not ALL sub-personal information manipulation has to be theorizing in
any sense that Ryle would recognizing. E.g. the modules that search for
low level structure in the optic array are not theorising except in a
very extended sense. The whole point about the notion of an architecture
is that the sub-modules have *different* capabilities and perform
*different* functions within the total system.

But "SOME do not theorise" does not imply "NONE theorise".

[AW]
> >What he is arguing against is the view that when you imagine a scene
> >there really is inside you some sort of ghostly analogue of the
> >real physical scene and an internal state of seeing that ghostly
> >analogue.

This is unnecessarily loaded. From our current standpoint we can drop
these pejorative expressions ("ghostly") and acknowledge that
information processing systems can and do manipulate internal structures
which are not physical structures, e.g. parse trees, graphs, theorems,
proofs, sentences, paragraphs, pages, chapters, etc. etc. So there is
nothing silly or incoherent in attributing such capabilities to
sub-personal modules, of which we are not conscious.

Moreover, even if the analogy with physical seeing is wrong, that does
not mean that we do not create and manipulate internal structures.

E.g. I have a standard example (which I think you met in March this
year) which is to ask people in how many different numbers of
intersection points the boundaries of a circle and a square in the same
plane can coincide. (The answer is that there are 7 possible numbers of
intersection points including tangents. A slightly harder task is to
list all possible combinations of numbers of tangents, intersections,
vertices on the circle).

Nearly everyone asked to do this without external aids such as pen and
paper, tackles the problem by visualising a circle and a triangle and
performing operations on the visualised figures, including sliding them
around into different configurations, distorting the shape of the
triangle, counting intersection points, etc. (I'll criticise the
"adverbial" interpretation of this, below, though at one level it does
not matter whether the adverbial reformulation is used or not: that's a
mere linguistic detail.)

[AW]
> ....But let us consider the implications for understanding
> cognitive science of Ryle's critique of the bad theory of imagining -- the one
> that involves queer inner mental pictures.

You and Ryle may wish to call these "queer". They are only queer for
people who don't know about structures in virtual machines.

[AW]
> .. First, I can't help noticing
> that Ryle's attack does not in any way depend on any scientific result
> about our brains -- I wonder why Sloman finds any merit in it?

Merit in what exactly? When I quoted the stuff about imagery from Ryle
it was only in the context of showing that he wanted to talk about
mental states and processes, unlike many behaviourists. However, I don't
accept his analysis of those mental states and processes.

Ryle's introduction of high level dispositions was a good move. I think
you are using another one of these invalid inferences based on a misuse
of quantifiers. I claim that SOME useful explanations will use science
to extend our knowledge. It does not follow that EVERYTHING with merit
depends on science. There can be two way traffic. (I am a philosopher).

I also think Ryle's dispositions provide a useful but still very shallow
level of description that needs to be explained in terms of deeper
implementation layers (and eventually brain phenomena), just as the
ability of an electronic circuit to detect radio signals and play the
output through loudspeakers, needs to be explained in terms of
oscillators, filters, amplifiers, etc. But the links are very complex
and subtle.

[AW]
> Shouldn't we leave it to SCIENCE to find out whether imagining should be
> understood Ryle's way or not? Might there really be such a thing as
> a confusion that is not addressed by scientific inquiry, but
> by conceptual analysis?

There are all sorts of confusions that science cannot answer, but why
assume that EVERYONE who is trying to explain and generalise NPR in the
framework of cognitive science is necessarily subject to confusions?

It seems to be something you are determined to continually reassert
instead of trying to help develop the science, which I think would be a
far more fruitful and satisfying application for your talents!

(Try it for a month or so. E.g. think about the sorts of design
requirements for a system able to use the internet to discuss philosophy
with remote users.)

THE ALLEGED AUTONOMY OF PHILOSOPHY
It may be that you have swallowed the dogma propagated by some teachers
of philosophy, namely that philosophy, including philosophy of mind, is
a discipline that is entirely separate from science and which is immune
to correction or revision by scientific investigation.

This is just historically false. E.g. Freud's work, inspired partly by
his empirical findings had a deep impact on philosophy, and Piaget's
studies of children showed that what appeared to be all-or-nothing
states of understanding of numbers and other concepts (e.g. "more than")
could come apart in various ways. Studies of brain damage show that
all sorts of variations in human "personal realms" are possible that
philosophers could not dream up in their arm chairs).

I suspect it is ultimately intellectual laziness (or fear of inadequacy)
in the philosophers who just don't want to find out about anything
outside *their* discipline, plus maybe wanting to maintain autonomous
philosophy in universities. (None of this is conscious motivation.)

(Likewise it is a mistake to think science cannot benefit from
philosophy.)

ADVERBIAL INTERPRETATIONS OF MENTAL CONTENTS
[AW]
> Set that aside. Ryle attacked the idea that in imagining there
> are any pictures before the mind. There is imagining, he says, but
> no images. Some more systematic philosophers have characterized such
> a view as an adverbial theory. On this view, when I say:
>
>   (1) I am imagining [it seems to me I'm seeing] a red rectangle,
>   twice as long as it is wide
>
> the accusative phrase should *not* be understood to refer to any object.
> Rather, it should be understood as a complex adverb modifying the
> main psychological verb. ...On the adverbial
> theory this queer language makes perfect sense. One can see why the
> sensation is "not a nothing" -- the accusatives are significant, they
> make a contribution to the meaning of the whole sentence. But also why
> it is "not a something either" -- the accusatives don't function to
> refer to objects -- they lack the "grammar of object and
> "designation".

I think it is implausible to make this linguistic philosopher's move in
connection with things like my circle and triangle task.

It may be motivated by philosophers' ignorance of the fact that
structures in virtual machines need not be anything like physical
structures and do not have to be directly implemented in physical
structures. But that does not mean they don't exist or that they cannot
be inspected, changed, etc.

In the case of imagined complex objects (like the surface containing an
overlapping circle and a triangle and various boundary intersection
points) the grammar of objects with parts and relationships to other
objects is perfectly apposite in talking about the CONTENT of the state
of imagining, as is talk of:
    attending to a part,
    noticing a relationship between parts,
    changing the relationships, e.g. moving the triangle to a
        different location, changing its shape or size,
    counting intersection points,
    etc.

It would certainly be a mistake to assume that this necessarily involves
replicas of physical triangles looked at by replicas of physical eyes
and manipulated by replicas of physical hands. We know from computer
science and software engineering that this is not necessary, and from
philosophical arguments that it could not work as a general account of
mental states and processes.

[AW]
> The grammatical fiction is evidently the superficial similarity of
> these words to nouns which truly do designate objects. We might say
> instead that sensation words are only pseudo-objectual.

Well I prefer to use the much less contorted language of structures in
virtual machines, which we already know has a well-justified use in
computer science and software engineering, and which we can extrapolate
to the description and explanation of NPR phenomena.

[AW]
> Some philosophers want to apply a similar explanation for the
> content-giving clauses of all intentional verbs. On this view, we might
> say, psychological language commits itself only to *representings* --
> events characterized by content -- but not to *representations* -- inner
> vehicles which bear that content.

There is no point in adopting such linguistic contortions if there is a
well-disciplined alternative that has powerful explanatory content and
can itself be explained in terms of subtle forms of lower level
implementation, and which, moreover, can contribute:

(a) to answering old philosophical questions about how mental processes
    are possible,
(b) to scientific questions about how individual and species differences
    arise, and explaining the differences between capabilities of
    different machines
(c) to engineering design tasks such as applied AI
(d) to cognitive engineering such as designing therapies, educational
    strategies, etc. to help people whose personal realms are not achieving
    the results they themselves want such as understanding mathematics or
    coping well in social interactions.

[AW]
> ....If the accusative phrase is simply non-referential,
> then it *also* does not refer to a 2-dimensional data structure -- a neural
> image -- in the back of the brain, for example.

What could a neural image be? (It certainly won't be a neurally
implemented picture in the brain observed by a neural eye in the brain).

Why are you assuming that I (or anyone) should think that if the phrase
refers it should refer to neural images with physical locations? You
already know that talk about a 2-D array in a computing system does not
necessarily involve reference to a 2-D structure physical structure in
the computer.

(Of course, I do assume that most of the virtual machine structures that
I want to talk about are *ultimately* implemented in brain mechanisms
(plus relationships to the environment) though probably in very subtle
and complex ways which will not be understood for many years to come. My
pessimism on this is based mainly on my work on vision.)

[AW]
> So how should we understand the relation between (1) as Ryle understands it
> and a cognitive scientific investigation like Stephen Kosslyn's into
> mental imagery?

I hold no brief for Kosslyn's models.

In the past I've accused him of confusing (a) the discovery that his
experimental subjects understand the instructions they are given (e.g.
"Imagine a banana in your left hand and an elephant a mile away") and
can understand the questions they are asked ("Which one takes up a
larger region of your visual field?") with (b) the alleged empirical
discovery that they create and inspect picture-like structures.

(I am not saying that they cannot do the latter: just that the
experimental result that subjects say the banana takes up a larger
region proves nothing more than they understand the task and the
question and have the knowledge to answer it. It tells us nothing about
HOW they answer it. Compare "Imagine a banana above an elephant. Is the
elephant below the banana?" Getting the answer "yes" would not imply
that a picture is created and inspected. It could be a simple logical
inference. By contrast, the triangle and circle task requires more
elaborate structure manipulation, and it is highly implausible that
anyone does it using logic in their heads: most people can't even write
down a logical proof on paper.)

[AW]
> If Ryle is right, one thing we *can't* say is that these inquiries are
> finding the unseen hidden referent of the accusative in a sentence like
> (1).

If you have met people who think that brain structures are unseen hidden
referents of statements like "I am imagining a triangle" then I agree
they are confused. But you should not generalise from them.

Brain mechanisms may *implement* the processes of imagining, but that's
a totally different matter.

I think part of your reaction to what I say is that you interpret it in
terms of some of the silly things other philosophers have said.

(Perhaps I misinterpret you in the same way.)

[AW]
> Another consequence, I believe, is that we can't suppose that *failure*
> to find any neural picture in the causal background of the assertion or
> judgment of (1) thereby renders it false.

I don't think that's relevant to anything I was talking about. Failure
to find a suitable 5-D structure in a computer would not render false
the statement that the system is using a huge 5-D array. What on earth
makes you think anyone might say otherwise?

[AW]
> > If John, say, asserts it, and
> it coheres with the other things John says and does, then it is true,

Well that's a (logical?) behaviouristic way of talking, which I don't
think has any hope of being made convincing in detail.

I don't think truth-conditions for any interesting mental state
descriptions can be cashed in such simplistic behaviourist terms, any
more than statements about the contents of a datastructure in a lisp or
prolog process can.

One problem is that describing someone as "saying" or "asserting"
already presupposes that the person has intentional states, and that's
part of what we ultimately want to explain.

[AW]
>...
> I believe these consequences fully support the outlook I expressed in other
> posts. Cognitive science is a genuine science -- there really are
> retinotopic maps scattered about our cortex,

You seem to be muddling cognitive science with neuroscience. This is
strange, given your environment.

Admittedly they are part of a common larger investigation.

But you also seem to assume that if a relationship exists between mental
contents and physical structures it must be a very simple one.

[AW]
> ...for example, and they might
> well be considered image-like representations. But they are not thereby
> *mental* representations.

Who on earth would say they are? Where did you get such an idea?

Is THAT what you have been arguing against all this time?

[AW]
> Finding these 'data structures"

Sounds to me as if you don't know what data structures are. You won't
find a silicon map of a 10^100 by 10^100 sparse 2-D array in any
computer. Similarly there's no reason to expect that you'll find neural
structures that can be identified with any datastructures in high level
mental modules.

[AW]
> ...simply cannot
> be understood to be finding the hidden inner objects that make a
> judgment like (1) true or false -- for no inner objects of any kind,
> whether mental or neural, are introduced by a judgment like (1).

Now you seem to be muddling up truth conditions with explanatory
theories!

Maybe you should temporarily abandon philosophy and spend some time
actually trying to implement intelligent systems, in order to develop a
new set of concepts for talking about relationships between levels.

[AW]
> ...But I do want to say that the reality of the subjective is what
> is captured in statements like (1), not in the rather different project
> of the cognitive science of sub-personal mechanisms.

Well now, I can't make any clear sense of "the reality of the
subjective".

I certainly claim that datastructures in virtual machines are real: that
a word processor really does have words, sentences, paragraphs, etc. in
it, and it really does change the contents of the pages when I tell it
to increase the line spacing. These are all real.

Are they subjective? Well, there's a clear sense in which only the
machine itself can access them or change them. It can tell me the
results by displaying an image on the screen or printing stuff on paper.
And it can accept instructions from me via keyboard keys or mouse
pointer, etc.

But you and I cannot see the data structures the way we see the screen,
and we cannot change the data-structures in anything like the way we can
pick up the keyboard.

In short: we are dependent on the computer and its software as an
intermediary for inspecting or changing the data structures.

In that sense subjectivity is already there in relatively unintelligent
computing systems. Richer forms of subjectivity would require richer
architectures, e.g. with more self-monitoring capability, a rich
motivational mechanism, etc.

I suspect most readers of comp.ai.philosophy are not content with simply
acknowledging the existence of these various kinds of subjectivity.

We want to know how to produce them, why they vary from one individual
to another or from one species to another, and why they can change so
dramatically as a result of brain damage, drugs, self-analysis, etc. and
how the develop over the years in a single individual.

All of this requires going deeper than the common-sense level of
description (the NPR) that seems to be all that interests you.

I.e. we want to go beyond the NPR (if you don't want to then just ignore
us).

Kant certainly tried to dig deeper. I think Wittgeinstein tried, failed,
and decided wrongly that it was impossible. Many philosophers tried and
made a mess of it (e.g. empiricist philosophers' attempts to model minds
as collections of interacting "ideas" or "impressions", and the
philosophers who tried to identify mental states or processes with brain
states or processes, because they did not know about implementation).

MENTAL STATES AND NORMS
In a later message Anders writes in response to my critique of the
intentional stance:

> Date: 6 Jul 1996 00:54:15 GMT

[AS]
> >                                    It is often thought that somehow our
> >ability to attribute these states to others is based on our observations
> >of their behaviour, and that attributions of mentality are useful
> >because they enable us to predict behaviour of others ...
> >
> >However, it turns out very difficult to make all this stick, not least
> >because (a) behaviour is grossly inadequate guide to the fine detail of
> >what's going on inside and (b) such predictions are often wrong because
> >people are often not rational.

[AW]
> Maybe it's because intentional explanation sets a *norm* for what people
> ought rationally to do, not a description of what they will in fact do. If
> a reliable person promises to give a paper in Pittsburgh, I might predict
> he will show up here at a certain time. But if he doesn't,
> the flaw might be with him, not with my prediction -- he reneged, say,
> and didn't do what he (morally?) ought to have done.
>
> Similarly for a particular ascriptions of beliefs and desires. They set
> a standard -- a rational ought -- to which the "design stance" stuff
> ought to conform, but need not in practice.

Actually I am not normally interested in moralising about what other
people *ought* to do.

When I talk or think about their beliefs, desires, fears, ambitions,
disappointments, emotions, attitudes, etc. in others, I am interested in
what's ACTUALLY going on inside them, not what ought to happen when they
manifest what's inside them.

Similarly, if I tell someone about my beliefs, desires, fears, etc.
I am giving them my views about what's currently going on in me
(about which I can sometimes be mistaken), not telling them anything
about how I ought to behave.

Of course they may have views about how such a person ought to behave,
and so do I, but that's a completely separate issue from what the
current mental states are, just as a question about the information in a
commercial database is different from what the system containing the
database *ought* to do with the information.

Enough for now. If you don't want to join us, why don't you wait for 20
or 50 or 100 years and see what we manage to explain? (It's not going to
be easy.)

Cheers.
Aaron
===