Andrews (1995) in discussing the paper of Latimer (1995) has argued that psychological models ``are metaphors rather than realistic descriptions of cognitive processes''.
Wiles (1995) distinguishes symbolic (traditional) and connectionist (neural net based) modelling by analysing the primitives used in each approach and the methods used to combine them.
I argue here that psychological modelling is indeed closely tied to the availability of appropriate metaphors, and that the chief value of connectionist research for psychology lies in the discovery of new metaphors. These new metaphors will include Wiles' primitives and the methods for dealing with them (including but exceeding their combinational principles).
As one develops more and more detailed models, the metaphor becomes more and more like the application domain, but the distinction remains. Even the mathematical modelling of well-understood physical processes involves a model which is quite distinct from the physical domain. For example the model often includes idealisations like the assumption that all of a body's mass is located at a dimensionless point.
Reductionism provides an extreme example of the need to distinguish the application domain from the metaphor: Assuming that all of the behaviour which is the subject of psychology arises from electro-chemical activity of the nervous system, then in principle physiology could provide us with a model. However, working with this model would not be doing psychology at all! Psychology must be about higher level processes that emerge from neurophysiology. At the moment our best means of theorizing about these processes is through the use of metaphor.
It has been argued, for example by Vroon and Draaisma (1990) that the history of psychology closely parallels the availability of metaphors. For example, in the eighteenth and early nineteenth centuries, the mechanistic view of the universe prevailed, and virtually the only metaphors available were mechanical. Typical of this time is Lamettrie's (1748) work L'Homme machine and the early studies of physiologists. The wide availability of steam power in the late nineteenth century coincided with the genesis of psychoanalytic theory, and even in the early work of Breuer and Freud (1974) the notion of a contained source of energy which needs to be released is apparent. Vroon and Draaisma present in detail a model in which the id corresponds to a boiler, an undirected reserve of raw energy; the ego corresponds to a regulator, which allows the energy to be channelled; and so on. And the importance of catharsis is readily apparent.
Continuing the argument of Vroon and Draaisma (1990), a telephone exchange, viewed as a black box, can be seen as a motivational metaphor for behaviourism; the availability of radio and radar paralleled the rise of parapsychology with notions of waves, transmission, coding and noise; and along with the computer revolution came modern cognitive psychology with models that either share many notions with computer science or are indeed described by programs.
Incidentally, some measure of the significance of these metaphors can be gained by seeing how frequently they have found their way into the common language: At the mechanistic level ``you can see the wheels turning'' or ``you can almost hear the ticking'' (about someone thinking); ``stoke the boilers'', ``build up a head of steam'' and ``letting off steam'' for the steam power level; at the radio level ``are you on my wavelength'' and ``he's sending out signals'' and so on. Notice that these are not idiomatic phrases --- the words are quite variable. Rather they are metaphors that are sufficiently widely established that we hardly even notice their use.
This completeness can be misconstrued as suggesting that we no longer need new metaphors --- whatever metaphor you might like to think of, if it's computational feasible at all, then it can be computed on an ordinary computer. So why not use an ordinary computer for all your modelling tasks?
In fact even the field of computer science is busily searching for new metaphors. We now have the technical ability to build reasonably cheaply computers which are capable of massive parallel processing (single machines with thousands of processors). However, such computers are proving to be of little value (except in fields where there is massively parallel data that can be processed identically) because of the difficulty of programming them to take advantage of the resources.
All of our traditional paradigms are based on sequential algorithms. Indeed our language is appropriate for describing sequential algorithms, and it provides some constructs to allow us to describe a small number of processes acting concurrently and interacting occasionally, but there is little understanding of how to deal with large numbers of processes interacting continuously. Even finding appropriate models for a modest number of interacting processes is currently exercising many computer scientists, see for example Pratt (1986), Rodriguez (1993), Pnueli (1981) and papers in Bakker, Roever and Rozenberg (1989).
Thus, computing, which has been a valuable source of metaphors, is currently ill-equipped to provide metaphors for the so-called emergent properties of massive interaction. So where might we find such metaphors? It seems that at least some relevant metaphors will come from connectionism.
I will not record here evidence for this claim since ample evidence appears in Wiles' catalogue of connectionist primitives (Wiles, 1995). But perhaps it is worth noting some obvious examples of metaphors which connectionism has shown us can arise in massively parallel systems: generalisation, content addressable memory, graceful degradation, self organization, etc. Of course these are the metaphors for which we have reasonably brief linguistic descriptions, but there are other metaphors which will only become accessible to us through connectionist analysis (the example of Liapunov functions is mentioned below).
The understanding of a connectionist model depends on obtaining some overall view of the model and its relationship to the input-output behaviour of the net. Examples of mathematical tools that have been used for this purpose include Liapunov functions and the analysis of error surfaces. These can be seen in detail in the work of Hopfield (1984) and the PDP group (Rumelhart & McClelland, 1986) where they are used to engineer nets with specified properties.
The precise analysis of error surfaces is still in its early stages but it is progressing well. Interesting new results have been obtained at Macquarie by Hamey (1994) and the hope is that a sufficiently detailed mathematical treatment will free modellers from at least some of the details of network architecture and allow them to focus on higher level properties (metaphors).
A further example of new non-primitive metaphors arising in connectionist research is the discovery of the nature of certain self-organizing nets by Kohonen (1982). Kohonen's nets were shown to organize themselves into units that could be described as centre-surround and orientation-selective in analogy with known properties of mammalian visual systems. This is important because it now seems probable that such neuronal arrangements are important in other parts of the brain rather than being merely specific properties of the visual system. This would suggest that a proper understanding of the nature of such networks may yield valuable metaphors for analysing other cognitive processes.
The question remains: What use should working psychologists make of connectionist models? Here I am rather less positive in my outlook. There is a real risk that the construction of connectionist models of psychological processes will be of immediate value in the development of connectionism rather than of psychology. This development is important, but psychology can't afford to have good psychologists stray too far from the major problems of their own discipline.
Nevertheless, cognitive psychologists should aim to stay well-informed about connectionist research and connectionism needs the interaction with psychologists to develop appropriate metaphors, so symposia such as this are very valuable.
de Bakker, J., de Roever, W., & Rozenberg, G. (1989). Linear time, branching time and partial order in logics and models for concurrency, Springer-Verlag Lecture Notes in Computer Science, 354..
Breuer, J., & Freud, S. (1974). Studies on hysteria, translated by Strachey and Strachey, Harmondsworth Penguin.
Hamey, L. G. C. (1994). XOR has no local minima, Macquarie Computing Reports.
Hopfield, J. J. (1984). Neurons with graded responses have collective computational properties like those of two-state neurons, Proceedings of the National Academy of Sciences, 81, 3088-3092.
Kohonen, T. (1982). Self-organized formation of topologically correct feature maps, Biological Cybernetics, 45, 53-69.
Latimer, C. R. (1995). Computer modelling of cognitive processes, Noetica: Open Forum, 1(1), http://psy.uq.edu.au/CogPsych/Noetica/.
Julien Offray de Lamettrie, J. O. (1748). L'Homme machine.
Pnueli, A.(1981). The temporal semantics of concurrent programs, Theoretical Computer Science, 13, 45-60.
Pratt, V. (1986). Modeling concurrency with partial orders, International Journal of Parallel Programming 15, 33-71.
Rodriguez, R. V. (1993). A relativistic temporal algebra for efficient design of distributed systems, Journal of Applied Intelligence 3, 31-45.
Rumelhart, D. E. & McClelland, J. L. (1986). Parallel Distributed Processing: Explorations in the microstructure of cognition}, Cambridge, MA: MIT Press.
Vroon, P. & Draaisma, D. (1990). De mens als metafoor (Man as metaphor).
Wiles, J. (1995). The connectionist modeller's toolkit: A review of some basic processes over distributed memories, Noetica: Open Forum, 1(3), http://psy.uq.edu.au/CogPsych/Noetica/.