How to Compose Contents
A Review of Jerry Fodor's In Critical Condition: Polemical Essays on Cognitive Science and the Philosophy of Mind
Markus Werning
Department of Philosophy
Chair of Theoretical Philosophy
Heinrich-Heine-University Düsseldorf
Universitätsstraße 1, Bldg. 23.21
40225 Düsseldorf
GERMANY
werning@phil-fak.uni-duesseldorf.de
Copyright (c) Markus Werning 2002
PSYCHE, 8(07), September 2002
Previously held: http://psyche.cs.monash.edu.au/v8/psyche-8-07-werning.html
KEYWORDS: concepts, compositionality, systematicity, productivity, connectionism, language-of-thought, functionalism, semantics.
REVIEW OF: Jerry Fodor (2000) In Critical Condition: Polemical Essays on Cognitive Science and the Philosophy of Mind.
Cambridge, MA: MIT Press, $18 pbk. 229 pp. ISBN: 026256128X. [Note:
the review refers to the hardback edition of this book].
ABSTRACT: The paper critically reviews Jerry Fodor's book In Critical Condition: Polemic Essays on Cognitive Science and the Philosophy of Mind. It focuses on Fodor's compositionality arguments and their
relevance to the following questions: (i) How should concepts be
individuated? (ii) What has semantics to do with epistemology? (iii)
Who is right in the debate over classical and connectionist theories
of cognition? (iv) How can the semantic properties of a mental state
be inherited from the semantic properties of the state's
constituents? The paper finally argues that Fodor's opposition to
functional role semantics might jeopardize his view that semantic
compositionality requires appropriate constituent relations between
complex and less complex concepts.
If your research interests, in any respect, are about things with
semantic properties, such as language, thought, brains, or computers,
many books may concern your particular subject, but only few authors go
after the big issues. One of them is Jerry Fodor. Although he is a
professional philosopher, his writings bear heavily on disciplines as
unlike as linguistics and neurobiology, or literary theory and
artificial intelligence. The reason for this immense scope is his work
on what's called compositionality: a necessary property of all systems
that show a certain systematic correlation between content bearers. An
example for the systematicity of the human mind is the law-like
correlation between its capacity to think that a red square is in a
green circle and its capacity to think that a red circle is in a green
square. A systematic correlation of this sort can only be achieved if
the mind comprises primitive content bearers and some means to compose
complex content bearers from primitives in a compositional way, i.e., in
such a way that the semantic values of the complex state are determined
by, and dependent on, the semantic values of the primitive states plus
syntax. The semantic value of the complex concept A RED SQUARE IN A
GREEN CIRCLE is a function of the semantic values of the primitive
concepts RED, GREEN, SQUARE, CIRCLE, and IN as well as the syntax scheme
[(adjective noun) preposition (adjective noun)]. The primitive concepts
contribute all there is to their semantic value to the complex concept
and the complex concept derives all there is to its semantic value from
its primitive constituents. Symmetric considerations apply to brain
states, natural languages, and computer architectures insofar as they
are systematic. Compositionality also is required for representational
systems to be productive: A system can impossibly generate infinitely
many representations from a finite basis unless complex representations
inherit their semantic values exclusively from the semantic values of
the primitive constituents and unless the primitive constituents
completely contribute their semantic values to the complex whole. The
only exception to this principle are finitely many idioms. The
expression "red herring" is an idiom precisely because its semantic
value - which has to do with its referents to draw attention away from
the central issue - is not determined by the semantic values of "red"
and "herring".
In Fodor's recent book you learn why compositionality is the critical
condition for a hell lot of theories which up to date have been
dominating their disciplines. It is amazing how mercilessly the
requirement of compositionality is blasting off long-standing traditions
in the philosophy of mind and language, in psychology, and computer
science. Take only the view that the meaning of words is a function of
their inferential role (The following considerations, mutatis
mutandis, also apply to mental concepts). Many of the most influential
philosophers of the 20th century hold that the inclination to draw the
right inferences from sentences is constitutive for word meaning: In
this regard, not only Wittgenstein, Quine, Davidson, Rorty, and Dennett,
but also Heidegger, Derrida, and Saussure play in the same team. Fodor
(pp. 27-33) shows that this view ultimately depends on the
analytic/synthetic distinction, i.e., the distinction between truth of
meaning and truth of empirical fact. The reason is that complex
expressions often have inferential roles that cannot be derived from the
inferential roles of their primitive constituents. The expression
"British cow" may license the inference to "dangerous" if you believe
that British cows are dangerous - say because you believe they are
likely to be infected by mad-cow decease or foot-and-mouth decease.
However, neither "British", nor "cow" licenses any inference to
"dangerous". The inferential role of "British cow" is not determined by
the inferential roles of "British" and "cow". To reconcile the principle
of compositionality with the view, that inferential role constitutes
meaning, one cannot but fall back on the distinction between inferences
that are constitutive for meaning and those that are not. For,
compositionality requires that all that's constitutive for the meaning
of a complex expression is inherited from what's constitutive for the
meaning of the primitive constituents (plus syntax). Since the
disposition to draw the inference "British cow --> dangerous" is not
inherited from the disposition to draw inferences about cows and British
things, it cannot be constitutive for the meaning of "British cow". Even
though not all inferences are constitutive, there ought to be, at least,
some that are. Otherwise, inferential role semantics would lack any
justification. Which are the good candidates, then, for
meaning-constitutive inferences? Fodor wants to exclude the trivial
inferences of the sort "British cows --> cows", of course. What else,
then, is left than the analytic truths like "unmarried man -->
bachelor"? If you make this move, the view that meaning is constituted
by dispositions to draw inferences, at the end of the day, depends on
the distinction between analytic and synthetic truth. And this is a
distinction most people, following W. V. Quine (1951, 1960), think
cannot be made. Fodor's argument goes even further, though. He lays the
axe to the root of this misconceived view of semantics. The idea that
meaning is inferential role originates in the halfhearted way the
linguistic turn was made by most philosophers of the passed century.
Modern philosophy was mostly about epistemology whereas recent
philosophy is mostly about meaning (or content). The question "How can
we know that P?" was substituted by the question "How can we say (think
or judge) that P?" The link between both projects were so-called
transcentendal arguments: They started out epistemologically: "If it
weren't the case that P, we couldn't know that Q; but we do know that Q;
therefore P"; and ended up semantically: "If it weren't the case that P,
we couldn't say (or think or judge) that Q; but we do say (or think or
judge) that Q; therefore P." Those semantical question were regarded to
be about certain epistemic capacities like the kind of inference
subjects are to draw, or about what objects subjects are to recognize
under which conditions. This way semantics was contaminated with
epistemology. About one quarter of Fodor's book is the attempt to
dismiss this historical load. In this context he also argues against the
existence of recognitional concepts (pp. 35-62). Not even the concept
RED is constituted by the possession of recognitional capacities. The
view that there are recognitional concepts is extremely wide-spread and
deep-rooted in philosophy and psychology. It's probably the main
heritage of empiricism. Again, Fodor employs a compositionality argument
to undermine this view. The notion of a capacity to recognize something
as a soandso depends on the notion of a good instance of a soandso. The
capacity to recognize certain objects as red depends on the fact that
these objects are good instances of red. No capacity is so perfect as to
recognize all and only red things as red. Goodinstancehood does,
however, not distribute, as Fodor calls it: If something is a good
instance of red hair, it need not to be a good instance of red and of
hair, and usually is not. For, red hair is not typically, but only
relatively red. The capacity to recognize something as red hair, thus,
is not inherited from the capacities to recognize something as red and
as hair, respectively. If recognitional capacities were constitutive for
concepts, they, however ought to, because complex concepts do inherit
their semantic values from their primitive constituent concepts. If you
are pretty good in recognizing hair and red things, you may well be
rather bad in recognizing red hair. Since the possession of the concept
RED HAIR, according to the principle of compositionality, is determined
by the possession of the concepts RED and HAIR (and some syntactic
tools), and the capacity to recognize red hair, is not determined by the
capacity to recognize red things and hair, the possession of the
concepts RED, HAIR, and RED HAIR is independent from the possession of
recognitional capacities.
Compositionality also looms large in Fodor's criticism of connectionism
(pp. 81-125, pp. 91-111 are co-authored with Brian McLaughlin). Here,
his main opponent is Paul Smolensky (1987, 1988) who challenged Fodor
and Pylyshyn (1988). Before this challenge connectionism was widely
acknowledged to provide explanations of brain functions on a
sub-cognitive level. Smolensky, however, put an end to this modesty. He
claimed that connectionism was strong enough to contribute to the
cognitive level of explanation. In contrast to the chemical or the
neurophysiological level, the cognitive level is defined as the level at
which representations and processes over representations figure in
explanations. If connectionists want to contribute to the cognitive
level of explanation, they have to subscribe to the principle of
systematicity. Over the years Smolensky has developed various accounts
of how connectionist architectures might be considered representational.
According to an earlier version (Smolensky 1987, 1988), representations
like COFFEE and CUP WITH COFFEE are activity vectors over units which
represent micro-features (units like BROWN, LIQUID, MADE OF PORCELAIN,
etc.). Units are nodes with a certain value of activation. In a
connectionist architecture these units are connected to other units. The
connections carry impeding or exciting weights. According to Smolensky,
CUP WITH COFFEE representations contain COFFEE representations as
non-classical constituents in the sense of a component vector. The
method of composition is vector addition. Furthermore, COFFEE and
presumably every other representation is supposed to be
context-dependent. The activity vector that is the COFFEE representation
in CUP WITH COFFEE is supposed to be not identical to the activity
vector that contributes COFFEE in GLASS WITH COFFEE. Smolensky calls
this a weakly compositional account of representation. Although the
exposition of this theory is pretty theoretical, it closely resembles
what many neurobiologists have in mind when they talk of neuronal
assemblies. Smolensky, thus, expresses a still very popular view. The
problem with weak compositionality, as Fodor points out, is that it does
not render systematicity. Assume that somebody is capable of the
representation CUP WITH COFFEE AND GLASS WITH TEA. Systematicity implies
that s/he is also capable of the representation CUP WITH TEA AND GLASS
WITH COFFEE. That somebody who is capable of the first representation is
also capable of the second is, however, not warranted by weak
compositionality. For, somebody who is capable of the vector COFFEE in
the context of CUP WITH COFFEE need not be capable of the vector COFFEE
in the context of GLASS WITH COFFEE. Recall that COFFEE vectors need not
be identical if they occur in different contexts. Weak compositionality,
hence, does not suffice for systematicity. Connectionism thus understood
does not reside on the cognitive level of explanation.
To match this objection, Smolensky (1995) developed a completely new
model, which he calls integrated connectionist/symbolic architecture
(ICS) and which he takes to be strongly compositional. Here, complex
vector representations are constructed from primitive vector
representations in a way that employs vector operations and role
vectors. The story is rather complicated and need not be told here. For
our purposes only two features of ICS deserve attention: First, ICS is
at least syntactically compositional insofar as the form of complex
representations is determined by, and dependent on, the form of the
primitive representations. By the form of representation I mean the
pattern of activity that corresponds to the representation in question
and is expressed by a vector. Second, in ICS primitive representations
aren't proper constituents of complex representations. The relation of
proper constituency is understood as a necessary co-tokening relation: A
state C is a proper constituent of a state W if and only if,
necessarily, whenever W is tokened, C is tokened. The dispute between
Fodor and Smolensky, then, boils down to two questions: (1) Is syntactic
compositionality sufficient for systematicity? (2) Is proper
constituency necessary for systematicity? If the answer is yes to the
first and no to the second question, Smolensky wins the battle. If the
answer is no to (1) and yes to (2), Fodor comes off triumphant. In all
other cases, not only Smolensky, but also Fodor seems to be in trouble.
If there is a yes to both questions, Fodor's compositionality arguments
fall to pieces because he always has in mind semantic compositionality
when he talks of compositionality. If syntactic compositionality
suffices for systematicity, Fodor lacks the premise that semantic
compositionality is necessary. If both questions deserve a negative
answer, Fodor is left without a positive argument for his
language-of-thought doctrine. Since it is essential to this doctrine
that complex concepts have primitive concepts as proper constituents,
there will be possible alternatives to the language-of-thought doctrine
if proper constituency is not necessary for systematicity.
Unfortunately, Fodor does not really articulate both questions. Though
he is a little bit more explicit about the second question, he is
altogether beating about the bush. Nevertheless, do I think that Fodor
wins. The reason for my optimism is that systematicity requires
semantic, rather than syntactic compositionality and semantic
compositionality requires primitive representations to be proper
constituents of complex representations. Let me first explain the
difference between semantic and syntactic compositionality. Consider the
following mappings between English expressions:
(i) ("unmarried", "man") --> "unmarried man",
(ii) ("unmarried", "man") --> "bachelor",
(iii) ("red, "herring") --> "red herring".
Number (i) is both syntactically and semantically compositional because
the form as well as the semantic value is conveyed through the mapping.
The mapping (ii) only is semantically compositional, provided that the
sentence "Unmarried men are bachelors" is analytic. The word "bachelor"
derives it semantic value, not its form, from the expressions
"unmarried" and "man". Finally, (iii) is syntactically, but not
semantically compositional because "red herring" is idiomatic.
Consider the two thoughts that the man with a red coat was distracted by
an old herring and that the man with an old coat was distracted by a red
herring. The thoughts are syntactically composed of the same concepts.
However, they aren't nomologically correlated. Someone who is capable of
thoughts about red coats and old herrings may well be unable to think of
red herrings. This is because the representation of a red herring fails
to be semantically composed of the representations of which it is
syntactically composed. The concept RED HERRING is idiomatic. Syntactic
compositionality does, thus, not suffice for systematicity. The answer
to question (1) is no. What's required for systematicity, apparently, is
semantic compositionality.
To win the battle against Smolensky without casualties, Fodor cannot set
his heart at rest, already, but has to argue further that proper
constituency is required for systematicity. A possible argument, which,
however, contains a pill to swallow for Fodor, runs as follows.
(I) The semantic value of a representational state is dependent on
nothing but the causal role of the state.
(II) The causal role of a state is dependent on nothing but the causal
roles of its proper constituents (and their relation to each other)
Hence:
(III) If the semantic value of a state A is dependent on nothing but the
semantic values of the states B and C, then B and C are proper
constituents of A.
The argument is valid. Its second premise is backed by deep-rooted
metaphysical intuitions. They even apply to microphysics where the
causal role of molecules is dependent on nothing but the causal roles of
their constituent atoms (and their relation to each other). The first
premise, however, is equivocal to functional role semantics (FRS), a
doctrine people like Fodor "who hate FRS root and branch" (p. 71) are
unlikely to use in arguments of their own. I do, however, not see how
the conclusion can be defended without using functional role semantics.
Maybe it's this threatening inconsistency that lets Fodor beat around
the bush in his argumentation against Smolensky. Notice that Fodor
really is in need for a good argument in favor of constituency relations
as a requirement for systematicity, and hence compositionality. If there
wasn't any argument of this sort, to which Fodor can consistently
appeal, the language-of-thought doctrine would come under fire. What's
going wrong here, probably, is an equivocation about what functional
role semantics means. Fodor uses this term both for inferential role
semantics and causal role semantics (cf. p. 70). We have already seen
that inferential role semantics does not comply with compositionality.
Whereas the causal role of a representational state determines its
inferential role, the inferential role need not determine the causal
role. There is more to the causal role of a representational state than
the fact that this state figures in dispositions to draw inferences.
Representations may cause all sorts of things: emotions, behavior, an
increase of heart beat etc. They may also be triggered by things other
than representations: retinal stimuli, acoustic waves, dopamine, etc.
Inferential role and causal role aren't isomorphic to each other.
Furthermore, causal role, apparently, composes well, whereas inferential
role, as we've seen, doesn't. The causal role of a complex state is
entirely determined by, and solely dependent on, the causal role of its
proper constituents and their relations to each other.
Fodor himself, from time to time, favors an externalist theory of
semantic value, which has causal elements, too, but only external ones.
His story goes something like this (cf. p. 63): The fact that a certain
internal state - call it DOG - represents dogs as dogs is determined by,
and dependent on, the fact that DOG stands in a certain causal or nomic
(i.e., law-like) relation to dogs. There are well know difficulties with
this position (Why aren't SMOKE and FIRE synonymous, although, by causal
law, there is smoke whenever there is fire and vice versa?). Those
complications may, however, be accounted for somehow. The problem I want
to highlight has to do with compositionality. Does nomic connectedness
distribute? Assume that not only DOG is nomologically connected to dogs,
but that also ROBOT is nomologically connected to robots. Does this
necessitate that ROBOT DOG is nomologically connected to robot dogs? The
laws that connect DOG to dogs and ROBOT to robots are at best
probabilistic. No biological system is so reliable that it causally
connects all and only dogs to the internal state DOG. No landscape is so
unruffled that it could not, in principle, hide a dog from your
perceptual apparatus. Probability, however, does not distribute
properly. Let P(SOANDSO) be the probability of a certain system to
trigger the internal state SOANDSO if and only if there is a soandso in
the system's environment (You may well substitute the logical constant
"if and only if" by "if" or by "only if" or some other logical
operation, e.g., counterfactual dependency). For the connection between
an internal state and an external object to count as nomic or causal the
corresponding probability should exceed a certain limit value L. Having
said this, it follows from mathematics that probability does not
distribute properly. The reason is that the probability of a conjunction
is usually lower than the probabilities of each conjunct. Therefore,
P(ROBOT DOG) is smaller than P(ROBOT) and P(DOG). The more conjuncts you
have, the less reliable the connection between the concept conjunction
and the conjunct object gets. It may well be that P(ROBOT) and P(DOG)
exceed the limit value L, but P(ROBOT DOG) doesn't. In fact, one can
always find a conjunction of internal states/concept C(1) ... C(n) such
that P(C(i)) exceeds L for each index i, whereas P(C(1) & ... & C(n)) is
lower than L. The externalist theory of semantic value is in a row with
inferential role semantics, the recognitional theory, and other theories
not mentioned here (e.g., prototype theory) insofar as it fails to
satisfy the compositionality requirement. The best candidate to comply
with compositionality seems to be causal role semantics, after all. But
Fodor won't agree.
In Critical Condition is a collection of reviews, journal articles,
and replies most of which have been published elsewhere within the last
decade. Thematically it is more heterogeneous than this review reveals.
Apart from compositionality, the book deals with metaphysical issues,
the innateness of concepts, the modularity of mind, and evolutionary
theory. You will also learn why we don't think in English but in
Mentalese, and whether science is biologically possible. Most of the
time, Fodor either attacks or replies to opponents, giving the book a
ping-pong format. Since many chapters originally address an either lay
or interdisciplinary audience, you need not be a heavy-duty philosopher
to follow the course of argument. Aside from philosophy, Fodor's talent
is humor. I haven't read a book written by a professional philosopher
that tells more jokes per page. This reflects the sort of pessimism the
title conveys: "Nietzsche is righter than Brecht: Sometimes the man who
laughs has heard the terrible news" (p. x).
References
Fodor, J. A. & Pylyshyn, Z. W. (1988). Connectionism and cognitive
architecture: A critical analysis. Cognition, 28, 3-71.
Quine, W. V. (1951). Two Dogmas of Empiricism. The Philosophical
Review, 60, 20-43.
Quine, W. V. (1960). Word and Object. Cambridge, MA: The MIT Press.
Smolensky, P. (1987). The Constituent Structure of Mental States: A
Reply to Fodor and Pylyshyn. Southern Journal of Philosophy, 26,
137-160.
Smolensky, P. (1988). On the proper treatment of connectionism. The
Behavioral and Brain Sciences, 11, 1-74.
Smolensky, P. (1995). Connectionism, constituency and the language of
thought. In Macdonald, C., & Macdonald, G. (Eds.), Connectionism.
Cambridge, MA: Blackwell.
