Contributions to discussion

J. P. Sutcliffe
University of Sydney

[1] Reply to the question: What is a psychological theory? posed by Simon Dennis during discussion of George Oliphant's paper.

Even if it may share with others some of its methods of enquiry, each field of scientific endeavour has its own defining subject matter and explicanda . For example, although computer programming is done in both fields, Computer Science and Psychology have different subject matters, with networks and other computing devices and their behaviour as such being explicanda for the former, and with human beings and their behaviour as such being explicanda for the latter.

Thus, to answer the question posed, one begins by specifying the subject matter of Psychology viz. the behaviour of living organisms. That subject matter, as noted, is distinct from that of Computer Science, and also from that of Physiology, Sociology, and the other sciences of man. While each may study "behaviour", their subject matters are distinct in that the behaving entities are different in each case: individual living organisms; computing machines; biological cells; social groups; and so on.

Relative to the subject matter of psychology, one poses questions of explanation and cause, and thence formulates propositions in psychological terms. Some such propositions may already be established, but otherwise one has only conjecture about their truth. In minimal form then, a psychological theory is a set of propositions bearing on matters of explanation and/or cause, such that their terms refer to one or another aspect of psychological subject matter, and the truth for at least one of them (called the hypothesis) is to be established.

[2] Comment, on the use of the fashionable terms network, model, and metaphor , made during discussion of George Oliphant's paper.

To examine the belief that one can use a network as a model or metaphor to account for human behaviour, it is instructive to consider the nature of the discourse which provides the context for such attempts. That contemporary discourse is more rhetorical than scientific.

With a responsibility and commitment to truth, the scientific way of proceeding is to state explicitly ones aims, and to set out the theory to be investigated as a series of clearly stated propositions: definitions, hypotheses, and supplementary assumptions. Then, in the terms of the hypothetico-deductive method, predictions deduced from the theory guide the investigation, such predictions being subjected initially to logical tests which, if passed, allow subsequent empirical tests. For the latter one would consider facts already established, or go on to make new observations, if necessary under experimental control. The mode of argument throughout is logical, that is, validity of argument is mandatory. The aim is to establish objectively what is the case , e.g. if a human being is a machine, then what sort of a machine is it? what are its component parts? how are they put together? and how then does it function?

. In contrast, the aim with Rhetoric - "the science of persuasion"- is to establish belief , whether or not the belief engendered in the person persuaded corresponds to what objectively is the case. Secondary to the matter of truth, what the rhetorician desires and values isthat others are persuaded of his or her beliefs . Although logical argument is included as one of its modes of persuasion, Rhetoric embraces many other methods. Much of contemporary discourse in Psychology and Cognitive Science is pseudo-scientific, being heavily larded with non-logical rhetorical devices. Bandwagon conformity pressures are applied with the use of: fashionable jargon ; the sledgehammer of forceful authoritarian repetition which pre-empts critical evaluation of what is being asserted; word magic as with the liberal use of "must be" words like "therefore" ; hand-waving appeals; and so on.

A major objection to the terms "model" and "metaphor" is that, should a difficulty arise, they are used in a spongy way to deflect criticism with statements such as "models are neither true nor false", "it is only a metaphor", and so on. Their "having heuristic value" has been proffered as a justification for playing with notions like "model" and "metaphor" ; but that defence is empty unless the ideas heuristically provoked are potentially objectively explanatory. Though one may use word magic and other rhetorical devices to persuade others that one is explaining human behaviour in the terms of the functioning of networks, there remains the question of substance for any such beliefs. Scientific responsibility is to be vigilant against rhetoric, and to examine the logic of any such claims. When using terms like metaphor and model to account for something, one should clearly define those terms. Then, dispensing with illogical rhetoric, one should (a) explicate the which is to be modelled; (b) expli

[3] Comment, on the source and content of the categories of input to and output from a network purported to account for human performance, made during the discussion of Kate Stevens' paper.

My comment can be made most simply with reference to your first network example which had just two outputs.

Using discriminant function analysis as a computational analogue of running a network with two output categories, one can see that for the network to give the correct outputs (for the objects to be correctly classified ), the inputs to the network must be such as to allow the intended association of the distributions-of-information-over-the-input-categories with the output categories (the object descriptions must be such that the objects are located within two disjoint clusters within the property space.) If the input information were not sufficient for the needed differentiation, the network could never completely "learn" to give the intended outputs (there would certainly be errors of classification of the objects described.)

Furthermore, for the network to give "correct" output for a given input, prior to the input the network (i) must have been assigned (as parts of itself) a repertoire of categories of input and a repertoire of categories of output; and (ii) it must have been supplied via the input categories with information relevant to differentation of the output categories. The network itself cannot fulfil either (i) or (ii); it is the (human) inventor of the network who carries out those activities. That is done by her/him knowing kinds of thing (the individual musical items and that any one of them is either in standard or modified form) and by choosing certain features to serve as input to the network in the light of knowledge that features of variation-in-sound-pressure serve to differentiate the standard and modified forms of musical item. Thus a "network" is not a non-human entity, comparison with whose behaviour can work heuristically towards an account of human behaviour. Rather, a "network" is a man-machine s

As the mechanical workings of a "network" appear themselves not to be well understood, and as the "man" part of a man-machine system is characteristically a "black box", the prospects are poor for gaining understanding of human cognition from consideration of the behaviour of networks assigned analogous tasks. One must approach the problem from the other direction: first develop the psychological theory before eventually understanding the operation of the (human) machine which when "built" may work along the lines of such theory.

[4] Comment, on correspondence between psychological and network variables, during the discussion of Simon Dennis' paper.

With reference to the first part of your talk, and to the previous talk given by Graeme Halford, it is not good enough to say offhand that the inputs to the network and to the human were "the same", and/or that the outputs from the network and from the human were "the same". In the language of the behaviourism of the 1930s, living organisms can manifest stimulus equivalence , and/or response equivalence . For example, in those terms the physically different stimuli "dog" and "chien" can be said to be "the same" inputs for a person who already can recognize the words and knows that both words have the same referent. Similarly, if the task is to indicate how a word should be pronounced, then, for an experienced user of the language, saying the word can be said to be "the same" output as selecting someone else's pronunciation of it from a menu of sounds. One can be gulled, by the efficacity of such equivalences, within one's own experience and between oneself and another person, into believing that such equivalences hold in all contexts. In fact, however, such equivalences are highly conditional upon the structure and learning experience of the organisms being considered. A machine will have neither stimulus equivalence nor response equivalence ! capabilities unless it has been s

In Halford's study, the "inputs" to the child comprised (visually presented) diagrams and (auditorily presented) instructions; whereas the correlative "inputs" to the network were quite different things. It was only for the investigator (under stimulus equivalence), and not for the network vis-a-vis the child, that the two sorts of input were "the same". Analogously, while the child's response to a given diagram and the correlative output of the network were "response equivalent" for the investigator, the two outputs, for the network and the child per se, were quite different things. In the latter case neither stimulus equivalence of the inputs used nor response equivalence for the resulting outputs hold, because the machine's structure and learning experience are vastly short of those for the child.

If two systems are identical, depending upon their organization they may give different outputs to different inputs or they may give the same outputs to different inputs. If the two systems are distinct they may give different outputs to different inputs, or they may give the same outputs to different inputs. System organization being indifferent to difference in input, one cannot conclude anything about the nature of one system's reaction to one set of inputs, from consideration of the outputs of another system's reaction to a different set of inputs.

There is need for much closer specification of what are being called inputs and outputs , and for closer scrutiny of the logic of comparisons of human and machine behaviours.