Take Home Exercise for Philosophy of Mind and
Cognition
(b)Critically evaluate the arguments for the
Language of Thought and/or the Map Theory
In this essay, I examine the
arguments for the several stances it is possible to take in the ‘debate’ (largely
implicit) between the Language of Thought Hypothesis and the Map Theory of
cognition. My ultimate conclusion is that, whilst so many fundamental issues
are still extraordinarily unclear, there is one particular intermediate stance in this debate that has significantly more
weight behind it than any other.
I first discuss how, for
both hard-line connectionists (those who think that the brain undergoes no ‘classical
computation’) and hard-line Fodorians (those who think the key parts of cognition
will be explained only by classical computational models), the stakes between
the “Language of Thought Hypothesis” and “Map Theory” of cognition are quite
clear: if the LoTH is the right theory of all cognition, then the hard-line
connectionists are completely wrong, and if complex cognition occurs without
any real LoT, then the adherents of Fodor’s computational-representational
theory are completely wrong. I argue that both these poles are probably wrong,
but that the connectionist extreme is much less implausible than the Fodorian
one. I secondly explore the intricate intermediate position that that all (or
nearly all) cognition in non-human animals involves connectionist ‘software’,
captured by “Map Theory”, and yet that the Language of Thought theory
represents a classical ‘program’ run only by humans. This thesis has not been
actually expounded in any literature I am aware of, although it seems to me to
be the hypothesis that has the most weight behind it of all. As I argue, it is
highly concordant with Noam Chomsky’s carefully considered hypothesis (since
the mid-1990s) for the origin of the human language faculty/explanation for the
“Great Leap Forward”; the notion that productive thought is bound up with
productive language is supported by evidence from developmental psychology; the
notion that the vast majority of biological ‘software’ is connectionist is
strengthened by well-known considerations about the ‘hardware’ of the brain and
the nature of evolution; and finally, despite Fodor’s many protestations, it
seems to me that Prototype Theory is the best account of the nature of our
concepts, and Prototype-structured concepts are better explained in terms of
“Map Theory” than the “Language of Thought”.
In The Philosophy of Mind and Cognition, Frank Jackson and David
Braddon-Mitchell are careful not to hitch the Map Theory to the connectionist
programme in AI, or the Language of Thought to more ‘classical models of
cognition’. They note that the truth is much more complex and that it may often
be unclear whether a ‘classical’
program is being implemented in a connectionist substrate (indeed, they claim
that there might be no “source code” for the brain at all) [Braddon-Mitchell
& Jackson, 2007: 219]. Nevertheless, they do suggest a certain harmony
between connectionism and the “Map Theory”, and between classical computation
and Fodor’s LoT. This makes perfect sense, because a hard-line connectionist
cannot possibly accept the LoTH, and the Map Theory is explicitly proposed as
the one alternative to the LoTH.
Paul and Patricia
Churchland, the influential reductive neuro-computationalists, are probably the
most prominent proponents of what I’ve called ‘hard-line’ connectionism. For
several decades, the central philosophical claim of these two
neuro-philosophers has been that classical cognitive science is (and always has
been) on the wrong track, because the brain’s hardware is connectionist[1] and
we can only understand the brain’s ‘software’ by investigating the nature of
the hardware, rather than constructing abstract, simple models of rules and
representations which may have no “psychological reality”. In numerous articles
and papers from the late 80s onwards, the Churchlands have argued that evidence
from neurobiology[2]
makes it clear that the Language of Thought (which they tend to lump together
with ‘Folk Psychology’ as part of the one Fodorian package), even if apparently explanatory, simply doesn’t
exist [P.S. Churchland, 1986; P.M. Churchland, 1989; P.S. & P.M.
Churchland, 1990; P.S. Churchland & Sejnowski, 1990].
It is crucial to point out,
of course, that there was always a key weakness in the Churchlands’ critiques
of classical cognitive science and the LoTH – a weakness that would ultimately allow
Fodor to evade their attacks without working up too much of a sweat at all.
This key weakness was that the Churchlands didn’t attempt to provide a serious
alternative, connectionist-based theory
of thought (as opposed to mere detailed accounts of bio-chemical processes
correlated to certain kinds of cognition) to the one they so roundly rejected. The
absence of such a theory in the Churchlandian canon meant that it remained
justified for Fodor to simply repeat the famous remark he made in his seminal
1975 work: that the LoTH is “the only game in town” [1975: 406]. And as Fodor
and Pylyshyn pointed out in their influential 1988 critique of connectionism
(which includes extensive reference to the Churchlands), without such an
alternative account of the actual nature of thought,
the neurobiological evidence marshalled in supposed contradiction of the LoTH cannot nullify the
abductive argument, because the neurobiological details do not in themselves itself
constitute an explanatory science of the mind [Fodor & Pylyshyn, 1988]. As
Fodor and Pylyshyn emphasise in this same article, no connectionist can explain
productivity and systematicity except by creating “Classical architecture” in
connectionist models, yet these are the two key features of thought that form
the basis of Fodor’s argument to begin with [Fodor & Pylyshyn, 1988: 33-40].
Fortunately for the
connectionists, however, Frank Jackson and David Braddon-Mitchell’s “Map
Theory” does provide such an alternative account of the ultimate nature of
thought. Evidently, if such a theory is at all tenable, it lends significant
credence to the hard-line connectionists, because the very fact of its
existence destroys Fodor’s key abductive argument. And unfortunately for Fodor, it does indeed seem to me that the Map
Theory is plausible as an alternative theory of thought – in fact, highly
plausible. The idea of thought as involving highly structured, non-sentential, ‘nth’-dimensional
map-like representations actually possesses a number of virtues over the LoTH: it
is (at least intuitively) much more neatly reconcilable with connectionist
networks and the ‘hardware’ of the brain, since a map is an inherently
‘distributed’ structure; it gives a far more plausible account than the LoTH of
how it is that our beliefs about people, places and events can so fluidly
change as we have new experiences and take in new information (our mental
‘maps’ are simply updated); it seems to explain memory and memory-retrieval in
a much more plausible way than a Language of Thought, since it is not necessary
to claim that we have a vast number of propositions stored in our memory which
we can retrieve at will; it would appear to explain certain weaknesses we have
in formal reasoning (for example, why content seems to matter to our ability to
carry out tasks of reasoning with formally identical structure [Jackson,
Braddon-Mitchell, 2007: 235]); it appears to have evolutionary considerations
in its favour over the LoTH, since a Language of Thought is a highly elegant
computational system which seems at odds with the haphazard makeshift nature of
evolution (it seems unlikely that a LoT would be the kind of system to slowly
crystallise, and it seems highly unlikely that our ancient Cambrian ancestors
had a Language of Thought, whereas they might have had primitive ‘map-like’
representations); and it can account (seemingly) for “systematicity”, since the
‘places’ in the cognitive maps can be switched. All in all, therefore, the
Churchlandian position is, in my view, massively strengthened by the “Map
Theory”.
With that said, of course, the
reason that I do not think that the hard-line connectionist position is likely
to be right is that the Map Theory,
even if it does seem to account for systematicity,
it doesn’t quite answer Fodor’s productivity
criterion. As a result, it cannot explain the apparent infinite generative
capacity of human cognition – our ability to think an infinite array of discrete
thoughts by finite (combinatorial) means. As I am about to suggest, however, it
may very well be that only humans
have this full-strength productivity. This would imply that the Map Theory
might be a far more general account of biological cognition, and that the
hard-line connectionist position might be the correct one for all creatures on
Earth except us. As I’m about to argue, I think this is actually the most
plausible stance of all.
Perhaps the major consideration
in support of the view that only Homo
sapiens has a LoT in Fodor’s sense (with a combinatorial syntax which
allows for infinite productivity) is that it accords with Chomsky’s hypothesis
about the evolution of the language faculty. Whilst undoubtedly a highly
controversial hypothesis – disputed even by other notable generativists (for
example, Ray Jackendoff and Steven Pinker [2005]) – the conjecture does have
nontrivial theoretical and empirical backing.
Chomsky largely eschewed
speculation about the evolution of the language faculty for the first few
decades of the Universal Grammar research programme, but he has become quite
vocal in espousing his ‘spandrelist’ hypothesis about the evolution of the
“Faculty of Language” ever since he wrote The
Minimalist Program in 1995. The central idea of Chomsky’s “Minimalist
Program” is that it is possible to distil the ‘Universal Grammar’ into one basic
computational operation called “Merge”, with two forms, “External” (for
separate objects A and B) and “Internal” (for two objects where at least one
object contains the other) [Chomsky, 1995]. This lends credence to this
spandrelist hypothesis because of the elegant computational simplicity of such
a system (“like a snowflake”), and the fact that this computational procedure
subserves thought as fundamentally (or more fundamentally) as it subserves
language.
The most rigorous defence of
this evolutionary hypothesis appeared only recently, in a 2015 book Chomsky
co-authored with the MIT computer scientist, Robert Berwick, entitled Why Only Us? : Language and Evolution. In
this short work, Chomsky and Berwick specifically defend the thesis that this
“Merge” operation first originated by a single mutation in a human individual
around 80,000-100,000 years ago, which gave rise to recursion and hierarchical
structure in that individual’s thoughts [Chomsky & Berwick, 2015]. Whilst
they don’t invoke Fodor explicitly, it is clear that they see this as the
moment the “productivity” property key to Fodor’s LoT actually emerged (for the
first time in the history of life on the planet). They hold that this capacity
would have had some selectional advantage, and would have spread throughout the
population before being secondarily externalised [Chomsky & Berwick, 2015].
Much of the book is taken up by arguments in defence of an anti-adaptationist,
‘messy’ view of evolution as involving “stochastic effects” and both gradual
and sudden developments, with more at work than simple natural selection in simple
genetically varied populations. However, there are also a number of positive
considerations adduced in favour of the hypothesis, many of them (in my view) highly compelling.
Firstly, recursion appears
to be a property that is either absolutely present or absolutely absent, so seemingly
couldn’t come about as a gradual adaptation [72]. Secondly, there is strong
evidence (disputed by some) that no other species on the planet is capable of
recursive communication – birdsong, they claim, never gets further than “linear
chunking” or iteration of “motifs” [142]. Thirdly, the fact that human linguistic
externalisation is modality-independent (sign language has the same level of
syntactic complexity as spoken language, and is learnt as rapidly by children
in the right environment) seems to constitute good evidence that the Faculty of
Language was first a Faculty of Thought [74]. Fourthly, the sudden emergence of
productivity around 80,000 years ago appears to be one of the best possible
explanations for what Jarred Diamond famously called the “Great Leap Forward”
[37]. Fifthly, the extent of language variation in the world can seemingly
explained by the fact that the means of externalisation did not actually
co-evolve with the core computational system, but are instead much more ancient
systems being co-opted for the task [82]. Sixthly, the evolutionary timeframe
of the change is certainly not ruled out by the reconstructed genomic evidence used
to compare Homo sapiens with its ancestors and cousins [Chapter 4]. And, finally,
there is a neuro-anatomical hypothesis for what actually happened to create the
key “Merge” mutation – the dorsal and ventral “fiber tracts” linking different
language-related areas of the brain formed a ring (which does not exist in
birds) [Chapter 3].
There is also evidence from
developmental psychology that lends support to the idea that language is bound
up with a special form of thought that marks humans out as unique (and this
naturally holds even if Chomsky and Berwick’s hypothesis is a fair way off the
mark in terms of timeframe and suddenness of evolution). As Antoni Gomila notes
in a critical review of Jerry Fodor’s 2008 book The Language of Thought Revisited, “there is no evidence for the
systematicity and productivity of thought before the development of language,
at the end of the second year” [Gomila, 2011: 151]. Gomila cites the child
psychologist Elizabeth Spelke for this claim, whose experiments suggest that babies
are born with modular packets of knowledge of particular contents (physical,
numerical, biological, intentional), but that such knowledge does not admit of
productive combination before language – instead it is context-dependent and encapsulated
[Spelke, 2003]. It seems that the role of language in development is to create
interaction between these different modules.
Another major virtue of this
idea that the LoT is human-specific is that it means we can still maintain the
benefits of a largely connectionist view of biological cognition (those
benefits espoused by the Churchlands, in particular, general biological
plausibility (in terms of physiology and evolution)), and the advantages of Map
Theory as a general account of biological cognition. It only means that we have
to say that humans may be the only species with a ‘classical program’.
Finally, adopting the Map Theory
as an account of the majority of human thought also seems to work far better
for explaining the fuzzy nature of human concepts, described by “Prototype
Theory” in cognitive linguistics – the research programme pioneered by Eleanor
Rosch’s studies about ‘categories’ in the 1970s. Map Theory would seem to
suggest that concepts are not discretely structured, but form part of wider
cognitive architectures. If one accepts the tight connection between
connectionist software and Map Theory, then there is certainly reason to think
that Map Theory is highly compatible with Prototype Theory since, as Gomila
notes, connectionist models which are based on Prototype Theory have met with
increasing success since the 1990s [2011: 149].
Overall, I think that the debate
between Fodor’s Language of Thought Hypothesis and the “Map Theory” of
cognition raise deep and interesting questions about the nature of the mind
which are not currently soluble – although some possibilities seem distinctly
more likely than others. I think that the LoTH is extremely unlikely to be a
true theory of all cognition, and thus that the ‘hard-line’ Fodorians are
seriously misled. I think that the Map Theory, most naturally joined with
connectionist models, is far less implausible as a general account of the
cognitive architecture of biological creatures – however I think it fails to
account for productivity in humans. These considerations have ultimately led me
to the view that the most likely hypothesis is that Homo sapiens is the only species to possess a Language of Thought,
implemented as a ‘classical program’ on top of the connectionist hardware and
software that all biological organisms probably possess.
Naturally, I acknowledge
that this claim, despite the justifications I have given for it, is still an extremely speculative one. I am also
aware that I have an aesthetic bias
for this view: I think it is a beautiful notion that humans might be the only
creature with a classical program that was key for our success in colonising
the planet and reaching civilisation… So it might be totally wrong.
Reference List
Braddon-Mitchell, D and
Jackson, F. (2007), The Philosophy of
Mind and Cognition, 2nd edition, Blackwell Publishing.
Chomsky, N. (1995), The Minimalist Program, The MIT Press,
Cambridge, M.A.
Chomsky, N & Berwick, R.
(2015), Why Only Us? : Language and
Evolution, The MIT Press, Cambridge, M.A.
Churchland, P.M. (1989), A Neurocomputational Perspective: The Nature
of Mind and the Structure of Science, The MIT Press, Cambridge, M.A.
Churchland, P.M. and P.S.
(1990), Could a machine think? Scientific American 262
(1):32-37.
Churchland, P.S. (1986),
Neurophilosophy: Towards a Unified Science of Mind-Brain, The MIT Press,
Cambridge, M.A.
Churchland, P.S and Sejnowski,
T. (1990), “Neural representation and neural computation”, Philosophical
Perspectives 4:343-382.
Fodor, J. (1975), The Language of Thought, Harvard
University Press, Cambridge, M.A.
Fodor, J and Pylyshyn Z.
(1988), “Connectionism and Cognitive Architecture: A Critical Analysis”, Cognition 28
(1-2):3-71.
Gomila, A. (2011), “The
Language of Thought: Still a Game in Town?” Teorema: Revista
Internacional De Filosofía, 30(1), 145-155.
Jackendoff, P and Pinker, S.
(2005), “What's special about the human language faculty?” Cognition 95
(2).
Spelke, E. (2003), “What
Makes Us Smart? Core Knowledge and Natural Language”, en Gentner, D. & Goldin-Meadow,
D. (eds.): Language in Mind. Advances in
the Study of Language and Thought, The MIT Press, Cambridge, M.A., 277-312.
[1]
The most general and simplistic reason for thinking this is that sensory
neurons can be easily analogised to the input nodes in a connectionist model,
output nodes can be analogised to the motor neurons and the hidden nodes can be
analogised to the web of neural connections in our nervous systems
[Braddon-Mitchell & Jackson, 2007: 223])
[2] The
most straightforward data included the “distributed representation” used by the
brain (held to be strong counter-evidence of the existence of discrete symbols)
and the fact of “graceful degradation” (certainly counter-evidence of digital hardware).
No comments:
Post a Comment