1. We’ve been discussing, in very broad terms, the class of computational architectures known as connectionist architectures.
2. In suggesting some of the arguments for connectionism, we’ve already hinted at one interpretation of the relationship between connectionist and so-called ‘symbolist’ or classical architectures that work by manipulating symbols or representations. This is that the two are:
(a) Rivals: Connectionist and classical architectures are rivals for explaining the way the mind works.
Others, however, believe that the two work at:
(b) Different levels of explanation: Thus connectionist architectures may offer a model of the physical or implementation level while classical or symbolist architectures may operate at the computational level. Notice that this explanation requires that the connectionist architecture implement the classical one.
A third alternative suggests that the two are true of:
(c) Different domains of cognition: Connectionist pattern recognizers may be true of some domains (e.g. face recognition or vision), while classical symbol manipulators may be true of others (e.g. reasoning).
3. Connectionist advocates of (a) and (c) have met with a wide array of criticisms. Here is a gloss of a few:
i. Connectionists do not offer a deep explanation of constraints on the space of possible architectures:
Various cognitive domains seem to exhibit nonaccidental constraints. For example, there are no languages that form the past tense by reversing the order of phonemes in a verb. However, connectionist pattern recognizers would seem (in virtue of not employing local representations) to be indifferent between this way of forming past tense and those employed in actual language. Thus, on connectionist architectures, it seems like an accident that no such languages exist.
ii. Arguments
from Systematicity and Productivity:
These are two concepts you’ll remember from our discussion of Fodor’s arguments for the language of thought hypothesis. Productivity, remember, suggests that one’s linguistic capacity can, in principle, produce and understand an infinite number of utterances. Systematicity, recall, suggests that there are systematic relationships between some thoughts/utterances and others, such that a language that could produce one could then produce another. E.g. any thinker that can think “John loves Mary” can think “Mary loves John.” On a symbolist story, this relationship is easy to explain. Since thoughts are comprised of symbol tokenings, and the symbols have discrete meanings, a thinker that can think ‘John’ in the subject position can also think ‘John’ in the object position. Again, critics claim, that thoughts would exhibit systematicity emerges as an accident on the connectionist view.
4. One may now wonder how good the preceeding anti-connectionist arguments are. To begin with,
(A). It is
possible to have a classical architecture that forms past tenses in ways not
actually found, or that fails to exhibit systematicity or productivity.
(B) The kind of
explanation of the phenomena critics seem to want is rooted in the structure of
thought. However, other kinds of
explanation may be possible. E.g.
Braddon-Mitchell and Fitzpatrick suggest evolutionary explanations may be
viable.
(C) Are thoughts
really as systematic as Fodor and Pylyshyn suggest? For other animals, it seems like the answer might be ‘no’. That is, it seems plausible to imagine that
while a lion can think “I want to eat the rat,” it might be impossible for it
to think “the rat wants to eat me.”