Friday, 9 September 2016

Gordon Pask: "A Discussion of the Cybernetics of Learning Behaviour" (1963)

At #altc this year (which I didn't attend) there was a keynote given by Lia Commissar (@misscommissar) about the brain and learning. By coincidence I stumbled across a volume in Stafford Beer's archive in Liverpool, edited by Norbert Wiener on "Nerve, Brain and Memory Models" from 1963. It followed a Symposium on Cybernetics of the Nervous System at the Royal Dutch Academy of Sciences in April, 1962. There is a stellar list of contributers:
W. R. Ashby, V. Braitenberg, J. Clark, J. D. Cowan, H. Frank, F. H. George, E. Huant, P. L. Latour, P. Mueller, A. V. Napalkov, P. Nayrac, A. Nigro, G. Pask, N. Rashevsky, J. L.Sauvan, J. P. Shade, N. Stanoulov, M. Ten Hoopen, A. A. Verveen, H. Von Foerster, C. C.Walker, O. D.Wells, N. Wiener, J. Zeman, G.W. Zopf
There is a long paper by Gordon Pask called "A discussion of the cybernetics of learning behaviour" which I thought would be relevant to the current vogue for everything 'neuro' in education. There are many other things there too, including a fascinating paper by Ashby and Von Foerster on "The essential instability of systems with threshold, and some possible applications to psychiatry". There is also a record of the conversation with Wiener afterwards. 

I've quoted the opening of Pask's paper below because it is an excellent summary of the neuroscience of the time. It was surprisingly advanced, and in many ways today's emphasis on MRI scanning technologies has meant that the field has become somewhat homogenised. One of the reasons why I'm interested is because the models of the brain taken by Stafford Beer in his Viable System Model very much belong to this period: what effect would more up-to-date understanding of the brain have had on his thinking? (I'm investigating this with people in Liverpool medical school).  

But Pask's contribution on Learning Behaviour is also interesting because it presents a very early (and rather formal) version of what became conversation theory. He relies quite heavily on Robert Rosen's work ("Representation of biological systems from the standpoint of the theory of categories" (1958) - Bulletin of Mathematical Biophysics; "A logical paradox implicit in the notion of a self-reproducing automaton" (1959), same journal). His championing of Ashby's approach to the brain is, I think, very important.

From "A discussion of the cybernetics of Learning Behaviour" - Gordon Pask, 1962

1.2 The approach of cybernetics

Some cybernetic models are derived from a psychological root, for example, Rosenblatt's (1961) perceptron and George's (1961) automata stem largely from Hebb's (1949) theory. Others, such as Grey Walter's (1953) and Angyan's (1958) respective tortoises, have a broader behavioural antecedent.

On the other hand, neurone models, like Harmon's (1961) and Lettvin's (1959), are based upon facts of microscopic physiology and have the same predictive power linked to the same restrictions as an overtly physiological construction.

Next, there are models which start from a few physiological facts such as known characteristics or connectivities of neurones and add to these certain cybernetically plausable assumptions. At a microscopic level, McCulloch's (1960) work is the most explicit case of this technique (though it does not, in fact refer to adaptation so much as to perception) for its assumptions stem from Boolean Logic (Rashevsky, (1960), describes a number of networks that are adaptive). Uttley (1956), using a different set of assumptions, considered the hypothesis that conditional probability computation occurs extensively in the nervous system. At a macroscopic level, Beurle (1954) has constructed a statistical mechanical model involving a population of artificial neurones which has been successfully simulated, whilst Napalkov's (1961) proposals lie between the microscopic and macroscopic extremes.

Cyberneticians are naturally concerned with the logic of large systems and the logical calibre of the learning process. Thus Willis (1959) and Cameron (1960) point out the advantages and limitations of threshold logic. Papert (1960) considers the constraints imposed upon the adaptive process in a wholly arbitrary network, and Ivahnenko (1962) recently published a series of papers reconciling the presently opposed idea of the brain as an undifferentiated fully malleable system and as a well structured device that has a few adaptive parameters. MacKay (1951) has discussed the philosophy of learning as such, the implications of the word and the extent to which learning behaviour can be simulated; in addition to which he has proposed a number of brain-like automata. But it is Ashby (1956) who takes the purely cybernetic approach to learning. Physiological mechanisms are shown to be special cases of completely general systems exhibiting principles such as homeostasis and dynamic stability. He considers the behaviour of these systems in different experimental conditions and displays such statements as 'the system learns' or 'the system has a memory' in their true colour as assertions that are made relative to a particular observer. 

No comments: