Janus: A Summing Up

by Arthur Koestler

  * It is interesting to note the intense reluctance of academic psychologists -- even those who have outgrown the cruder forms of behaviourist S-R theory -- to come to grips with reality. Thus Professor G. Miller writes in an article on psycholinguistics: 'As psychologists have learnt to appreciate the complexities of language, the prospect of reducing it to the laws of behaviour so carefully studied in lower animals [he means Skinner's rats] has grown increasingly remote. We have been forced more and more into a position that non-psychologists probably take for granted, namely, that language is rule-governed behaviour characterized by enormous flexibility and freedom of choice. Obvious as this conclusion may seem, it has important implications for any scientific theory of language. If rules involve the concepts of right and wrong, they introduce a normative aspect that has always been avoided in the natural sciences. . . . To admit that language follows rules seems to put it outside the range of phenomena accessible to scientific investigation.' [20] What a very odd notion of the purpose and methods of 'scientific investigation'!

  The nature of the code which regulates behaviour varies of course according to the nature and level of the hierarchy concerned. Some codes are innate -- such as the genetic code, or the codes which govern the instinctive activities of animals; others are acquired by learning -- like the kinetic code in the circuitry of my nervous system which enables me to ride a bicycle without falling off, or the cognitive code which defines the rules of playing chess.

  Let us now turn from codes to strategies. To repeat it once more: the code defines the permitted moves, strategy decides the choice of the actual move. The next question is: how are these choices made? We might say that the chess-player's choice is 'free' -- in the sense that it is not determined by the rule-book. In fact the number of choices confronting a player in the course of a game of forty moves (while calculating the potential variations which each move might entail two moves ahead) is astronomical. But though his choice is 'free' in the above sense of not being determined by the rules, it is certainly not random. The player tries to select a 'good' move, which will bring him nearer to a win, and to avoid a bad move. But the rule-book knows nothing about 'good' or 'bad' moves. It is, so to speak, ethically neutral. What guides the player's choice of a hoped-for 'good' move are strategic precepts of a much higher complexity -- on a higher level of the cognitive hierarchy -- than the simple rules of the game. The rules a child can learn in half an hour; whereas the strategy is distilled from past experience, the study of master games and specialized books on chess theory. Generally we find on successively higher levels of the hierarchy increasingly complex, more flexible and less predictable patterns of activity with more degrees of freedom (a larger variety of strategic choices); while conversely every complex activity, such as writing a letter, branches into sub-skills which on successively lower levels of the hierarchy become increasingly mechanical, stereotyped and predictable.* The original choice of subjects to be discussed in the letter is vast; the next step, phrasing, still offers a great number of strategic alternatives but is more restricted by the rules of grammar; the rules of spelling are fixed with no elbow-room for flexible strategies, and the muscle-contractions which depress the keys of the typewriter are fully automatized.

  * Cf. the ethologist's 'fixed action patterns'.

  If we descend even further down into the basement of the hierarchy, we come to visceral processes which are self-regulating, controlled by homeostatic feedback devices. These, of course, leave little scope for strategic choices; nevertheless, my conscious self can interfere to some extent with the normally unconscious, automated functioning of my respiratory system by holding my breath or applying some Yoga technique. Thus the distinction between rules and strategies remains in principle valid even on this basic physiological level. But the relevance of this distinction will only become fully apparent in later chapters when we apply it to such fundamental problems as the theory of evolution; free will versus determinism; and the pathology and creativity of the human mind.

  12

  As already mentioned, the purpose of this chapter is not to provide a manual of hierarchies, but to convey some idea of the conceptual framework on which this inquiry is based, and to give the reader the 'feel' of hierarchic thinking as opposed to the current reductionist and mechanistic trends. To conclude this summary survey, I must mention, however briefly, a few more principles which all hierarchic systems have in common.

  One obvious point is that hierarchies do not operate in a vacuum, but interact with others. This elementary fact has given rise to much confusion. If you look at a well-kept hedge surrounding a garden like a living wall, the rich foliage of the entwined branches may make you forget that the branches originate in separate bushes. The bushes are vertical, arborizing structures. The entwined branches form horizontal networks at numerous levels. Without the individual plants there would be no entwining, and no network. Without the network, each plant would be isolated, and there would be no hedge and no integration of functions. 'Arborization' and 'reticulation' (net-formation) are complementary principles in the architecture of organisms and societies. The circulatory system controlled by the heart and the respiratory system controlled by the lungs function as quasi-autonomous, self-regulating hierarchies, but they interact on various levels. In the subject-catalogues in our libraries the branches are entwined through cross-references. In cognitive hierarchies -- universes of discourse -- arborization is reflected in the 'vertical' denotation (classification) of concepts, reticulation in their 'horizontal' connotations in associative nets.

  The complementarity of arborization and reticulation yields relevant clues to the complex problem of how memory works.*

  * The section that follows is a summary of The Act of Creation, Book II, Ch. X, The Ghost in the Machine, Ch. V and VI, and of a paper read to the Harvard Medical School Symposium on 'The Pathology of Memory'. [21]

  13

  In Stevenson's novel Kidnapped, Alan Breck makes the casual remark: 'I have a grand memory for forgetting, David.' He speaks for all of us, and not only those afflicted with aphasia or senility. Painful as it is, we have to admit that a large proportion of our memories resembles the dregs in a wine glass, the dehydrated sediments of experiences whose flavour has gone -- or, to change the metaphor, they are like dusty abstracts of past events on the shelves of a dimly lit archive. Fortunately this applies only to one type or category of memories, which I shall call abstractive memory. But there is another category, derived from our capacity to recall past episodes, or scenes, or details of scenes, with almost hallucinatory vividness. I shall call this the spotlight type of memory, and I shall contend that 'abstractive memory' and 'spotlight memory' are different classes of phenomena, based on different neural mechanisms.

  Take abstractive memory first. The bulk of what we can remember of our life history, and of the knowledge we have accumulated in the course of it, is of the abstractive type.

  The word 'abstract' has, in common usage, two main connotations: it is the opposite of 'concrete', in the sense that it refers to a general concept rather than a particular instance; in the second place, an 'abstract' is a condensation of the essence of a longer document. Memory is abstractive in both senses. I watch a television play. The exact words of each actor are forgotten within a few seconds; only their abstracted meaning is retained. The next morning I can only remember the sequence of scenes which constituted the story. A month later, all I can remember is that the play was about a gangster on the run. Much the same happens to the mnemic residues of books one has read and whole chapters of one's own life-story. The original experience has been stripped of detail, skeletonized, reduced to a colourless abstract before being confined to the memory store. The nature of that store is still a complete mystery in brain-research, but it is obvious that if stored knowledge and experience are to be retrievable (for otherwise they would be useless), they must be ordered according to the hierarchic principle -- like a thesaurus or a library subject-catalogue,
with headings and sub-headings but also with a wealth of cross-references to assist the process of retrieval (the former representing arborization, the latter the reticulation of the hierarchic structure). If we pursue for a moment the metaphor of a library representing our memory stores, we arrive at rather depressing conclusions. Quite apart from the countless volumes that are left to rot away or fall to dust, there is a hierarchy of librarians at work who ruthlessly condense long texts into short abstracts and then make abstracts of the abstracts.

  This process of sifting and abstracting actually starts long before a lived experience is confined to the memory store. At every relay station in the perceptual hierarchy through which the sensory input must pass before being admitted to consciousness it is analysed, classified and stripped of irrelevant detail.* This enables us to recognize the letter R in an almost illegible scrawl as 'the same thing' as a huge printed R in a newspaper headline, by a sophisticated scanning process which disregards all details and abstracts only the basic geometrical design -- the 'R-ness' of the R -- as worth signalling to higher quarters. This signal can now be transmitted in a simple code, like a message in Morse, which contains all the relevant information -- 'it's an R' -- in a condensed, skeletonized form; but the wealth of calligraphic detail is of course irretrievably lost, as the inflections of the human voice are lost in the Morse message. The wistful remark 'I have a memory like a sieve may be derived from an intuitive grasp of these filtering devices which operate all along the input channels and storage channels of the nervous system.

  * The psychologist distinguishes on the lower levels of the hierarchy lateral inhibition, habituation, and efferent control of the receptors; on the higher levels the mechanisms responsible for the visual and auditory constancy phenomena, and the scanning and filtering devices that account for pattern recognition and enable us to abstract universals.

  Yet even the chosen few among the multitude of potential stimuli incessantly bombarding our receptor organs which have successfully passed all these selective filters and have attained the status of consciously perceived events, must submit, as we have seen, to further rigorous stripping procedures before being admitted to the permanent memory store; and as time passes, they will suffer further decay. Memory is a prize example of the law of diminishing returns.

  This retrospective impoverishment of lived experience is unavoidable; 'abstractive memory implies the sacrifice of particulars. If, instead of abstracting generalized concepts, like 'R' or 'tree' or 'dog', our memories consisted of a collection of all our particular experiences of R's and trees and dogs encountered in the past -- a storehouse of lantern slides and tape -- recordings -- it would be a chaotic jumble, completely useless for mental guidance, for we would never be able to identify an R or understand a spoken sentence. Without hierarchic order and classification, memory would be bedlam (or the parroting of sequences learned by rote, and reinforced by conditioning, which is the behaviourist's model -- or caricature -- for remembering).

  To say it again: the loss of particulars in abstractive memory is unavoidable. Fortunately this is not the whole story, for there are several compensating factors which, at least in part, make up for the loss.

  In the first place, the abstractive process can acquire a higher degree of sophistication by learning from experience. To the novice, all red wines taste alike, and all Japanese males look the same. But he can be taught to superimpose more delicate perceptual filters on the coarser ones, as Constable trained himself to discriminate between diverse types of clouds and to classify them into sub-categories. Thus we learn to abstract finer and finer nuances -- to make the trees of the hierarchies of perception grow new shoots, as it were.

  Moreover, it is important to realize that abstractive memory is not based on a single hierarchy but on several interlocking hierarchies pertaining to different sensory realms such as vision, hearing, smell. What is less obvious is that there may exist several distinct hierarchies with different criteria of relevance operating within the same sense modality. I can recognize a melody regardless of the instrument on which it was played; but I can also recognize the sound of an instrument regardless of the melody played on it. We must therefore assume that melodic pattern and instrument sound (timbre) are abstracted and stored independently by separate filtering hierarchies within the same sensory modality but with different criteria of relevance. One abstracts melody and disregards timbre, the other abstracts the timbre of an instrument and disregards melody as irrelevant. Thus not all the detail discarded as irrelevant by one filtering system is irretrievably lost, because it may have been retained and stored by another filtering hierarchy with different criteria of relevance.

  The recall of an experience would then be made possible by the cooperation of several interlocking hierarchies, which may include different sense modalities, for instance, sight and sound or odour, or different branches within the same modality. You may remember the words of the aria 'Your Tiny Hand is Frozen', but have lost the tune. Or you may remember the tune after having forgotten the words. And you may recognize the unique timbre of Caruso's voice on a gramophone record, regardless of the words and the tune he is singing. But if two, or all three of these features have been abstracted and stored, the recall of the original experience will have more dimensions and be the more complete.

  The process could in some respects be compared to multi-colour printing by the superimposition of several colour-blocks. The painting to be reproduced -- the original experience -- is photographed through different colour-filters on blue, red and yellow plates, each of which retains only those features that are 'relevant' to it: i.e., those which appear in its own colour, and ignores all other features; then they are recombined into a more or less faithful reconstruction of the original input. Each hierarchy would then have a different 'colour' attached to it, the colour symbolizing its criteria of relevance. Which memory-forming hierarchies will be active at any given time depends, of course, on the subject's general interests and momentary state of mind.

  Although this hypothesis represents a radical departure from both the behaviourist and the Gestalt schools' conceptions of memory, some modest evidence for it can be found in a series of experiments carried out in cooperation with Professor J. J. Jenkins in the psychological laboratory of Stanford University;* and more tests on these lines can be designed without much difficulty.

  * See Appendix II. This is a rather technical paper of possible interest to experimental psychologists, which the general reader can safely ignore. The gist of the experiment was to show to each subject for a fraction of a second only (by means of an apparatus called a tachistoscope) a number of seven or eight digits, and then let him try to repeat the sequence. The results of several hundred experiments show that a highly significant number of errors (approximately fifty per cent) consisted in the subject correctly identifying all numbers in the sequence, but inverting the order of two or three neighbouring digits. This seems to confirm that the identification of individual digits, and the determination of their sequential order, are carried out by different branches of the perceptual hierarchy.

  14

  The 'colour printing' hypothesis may provide part of the explanation of the complex phenomena of memory and recall, but it is based solely on the abstractive type of memory which by itself cannot account for the extreme vividness of the 'spotlight' type of memory mentioned at the beginning of this chapter. It is a method of retention based on principles which seem to be the exact opposite of memory formation in abstractive hierarchies. It is characterized by the recall of scenes or details with almost hallucinatory clarity. They are rather like photographic close-ups, in contrast to abstractive memory's aerial panorama seen through a haze. The emphasis is on detail, which may be a fragment, torn from its context, that survived the decay of the whole to which it once belonged -- like the single lock of hair on the shrivelled mummy of an Egyptian princess. It may be auditory -- a line from an otherwise forgotten poem, or a chance remark by a stranger over
heard in a bus; or visual -- a wart on Nanny's chin, a hand waving farewell from the window of a departing train; or even refer to taste and smell, like Proust's celebrated madeleine (the French pastry, not the girl). Though often trivial from a rational point of view, these spotlighted images add texture and flavour to memory, and have an uncanny evocative power. This suggests that, although irrelevant by logical criteria, they have some special emotive significance (on a conscious or unconscious level) that caused them to be retained.

  Nobody, not even computer designers, thinks all the time in terms of abstractive hierarchies. Emotion colours most of our perceptions, and there are indications that our emotive reactions also involve a hierarchy of levels -- including archaic structures in the brain which are phylogenetically much older than the structures concerned with abstract conceptualizations. One might speculate that in the formation of 'spotlight memories' these older levels in the hierarchy play a dominant part.

  There are some further considerations in favour of such a hypothesis. First, from the neurophysiologist's point of view, they receive strong support from the Papez-MacLean theory of emotions.* Second, from the standpoint of the communication-theorist, abstractive memory generalizes and schematizes, while spotlight memory particularizes and concretizes -- which is a much more primitive method of storing information.** Third, from the standpoint of the psychologist, abstractive memory would be related to insightful learning and spotlight memory to a process resembling imprinting. But imprinting in Konrad Lorenz's geese is restricted to a critical period of a few hours, and apparently results in a very coarse and vague imprint. On the human level, imprinting may take the form of eidetic imagery. According to Jaensch [22] and Kluever [23], a considerable proportion of children have the eidetic faculty -- they are able to 'project' a photographically accurate, coloured image of a previously fixated picture onto a blank screen and to repeat this after long intervals, sometimes even years. Penfield and Roberts' [24] experiments, evoking what is claimed to be total recall of past scenes by electrical stimulation of the patient's temporal lobes, may be a related phenomenon.