@Article{Guarini2009,
author = {Guarini, Marcello},
title = {Computational theories of mind, and Fodor's analysis of neural network behaviour},
journal = {Journal of Experimental \& Theoretical Artifical Intelligence},
year = {2009},
volume = {21},
number = {2},
pages = {137--153}
}
The article can be found here.
Fodor has argued that classical and connectionist approaches to modelling cognition will have difficulty with performing global processes. A global process is a process where a representation might have to interact with a large number of representations, for example, in belief revision (when adding a new belief to our collection of beliefs, how do we decide which of our many beliefs should be affected by the new one?). Furthermore, he believes that there are more problems with the connectionist approach than the classical approach. Guarini argues that Fodor's concerns with the connectionist approach are not warranted because they are based on an outdated and incorrect understanding of neural networks (p. 137). Guarini also points out, but does not fully explore, other papers that argue against Fodor's claim that global processes are difficult to computationally model (p.150).
Guarini first provides a small example of a neural network that identifies certain actions as morally acceptable or unacceptable (p.139). In this network, the 'theory' of what is right and wrong is stored in the synaptic weights of various nodes. Changing even a single belief would involve changing the synaptic weights of the entire network. However, Guarini argues that what is more important is that even though every synaptic weight in the network will have changed, the theory that it represents will have only changed in one respect (it will have one changed belief).
Next, Guarini argues that Fodor makes several problematic assumptions (p.142), all of which might be based on the assumption that all connectionist networks are ultralocal (p.145). An ultralocal network is a network where individual nodes represent propositions, and the only way to change meaning in an ultralocal network is to add a node. Furthermore, in an ultralocal network a theory would be a combination of nodes, or propositions. Fodor thinks it is a problem that adding a single proposition to the network (like a new belief) will require changing the meaning of everything else in the network. Guarini argues that Fodor does not properly distinguish between “node as part of network” and “node as part of syntactic representation” (p.146). Guarini argues that nodes to not need to be syntactically or semantically interpreted, so it is possible to add a node to a network in such a way that it would not affect every single distributed syntactic representation. Guarini cites several recent examples of connectionist networks that represented propositions as patterns of distributed activation across an entire neural network (p.147). He admits that he has not shown that connectionist networks can necessarily succeed on a large scale at performing global processes, but that he has dealt with Fodor's argument that, in principle, connectionist networks cannot succeed, and so work on connectionist networks should continue.
Fodor also argued that researchers using connectionist networks are interested in modelling total theories ('the totality of one's epistemic commitments', or everything that one believes). Guarini replies that this is simply not the case; while connectionist networks model theories, they are domain specific, and thus not subject to the criticism that total theories are too large to be computational units (p.150). Guarini suggests that many domain specific theories operating at the same time could do the work of a total theory. Guarini also suggests that cognition might be partially linguistic and classical, as Fodor envisions it, and partially governed by domain specific networks. Guarini closes the article by admitting that Fodor's concerns about large-scale global processes have merit, but that restricting the domains of connectionist networks is enough to allow research into them to continue.