Definitions and Interpretations: Comments on the Symposium on Connectionist Models and Psychology

Ellen Watson
Department of Philosophy
University of Queensland

At a number of points during the seminar, discussion became paralysed over matters of vocabulary. By the end of the day, anyone who still dares to use words such as 'frame', 'schema', 'psychological theory', 'functional architecture', 'account for', 'metaphor', 'rule', 'subsymbol', of even 'model' had to do so with apologies or caveats. Because we kept coming back to these questions, and because they have a potential to be so paralysing if not answered, permit me to draw the following conclusion. Whatever neural networks have to contribute to psychology, and whatever psychologists are able to contribute to the legitimacy of neural networks or computational modelling in general, philosophers have a number of things that they can contribute to both. Philosophers do not (usually) perform experiments or run simulations, but we do specialise in defining terms, uncovering assumptions, specifying the problem to be tested, and helping to determine whether that problem has been tested after all. The seminar focussed on the possibilities for interdisciplinary cooperation between computer scientists and psychologists; I would propose that philosophers also have a useful role they might play in this program of cooperative research.

Metaphor

Mike Johnson argued that new metaphors are needed, but who is it that needs them? He suggested that psychology needs a new machine metaphor, because these have been shown to drive psychological theory throughout its history (even back to Aristotle and Descartes). But with neural networks the technology (and attendant mathematics) has outstripped our ability to understand it. We need metaphors in order to understand the neural networks themselves, with their vectors, tensor products, hidden units and "subsymbols".

Sally Andrews suggested that metaphors are necessary and invaluable because in trying to link brain and behaviour we a re trying to link two incommensurable levels of analysis. Peter Slezak questioned why psychologists hedge their theories more than other scientists by calling them metaphors, and why they don't adopt principles of scientific realism and put their theories forward as, maybe not perfect, but at least purported descriptions of reality.

To navigate the waters between Andrews and Slezak, for and against metaphors, we need to sort out who is to use the metaphors, for what, where they appear in the theory, at what level, and what truth bearing descriptions could potentially replace them -- are they unavoidable and the limit of psychology, or are they holding a place for more detailed and well worked out theories of human mind? Philosophy of science could help us sort out all of these questions, because it investigates the nature of theories, their relation to evidence and their relation to the phenomena to be explained. Andrews' justification of metaphor suggests that neuronal and behavioral levels of explanation really are incommensurable; Slezak's comments suggest that psychologists treat their theories differently than those in other sciences. Do psychologists want to continue making these assumptions? Making the assumptions more explicit might make the choice more clear.

What is a psychological theory?

Philosophy of science might be able to help here, too, since as I mentioned above, some of the central questions of philosophy of science are 'What is a theory?' and, 'What is the relationship between theory and evidence?' However, here we start travelling in a circle (or maybe pulling ourselves up by our bootstraps) because Paul Churchland's recent book contains an argument based on connectionist assumptions about the nature of mental states (Churchland, 1989). If we think of theories in the context of philosophy of mind, and if we subscribe to Churchland's form of connectionism, then theories turn out to be collections of vectors in n-dimensional weight space (sound familiar?). At this point, then, we need to turn back to the philosophy of science and to epistemology to see if the circle in which we are caught is vicious or virtuous. What happens when your philosophy of mind grounds the science on which you have based your philosophy of mind? This is another question with which philosophers have grappled. As Cyril Latimer pointed out, philosophers have been asking questions about the nature of objects and their properties and our perception and representation of these properties since philosophy began. Although philosophers are far from coming up with the last word on the subject, we have made some mistakes that contemporary cognitive scientists shouldn't have to make again, such as mistaking a prototypical representation for a specific image (as Berkeley did), or forgetting to specify how a cognitive machine can automatically extract features out of a holistically observed scene (as Locke did long before schema theorists).

What is a rule?

Peter Slezak brought up the question of whether a model has a rule, and made an analogy to the work of Chomsky and his claim that human beings have rules of grammar. When I learned my Chomsky in a philosophy of language seminar, the instructor was careful to point out that there are two ways that something can "have" a rule. One is to have the rule actually inscribed inside; in this case, the system looks up the rule and applies it. John Searle attributed this picture of rule-following to strong AI in his Chinese room paper (Searle, 1980), and rightfully criticised it for invoking irreducible homunicularism. The other way a system can "have" a rule is that it can obey the rule, i.e. have its behaviour align with it, without having to look up that rule anywhere inscribed and without having to look it up. In this sense, planets obey "rules" that describe their orbits (also known as "laws"), without far as we know) representing those rules to themselves. You can't have a debate about the nature of rules in a particular system without delineating which sense of "have" you have in mind.

Subsymbol

The term 'subsymbol' is similarly ambiguous. Since the publication of Paul Smolensky's article, "On the proper treatment of connectionism," critics have been trying to get Smolensky himself to be clear about what he means. Are subsymbols still symbols, or not? In my opinion, the whole excitement about neural networks is that they are able to process information without internal programming that one can read off in something like English propositions, and therefore show that people like Jerry Fodor are wrong to argue that there must necessarily be a language of thought. However, one must be cautious -- this way could lie behaviourism (Max Coltheart was dangerously close to being called a behaviorists when Sally Andrews accused him of having said that a system has a rule depending on its performance on pronouncing non-words). If neural nets are to serve the purpose of suggesting formal structures for carrying out tasks, which several people claimed for them, including Max Coltheart and George Oliphant, then we need some language in which to read off the properties of hidden units. This seems to me the central issue concerning neural nets with respect to explanatory applications in cognitive psychology, and (therefore?) the most difficult to resolve. It is here that discussions of the interface between neural net research and cognitive psychology suddenly makes contact with the entire philosophical tradition that investigates the nature of meaning and representation itself.

Conclusions

The above topics represent some of the areas in which philosophers might contribute some of their experience and expertise to discussions in connectionism and psychology. However, the benefits of collaboration would definitely flow both ways. As Cyril Latimer said, modelling and real applications enforce theoretical rigour, and this would be true for philosophers as well as scientists. Philosophers can learn a lot from psychologists and AI researchers who try to build models and put some of the theories in practise. These notes are an appeal for continuing the dialogue and the cooperative enterprise.

References

Fodor, Jerry A. (1975) The language of thought. Cambridge, MA: Harvard University Press.

Churchland, P. M. (1989) On the nature of theories: A neurocomputational perspective, A Neurocomputational Perspective: The Nature of Mind and the structure of science. Cambridge, MA: MIT Press.

Churchland, P. M. (1989) On the nature of explanation: A PDP approach, A Neurocomputational Perspective: The Nature of Mind and the structure of science. Cambridge, MA: MIT Press.

Searle, J. R. (1980) Minds, Brains and Programs, The Behavioral and Brain Sciences, 3, 417-457.

Smolensky, P. (1988). On the proper treatment of connectionism, The Behavioral and Brain Sciences, 11,, 1-74.