10.6.4—
Comparisons with other Research Programmes

Finally, my alternative approach yields a strategy for comparing "traditional" computational approaches to psychology that center on notions

of "rules" and "representations" with rival approaches arising from other sources such as neuroscience and neural network theories. For most of the time that the computer metaphor has been exploited in psychology, other approaches have also been explored, even if they have only recently been brought to the awareness of a broad audience of philosophers. For example, the Mathematical Theory of Communication of Shannon and Weaver (1949) has been explored by Sayre (1969, 1976, 1986) as an alternative basis for characterizing mental processes from the early 1960s to the present, and "neural network" approaches based on attempts to provide a mathematical characterization of the interactions of large numbers of cells in the brain were pioneered by McCulloch and Pitts (1943) and have been developed over the space of more than three decades by researchers such as Grossberg (1982) and Anderson (1973), as well as more recent researchers better known to philosophers such as Rumelhart and McClelland (1986) and Smolensky (1988).

Each of these approaches has its own preferred mathematical tools (sometimes including new mathematical machinery developed to solve particular problems) and its characteristic formally exact models of processes underlying cognition. Much of the debate between proponents of different models centers upon the features gained and lost by different kinds of mathematical apparatus: for example, the use of differential versus difference equations, or additive versus multiplicative shunting. (A good survey of mathematical differences in neural modeling is found in Levine [1991].) A second area of difference, both within the neural network camp and between its members and traditional artificial intelligence comes in the relationships assumed between the project of mathematical modeling and the project of connecting the model upwards to the data supplied by psychophysics and downward to that supplied by neuroscience. Some models are designed only to fit particular data curves, while others are intended additionally to be neurologically plausible.

In short, there is a great deal to be understood about the major research programmes in this area by looking at their mathematics, looking at their commitments to forming ties to other domains, and looking at their strategies for doing so. This kind of approach has some hope of shining light on individual theories, and also of clarifying the real differences between them. By contrast, most of what has come out of the philosophy of psychology with respect to neural networks so far has been centered on one of two issues: (1) whether connectionist theories are really the same as (or compatible with, or reducible to, or implementable

through) computational theories (and vice versa), and (2) how the availability of connectionist models vitiates Fodor's "only game in town" argument to the effect that we need intentional states because the only models we have to explain behavior require them as theoretical posits. Very little has been said of a philosophical nature about neural network theories in their own right, as opposed to how they compare to the kind of view espoused in CTM. Since I have argued that CTM does not in fact bear any philosophical fruit after all, I find comparisons with neural networks along that axis to be pretty much beside the point. Looking in detail at their mathematical repertoires and their commitments to kinds of interdomain connections in various directions, by contrast, gives us a concrete project in philosophy of science that we can sink our teeth into.