Q: INPUT PROBLEM
FODOR:
"Really massive modularity is a coherent account of cognitive
architecture only if the input problem for each module (the problem
of identifying representations in its proprietary domain) can be
solved by inferences that aren't abductive (or otherwise holistic);
that is, by domain-specific mechanisms."
Kovacs:
Fodor's real problem with abduction here is that if an encapsulated
module can not determine (or compute) which module should operate with
a given representation as an input, here is what we've got:
FODOR:
"What we've got so far is, in effect, an argument that each modular
computational mechanism presupposes computational mechanisms less
modular than itself, so there's a sense in which the idea of a
massively modular architecture is self-defeating".
Kovacs:
The problem with the argument is that it presupposes that the input
analysis can only be done by a computational mechanism. If we agree
with this, the problem is gigantic: we have fast encapsulated modules,
but for activating them, we need slow global abduction, that is, we
have to search all our mind (knowledge) to decide which one to
activate.
Kovacs:
However, if we do not insist on a computational system for doing this
analysis, we are saved. For example, the receptor of a given hormone
(for example ACTH) can analyze it's input mechanistically, and
obviously does not have to "know" anything about all the hormones in
our blood stream. If it was an input analysis in the Fodorian sense,
this receptor should be able to recognize all the hormones, in other
words it should know everything the input is not in order to know what
the input is. Nevertheless, it is fortunately able to distinguish
between adrenaline and ACTH without knowing anything about the former.
Why wouldn't the mind work this way, that is, able to determine which
module to activate by the attributes of the input? It seems that Fodor
has an a priori argument:
FODOR:
{about CDM (Cheater Detection Module) as an example of the input
problem}: How does a module decide whether what it's looking at is
a social exchange? Since, Cherished Reader, you are not an
empiricist, you presumably do not believe that social exchanges
have proprietary sensory telltales.
Kovacs:
Why wouldn't I be an empiricist?
FODOR:
If you were an empiricist, I expect you would have stopped reading
some time back; probably, come to think of it, you wouldn't have
started.
Kovacs:
Let's summarize it. The mind is unable to analyze inputs by its
proprietary sensory telltales because it would smell like empiricism,
and an empiricist wouldn't read Fodor. This seems a bit weird to me, or
at least does not sound like a strong argument. And unfortunately I
could not find anything else against the empiricist view besides that
it's empiricist. But OK, we read Fodor hence we are rationalists. Let's
try something else.
Kovacs:
Let's suppose that our exemplary general input is something like the
word "exhibition", except that it's not a word and has nothing to do
with real word linguistic processing. Can the mind activate the
"exhibition-module" and nothing else by having a module for doing
this? Or, according to Fodor, is a modular system unable to do the
job? Let's imagine a module, which reduces every opportunity letter by
letter. By "e" we already know that our module is neither the
"demonstration-module" nor the "fraternity-module", and by "x", we know
moreover that it's not even the "employment-module". By using this we
could be very quick and avoid searching and matching all the possible
inputs (words). Can it be executed like a Turing-based mechanism? If
you read this on a computer, just push the F1 button and try your
Turing-based computation in the "Help module".
Kovacs:
The fact Fodor refers to, that AI researchers could not make a
Turing-machine for input-analysis for the whole mind - not even for
linguistic perception - does not prove that the mind works in
mysterious ways - only that it is more complicated than a simple
"search by word". But on the other hand we have strong experimental
data supporting this sort of selection between representations. Sounds
like a network? Might be, and we know that Fodor doesn't like networks
(they again smell like empiricism), but that is still a weak argument
(and at least the previous example is absolutely neglecting
semantics).
Q: NETWORKS
FODOR:
Networks have much the same trouble with abductive inference that
Classical architectures do (though, to repeat, for a slightly
different reason). Consider, for example, the worry about how
classical models are to reconstruct the effects of the "centrality"
of representations on their cognitive roles. Let me remind you what
the worry was: Prima facie, the centrality of a representation
changes as you go from one belief system to the next, but the local
syntax of the representation doesn't; local syntax is contex
independent. So, assuming that cognitive processes are sensitive
exclusively to local syntax, how does Classical psychology recover
the fact that the same belief may have different centrality in
different theories? Nobody knows. Well, the present point is that
if Classical models aren't able to to answer this question,
networks aren't even able to ask it. For, to repeat, the
type-individuation conditions that network architectures afford are
incompatible with a node being identified transtheoretically. And,
if it can't even be true that the same representation occurs in
more than one theory, then of course it can't be true that the same
representation survives the transition from one theory to the
next.
Kovacs:
But why is centrality indefinable to a single node in the network?
FODOR:
The smallest unit of a connectionist representation for which a
type/token relation is definable is a whole network.
Kovacs:
I'm not much of a connectionist, but I don't see the argument clearly.
Let's say that a single node can be part of several networks at the
same time, so these networks are interconnected and not discrete. In
this super-network the number of connections a node has in each single
network defines the centrality thereof. Hence its centrality is
obviously different as we change from one network to the other, because
it has different number of connection in each. In those it has several
it counts as a central node, in the ones in which it has less, it's a
peripheral one, and all at the same time. This could be a description
of the centrality-problem in a network-system, so we are now able to
ask our questions (though we might be unable to answer them). However,
according to Fodor, networks are not interconnected, hence a node can
only be part of a single network. But I don 't understand why? It seems
a priori to me, not proven, and is pretty far from how our nervous
system as a network works.
Q: OUTPUT
FODOR:
{in his previous review of Pinker's book in the London Review of
Books}: Eventually the mind has to integrate the results of all
those modular computations and I don't see how there could be a
module for doing that.
Kovacs:
Now this is much stronger. Let's call it the output-problem: how could
a module be able to integrate the different outputs of different
modules without having a much wider scope than those modules? I don't
want to argue against this, I just really miss this argument in his
recent book, because I think it is a much better and much less a priori
argument against a massively modular mind, that is a mind without
global cognition. Why is it completely missing here, when there is a
detailed but weaker argument dealing with the input-problem?
Q: HOW THE MIND SUPERVENES ON THE BRAIN
FODOR:
Nothing at all is known about the laws according to which cognition
supervenes on brain structures, or even about which brain
structures it is that cognition supervenes on.
Kovacs:
This again seems too radical. OK, we do not know much about this
supervening, but "nothing"? Fodor himself differs from this just five
lines below:
FODOR:
Nothing we know about how cognitive structure supervenes on neural
structure impugns the possibility that quite small variations in
the latter may produce very large reorganizations in the former.
Kovacs:
So we either know at least some things "about how cognitive structure
supervenes on neural structure", and the quotation stating the contrary
is incorrect, or we really don't know anything, but in this case the
last quotation is a truism.
Q: CONSISTENCY
FODOR:
{this time about evolutionary psychology}: The argument from
consistency is too weak because, though it's true - indeed, true a
priori - that psychology needs to be consistent with the rest of
science, that doesn' t buy anything much that's of methodological
interest. In particular, consistency is not what "makes different
fields relevant to each other". To the contrary, mere consistency
is cheap; any two theories that are both true ipso facto achieve
it. If, for example, you are convinced that your favorite botanical
theory and your favorite astrophysical theory are both true, you
don't need anything further to justify assuming that they're
consistent. Likewise, the laws of quantum mechanics (if true) are
ipse facto compatible with the truth that Columbus, Ohio is bigger
than Urbana-Champaign. It does not follow that quantum mechanics
has much to say to demography or vice versa. Likewise, mutatis
mutandis, for your favorite theory about how the mind works and
your favorite theory about how evolution works.
Kovacs:
The example is not valid, because in the case of evolutionary
explanations of the mind, there is always a causal order; natural
selection caused the way the mind works at the time being. However, the
theory of quantum mechanics does not explain why Columbus is the bigger
city. Regardless of the example (he could have surely found more
relevant ones) the argument seems to me rather convincing.
Kovacs:
Finally I just would like to mention that not counting the example
above, Fodor's criticism of Cosmides' and Tooby's views of evolutionary
psychology is really interesting and itself alone worth reading the
book. In my opinion this part is great and is a must for everyone
interested in the topic. However, its essence, that different
conceptions of the mind shouldn't be connected a priori, could be in
part related to Fodor himself. Sometimes, as I tried to show above, he
himself rejects theories on the same basis. Though innate mechanisms
are consistent with computation, and both are parts of the rationalist
tradition, this has nothing to do either with modularity (see
Karmiloff-Smith for a not innate modularity) or why connectionist
network theories and empiricist solutions to the input-problem would a
priori be false. Horribile dictu, theories which smell like empiricism
could explain some facts about how the mind works while at the same
time others smelling like rationalism could explain others. After all,
juxtaposed traditions in the history of philosophy are not necessarily
consistent with how the mind works.
This archive was generated by hypermail 2b30 : Wed Jun 13 2001 - 18:38:04 BST