Some Loose Ends in Cybernetic Thinking*

There are various contexts in which there seem to be “loose ends” in aspects of Cybernetics and I will try to discuss a set of them – five main headings to be exact, though I can think of many others – with a few subheadings. The emphasis will be on rather old-fashioned (“first-order”) treatments since I think there are many questions raised there that have either not been answered, or whose further consideration seems likely to be profitable. The five main topics are:

  1. How does a learning system decide when to classify and when to fine-tune?

  2. Is there anything to be learned from computer programming methods about the evolution of language?

  3. Have “administrative methods” been under-utilised in AI?

  4. Do neural processes depend on accepted neural doctrine?

  5. Is biological evolution really neo-Darwinian?

*By Alex Andrew – Talk to the UK Cybernetics Society on 27:10:03

  1. How does a learning system decide when to classify and when to fine-tune?

Schemes for artificial learning systems come under two main headings discrete and continuous. A discrete scheme (such as Uttley's Conditional Probability Computer, modelling a conditioned reflex), requires the assumption that the significantly different states of the controlled system can be treated separately. For control of an industrial plant, or a task such as riding a bicycle, the number of distinguishable states is unmanageable. Learning responses separately for each is impractical. It is essential to exploit relationships between states due to continuity of the variables.

Continuous-variable learning schemes have been built or proposed by, among others, Gabor et al (1961) and Andrew (1959), and a large part of the operation of Samuel's checker-playing program (1963) has the same character. These schemes go beyond mere parameter-adjustment and have a self-organising feature in that they select terms to be included in the function being optimised.

All these schemes assume that the learning device is set to do one particular task and that the conditions are reasonably stable. Of course, a person or animal faces different tasks, and a given task may have different characteristics on different occasions. Giving a talk, for example, is somewhat different depending on the type of topic, as well as on the size and amenability of the audience, and on the available audio-visual aids, etc. A learning system needs to have a means of classifying situations, and then associating the results of learning of the fine-tuning kind with the appropriate situation class. Fairly certainly, the classification should be subject to review during the fine-tuning stage. An extra complication, though, is that aspects of fine-tuning may be common to a range of situation types. For example, if the tasks to be learned by a person or a robot include walking, cycling, and driving a car, these are three situations that have to be treated differently, and yet there are aspects of balance and of sensory input interpretation, and of muscle coordination that are common to all three and for which the distinctions between them should be ignored, and presumably are ignored in living systems.

Donald MacKay (1959) has advocated a form of Information Theory in which he distinguishes logons, or units of structural information, from metrons, or units of metrical information. The sensory inputs to a versatile learning system are metrical, and the system must derive structure for itself. There is a connection with the early work of Gordon Pask, who maintained that any process to be described as “learning” (as distinct from adaptation) must introduce a new descriptive language.

A related “loose end” is in the distinction between negative feedback of a continuous kind, and a similar action having more structure, perhaps best represented by the “means-ends analysis” of the General Problem Solver of Newell, Shaw and Simon. In some early discussions of the place of Cybernetics its interdisciplinary nature was emphasized by suggesting that the reflex arc of the neurophysiologist was essentially the same as the negative feedback loop of the control engineer. This is rather misleading since a spinal reflex in an animal (withdrawal reflex or scratch reflex) can be quite structured, though there are other biological control loops that correspond better to the engineer's version.

In the Introduction to Cybernetics, Norbert Wiener touches on this distinction between continuous and structured control, and seems to feel embarrassment at the need to gloss over it.

“Now, suppose that I pick up a lead pencil. To do this, I have to move certain muscles. However, for all of us except a few expert anatomists, we do not know what these muscles are, and even among the anatomists, there are few, if any, who can perform the act by a conscious willing in succession of the contraction of each muscle concerned. On the contrary, what we will is to pick the pencil up. Once we have determined on this, our motion proceeds in such a way that we may say roughly that the amount by which the pencil is not yet picked up decreased at each stage. This part of the action is not in full consciousness.

“To perform an action in such a manner, there must be a report to the nervous system, conscious or unconscious, of the amount by which we have failed to pick up the pencil at each instant.”

MacKay, D.M. (1959) (then of Dept. of Physics, Kings College, London) “Operational aspects of intellect” in: Mechanisation of Thought Processes (NPL Symposium in 1958) HMSO, London, vol. 1, pp. 37-73

Wiener, Norbert (1948, 1961) Cybernetics or control and communication in the animal and the machine, MIT Press and Wiley, New York. (Text referred to is on p. 7, in Introduction section, of second edition, 1961)

Andrew, A.M. (1959) “Learning machines” in: Mechanisation of Thought Processes (NPL Symposium in 1958) HMSO, London, vol. 1, pp. 473-509

Gabor, D., Wilby, W.P.L. and Woodcock, R. (1961) “A universal non-linear filter, predictor and simulator which optimizes itself by a learning process” Proc. IRE (London) B13, pp. 422-435

Samuel, A.L. (1963) “Some studies in machine learning using the game of checkers” in: Computers and Thought, ed. E.A. Feigenbaum and J. Feldman, McGraw-Hill, N.Y. pp. 71-105

  1. Is there anything to be learned from computer programming methods about the evolution of language?

It is difficult to imagine how human language could have evolved to its present form, with syntactic structure. There may be a hint in the observation that push-down stacks, or last-in-first-out stores, are useful both in the parsing of linguistic communications and in managing nested procedure calls. This links syntax to an aspect of everyday living, since human performance of a complex task amounts to a set of nested procedure calls. This is well illustrated in discussions of the General Problem Solver of Newell, Shaw and Simon, and a person can be seen as making a return from executing a procedure at any point where he or she might say: “Now, where was I?”

It has to be admitted that formal parsing is more useful with programming languages than with natural utterances, but even so this association of the two distinct uses of stacks in computing could contain a hint as to how language evolved.

The idea receives indirect support from a recent paper by Arbib (2002) in which he discusses studies of the primate brain in which it has been found that there are “mirror areas” in which excitation is similar when an action is observed and when the same action is performed by the individual. Arbib attaches importance to the proximity of these areas to others necessary for speech in humans, and suggests that speech developed from sign language and hence from activity. The dual role of push-down stacks would be consistent with this.

Arbib, Michael (2002) “The mirror system, imitation, and the evolution of language”, in: Imitation in Animals and Artifacts, edited by Kerstin Dautenhahn and Chrystopher L. Nehaniv, MIT Press, Cambridge, Mass. (Book reviewed in Kybernetes 32, no. 7/8, pp. 1189-1191)

  1. Have “administrative methods” been under-utilised in AI?

In an early paper, Marvin Minsky (1959) mentions “Administrative methods” as features he expects to find in AI programs. It can be argued that they are features, in rudimentary forms, of the General Problem Solver but they have tended to be forgotten in later work.

Minsky describes the methods under three headings:

  1. Difficulty estimates.

  2. Sub-problem utilities.

  3. Methods for selecting methods.

Each of these is employed to guide the flow of control in problem-solving.

It seems possible that, in human thinking, paradox is detected by an administrative “circularity detector”, which could have been evolved in the context of hunter/gatherer activity.

McCulloch's “redundancy of potential command” implies an administrative structure, admittedly of an unconventional kind.

Minsky, M.L. (1959) “Some methods of artificial intelligence and heuristic programming” in: Mechanisation of Thought Processes (NPL Symposium in 1958) HMSO, London, vol. 1, pp. 3-36

  1. Do neural processes depend on accepted neural doctrine?

Cells of the body, including neurons, have within them microtubules. There is reason to think that these have a communication function, especially since they are directed, in the sense that their two ends are not equivalent. There seems to be no doubt that molecules are conveyed along them, some in one direction and some in the other, which supports the idea that they provide slow communication channels.

There are suggestions, however, that microtubules also convey rapid signals in a way that depends on quantum theory for its explanation (Pribram, 1963, also Dayhoff et al in the same work, and Jibu and Kusio, 1994). Dayhoff et al seem to be sure that microtubules underlie backpropagation.

The treatments in terms of quantum theory seem highly speculative, but support for the idea that microtubules play some major role comes from observations by Allison and Nunn (1968) and Allison et al (1970) that substances used as inhalational anaesthetics have a strong but reversible effect on microtubules. Their experiments were on actinosphaerium nucleophilum which are protozoans found in pond water.

Sejnowsky (1985) has shown that complex neural computations are carried out in a time that only allows a remarkably small neural “depth”. Of course, this is not a problem if the computation is performed on a high-speed network of microtubules within the neurons, rather than by the network of neurons as in classical theory. Although this highly speculative, it is salutary to realise that our understanding is still such that there can be doubt.

Pribram, Karl H. (1993) (ed) Rethinking Neural Networks: Quantum Fields and Biological Data, Lawrence Erlbaum, Hillsdale, New Jersey.

Dayhoff, Judith E., Hameroff, Stuart, Swenberg, Charles E. and Lahoz-Beltra, Rafael (1993) “The Neuronal Cytoskeleton: A Complex System that Subserves Neural Learning”. Chapter 12, (pp. 389-441) of book edited by Pribram

Jibu, Mari and Yasue, Kusio (1994) “Is Brain a Biological Photonic Computer with Subneuronal Quantum Optical Networks?”, in Cybernetics and Systems '94, edited by Robert Trappl, World Scientific, Singapore, pp. 763-770

Allison, A.C. and Nunn, J.F. (1968) “Effects of general anæsthetics on microtubules: A possible mechanism of anæsthesia”, The Lancet, Dec. 21, pp. 1326-1329

Allison, A.C., Hulands, G.H., Nunn, J.F., Kitching, J.A. and Macdonald, A.C. (1970) “The effect of inhalational anaesthetics on the microtubular system in actinosphaerium nucleophilumJ. Cell Sci. 7, pp. 483-499

Sejnowski, T.J. (1985) “Open questions about computation in cerebral cortex” in: Parallel Distributed Processing: Explorations in the Microstructure of Cognition, vol. 2, edited by James L. McClelland, David E. Rumelhart and the PDP Research Group, MIT Press, Cambridge, Mass., pp. 372-389.

  1. Is biological evolution really neo-Darwinian?

The currently accepted view of evolution is that it depends entirely on chance mutations and selection. The view is eloquently defended by Richard Dawkins, who, however, dismisses Lamarckianism essentially on the grounds that there is no evidence for it and no need for it. On the other hand, Musès (1996) has argued that chance mutation alone is insufficient, and arguments are advanced by Koestler (1971) that seem rather convincing.

These writers do not oppose Darwinian ideas in general, and in fact point out that Darwin accepted Lamarckian principles. The principle of inheritance of acquired characteristics was imposed as part of a political doctrine by Lysenko, but now the neo-Darwinian view, that everything depends on chance mutations and selection, is defended with almost religious fervour. Koestler makes a good case for supposing that there may be some additional factor linking evolutionary change to the animal's awareness of what would be useful.

Koestler's book is largely about the Austrian experimental biologist, Dr Paul Kammerer, who died, apparently by suicide, on September 23, 1926. Kammerer experimented with various amphibia including the so-called midwife toad, and also the sea squirt Ciona intestinalis and found evidence for inheritance of acquired characteristics.

The midwife toad, Alytes obstetricans, differs from other toads in mating on dry land. The males of other toad species grow “nuptial pads” on their forelimbs that help them grasp the female. The males of Alytes do not have these and are successful without them because they mate on dry land. Kammerer claimed he could breed the toads under conditions where they mated in the wet, and that after several generations the males developed nuptial pads, and the characteristic was inherited. The results are generally considered to be discredited by the fact that one male toad showing the pads was found to have been “doctored” by injection of indian ink. However, Kammerer himself seems to have been worried by the atypical appearance of this specimen and it is possible that the ink was injected by an over-zealous assistant. It was shortly after this humiliation that Kammerer died, apparently by his own hand though there are some doubts.

The sea squirt Ciona is a small immobile aquatic animal that takes in sea water through one tube termed a “siphon” and expels it, after filtering for nutrient, through another. It has been found that, if the siphons are cut off, they regrow larger than before. Kammerer claimed that the increased siphon length was inherited.

Apart from these controversial claims by Kammerer, Koestler makes a strong case for considering alternatives to the accepted neo-Darwinian view. (There are surprising connections with Cybernetics in that one of the main opponents of Kammerer was William Bateson, the father of Gregory Bateson, and reference is made to von Bertalanffy as a critic of neo-Darwinian orthodoxy.) The following is an extract from page 131 of Koestler's book:

Some classical examples quoted over and again in the literature seem almost to cry out for a 'Mini-Lamarckian' explanation:

There is, for example, the hoary problem why the skin on the soles of our feet is so much thicker than elsewhere. If the thickening occurred after birth, as a result of stress, wear and tear, there would be no problem. But the skin of the sole is already thickened in the embryo which has never walked, barefoot or otherwise. A similar, even more striking phenomenon are the callosities on the African warthog's wrists and forelegs, on which the animal leans while feeding; on the knees of camels; and oddest of all, the two bulbous thickenings on the ostrich's undercarriage, one fore, one aft, on which that ungainly bird squats. All these callosities make their appearance, as the skin on our feet does, in the embryo. They are inherited characters. But is it conceivable that these callosities should have evolved by chance mutation just exactly where the animal needed them? Or must we assume that there is a causal, Lamarckian connection between the animal's needs and the mutation which provides them?

Musès, Charles (1996) “Consciousness: the Holy Grail of Science”, Kybernetes 25, no. 7/8, pp. 109-129

Koestler, Arthur (1971) The Case of the Midwife Toad, Hutchison, London

Dawkins, Richard (1976) The Selfish Gene, OUP, Oxford, also The Blind Watchmaker, etc.