The Cybernetic Brain

by Andrew Pickering

  Warren McCulloch (1988) notably described his cybernetics as "experimental epistemology," meaning the pursuit of a theory of knowledge via empirical and theoretical analysis of how the brain actually represents and knows the world. We could likewise think of Ashby's cybernetics as experimental ontology. I noted earlier that the general performative vision of the world does not imply any specific cybernetic project; that such projects necessarily add something to the vision, both pinning it down and vivifying it by specifying it in this way or that. The homeostat can certainly be seen as such a specification, in the construction of a definite mechanism. But in Ashby's reflections on time to equilibrium, this specification reacted back upon the general vision, further specifying that. If one recognizes the homeostat as a good model for adaptation, then these reflections imply something, not just about the brain but about the world at large as well: both must consist of sparsely connected dynamic entities.

  We are back to the idea that ontology makes a difference, but with a twist. My argument so far has been that the nonmodern quality of cybernetic projects can be seen as the counterpart of a nonmodern ontology. Here we have an example in which one of these projects fed back as a fascinating ontological conclusion about the coupling of entities in the world. It is hard to see how one could arrive at a similar conclusion within the framework of the modern sciences.27

  DAMS

  AS A SYMBOL OF HIS INTEREST IN RELATIONS HE CARRIED A CHAIN CONSTRUCTED OF THREE SIMPLER CHAINS INTERLOCKED IN PARALLEL; HE ENJOYED WATCHING MICROSCOPIC ECOSYSTEMS (CAPTURED WITH FISHPOLE AND BOTTLE FROM THE BONEYARD CREEK IN URBANA) FOR THE RICHNESS OF INTERACTION THEY DISPLAYED, AND HE BUILT A SEMI-RANDOM ELECTRONIC CONTRAPTION WITH 100 DOUBLE TRIODES AND WATCHED IT FOR TWO YEARS BEFORE ADMITTING DEFEAT IN THE FACE OF ITS INCOMPREHENSIBLY COMPLEX BEHAVIOR.

  ROGER CONANT,"W. ROSS ASHBY (1903–1972)" (1974, 4)

  Figure 4.8.Photograph of DAMS. (By permission of Jill Ashby, Sally Bannister, and Ruth Pettit.)

  The 1952 first printing of Design for a Brain included just one footnote: on page 171 Ashby revealed that he was building a machine called DAMS. In the 1954second printing of the first edition the footnote was removed, though the entry for DAMS could still be found in the index and a citation remained on page 199 to the only publication in which Ashby described this device, the paper "Statistical Machinery" in the French journal Thalès (Ashby 1951). In the second edition, of 1960, both the index entry and the citation also disappeared: DAMS had been purged from history. Despite the obscurity to which Ashby was evidently determined to consign it, his journal in the 1950s, especially from 1950 to 1952, is full of notes on this machine. It would be a fascinating but terribly demanding project to reconstruct the history of DAMS in its entirety; I will discuss only some salient features.

  I opened the book with Ashby's suggestion that "the making of a synthetic brain requires now little more than time and labour" (1948, 382), and he evidently meant what he said. DAMS was to be the next step after the homeostat. Its name was an acronym for dispersive and multistable system. A multistable system he defined as one made up of many interconnected ultrastable systems. A dispersive system was one in which different signals might flow down different pathways (Ashby 1952, 172). This gets us back to the above discussion of times to reach equilibrium. Ashby conceived DAMS as a system in which the ultrastable components were linked by switches, which, depending on conditions, would either isolate components from one another or transmit signals between them. In this way, the assemblage could split into smaller subassemblies appropriate to some adaptive task without the patterns of splitting having to be hard wired in advance. DAMS would thus turn itself into a sparsely connected system that could accumulate adaptations to differing stimuli in a finite time (without disturbing adaptive patterns that had already been established within it).

  At the hardware level, DAMS was an assemblage of electronic valves, as in a multihomeostat setup, but now linked not by simple wiring but by neon lamps. The key property of these lamps was that below some threshold voltage they were inert and nonconducting, so that they in fact isolated the valves that they stood between. Above that threshold however, they flashed on and became conducting, actively joining the same valves, putting the valves in communication with one another. According to the state of the neons, then, parts of DAMS would be isolated from other parts by nonconducting neons, "walls of constancy," as Ashby put it (1952, 173), and those parts could adapt independently of one another at a reasonable, rather than hyperastronomical, speed.

  Not to leave the reader in undue suspense, I can say now that DAMS never worked as Ashby had hoped, and some trace of this failure is evident in the much-revised second edition of Design for a Brain. There Ashby presents it as a rigorous deduction from the phenomenon of cumulative adaptation to different stimuli, P1, P2, and so on, that the step mechanisms (uniselectors in the homeostat, neon tubes in DAMS) "must be divisible into non-overlapping sets, that the reactions to P1 and P2 must each be due to their particular sets, and that the presentation of the problem (i.e., the value of P) must determine which set is to be brought into functional connexion, the remainder being left in functional isolation" (1960, 143). One can see how this solves the problem of accumulating adaptations, but how is it to be achieved? At this point, Ashby wheels on his deus ex machina, a "gating mechanism," , shown in figure 4.9. This picks up the state of the environmental stimulus P via the reaction R of the organism to it and switches in the appropriate bank of uniselectors, neons, or whatever that the essential variables (the dial on the right) can trigger, if necessary, to preserve the equilibrium of the system. But then the reader is left hanging: What is the go of this gating mechanism? How does it do its job? Almost at the end of the book, eighty-four pages later, Ashby acknowledges that "it was shown that . . . a certain gating-mechanism was necessary; but nothing was said about how the organism should acquire one" (1960, 227). Two pages later, Ashby fills in this silence, after a fashion (1960, 229–30): "The biologist, of course, can answer the question at once; for the work of the last century . . . has demonstrated that natural, Darwinian, selection is responsible for all the selections shown so abundantly in the biological world. Ultimately, therefore, these ancillary mechanisms [the gating mechanism and, in fact, some others] are to be attributed to natural selection. They will, therefore, come to the individual (to our kitten perhaps) either by the individual's gene-pattern or they develop under an ultrastability of their own. There is no other source." Within the general framework of Ashby's approach to the brain and adaptation, these remarks make sense. We need a gating mechanism if multiple adaptations are to be achieved in a finite time; we do adapt; therefore evolution must have equipped us with such a mechanism. But what Ashby had been after with DAMS was the go of multiple adaptation. What he wanted was that DAMS should evolve its own gating mechanism in interacting with its environment, and it is clear that it never did so. To put the point the other way around, what he had discovered was that the structure of the brain matters—that, from Ashby's perspective, a key level of organization had to be built in genetically and could not be achieved by the sort of trial-and-error self-organization performed by DAMS.28

  Figure 4.9.The gating mechanism. Source: W. R. Ashby, Design for a Brain: The Origin of Adaptive Behaviour, 2nd ed. (London: Chapman & Hall, 1960), 144, fig. 10/9/1. (With kind permission from Springer Science and Business Media.)

  _ _ _ _ _

  Though DAMS failed, Ashby's struggles with it undoubtedly informed his understanding of complex mechanisms and the subsequent development of his cybernetics, so I want to pursue these struggles a little further here.29 First, I want to emphasize just how damnably complicated these struggles were. DAMS first appeared in Ashby's journal on 11 August 1950 (pp. 2953–54) with the words "First, I might as well record my first idea for a new homeostat [and, in the margin] found a month ago." The next note, also dated 11 August 1950, runs for twenty pages (pp. 2955–74) and reveals some of the problems that Ashby h
ad already run into. It begins, "For a time the construction of the new machine (see previous page) went well. Then it forced me to realise that my theory had a yawning hole in it" (p. 2955).

  This yawning hole had to do with DAMS's essential variables, the parameters it should control. In the original homeostat setups all of the currents were essential variables, capable of triggering discontinuous changes of state via the relays and uniselectors. But there was no reason why all of the currents in DAMS should be essential variables. Some of them should be, but others would have simply to do with making or breaking connections. Thus, a new problem arose: how the environment should be supposed to connect to DAMS's essential variables, and how those variables might act back onto the environment.30 The homeostat offered no guidance on this, and the remainder of this entry is filled with Ashby's thoughts on this new problem. It contains many subsequently added references to later pages which develop these early ideas further. In a passage on page 2967, for example, one thought is linked by an asterisk to a note at the bottom of the page which says, "May '51. Undoubtedly sound in aim, but wrong in the particular development used here," while in the margin is a note in black ink, "Killed on p. 2974," and then another note, "Resurrected p. 3829," in red. The next paragraph then begins, "This was the point I reached before I returned to the designing of the electrical machine, but, as usual, the designing forced a number of purely psychological problems into the open. I found my paragraph (2) (above) [i.e., the one just discussed here] was much too vague to give a decisive guide." The penultimate paragraph of the entire note ends (p. 2974), "I see no future this way. The idea of p. 2967 (middle) [i.e., again the one under discussion here] seems to be quite killed by this last figure." But then a marginal note again says, "Resurrected p. 3829" (i.e., 17 May 1952).

  The substantial point to take from all this is that the construction of DAMS posed a new set of problems for Ashby, largely having to do with the specification of its essential variables and their relation to the environment, and it was by no means clear to him how to solve them.31 And what interests me most here is that in response to this difficulty, Ashby, if only in the privacy of his journal, articulated an original philosophy of design.

  "The relation of the essential variables to a system of part-functions [e.g., the neon tubes] is still not clear, though p. 3074 helps. Start again from first principles," Ashby instructed himself on 28 January 1951, but a second note dated the same day recorded that DAMS was "going to be born any time" (pp. 3087–8). Six weeks later Ashby recorded that "DAMS has reached the size of ten valves, and," he added, "has proved exceedingly difficult to understand." He continued (14 March 1951, pp. 3148–51),

  But while casting around for some way of grasping it I came across a new idea. Why not make the developent of DAMS follow in the footsteps marked out by evolution, by making its variations struggle for existence? We measure in some way its chance of "survival," and judge the values of all proposed developments by their effects on this chance. We know what "survival" means in the homeostat: we must apply the same concept to DAMS. . . .

  The method deserves some comment. First notice that it totally abandons any pretence to "understand" the assembly in the "blue-print" sense. When the system becomes highly developed the constructor will be quite unable to give a simple and coherent account of why it does as it does. . . . Obviously in these circumstances the words "understand" and "explain" have to receive new meanings.

  This rejection of the "blue-print" attitude corresponds to the rejection of the "blue-print" method in the machine itself. One is almost tempted to dogmatise that the Darwinian machine is to be developed only by the Darwinian process! (there may be more in this apothegm than a jest). After all, every new development in science needs its own new techniques. Nearly always, the new technique seems insufficient or hap-hazard or plain crazy to those accustomed to the old techniques.

  If I can, by this method, develop a machine that imitates advanced brain activities without my being able to say how the activities have arisen, I shall be like the African explorer who, having heard of Lake Chad, and having sought it over many months, stood at last with it at his feet and yet, having long since lost his bearings, could not say for the life of him where in Africa Lake Chad was to be found.

  This is a remarkable passage of ontological reflection, which gets us back to the cybernetic discovery of complexity from a new angle. Like Walter's tortoise, the homeostat had been designed in detail from the ground up—the blueprint attitude—and this approach had been sufficient, inasmuch as the two machines did simulate performances of the adaptive brain. My argument was, however, that when constructed, they remained to a degree impermeable Black Boxes, displaying emergent properties not designed into them (the tortoise), or otherwise opaque to analysis (the multihomeostat setup). But it was only with DAMS that Ashby had to confront this discovery of complexity head-on. And in this passage, he takes this discovery to what might be its logical conclusion. If, beyond a certain degree of complexity, the performance of a machine could not be predicted from a knowledge of its elementary parts, as proved to be the case with DAMS, then one would have to abandon the modern engineering paradigm of knowledge-based design in favor of evolutionary tinkering—messing around with the configuration of DAMS and retaining any steps in the desired direction.32 The scientific detour away from and then back to performance fails for systems like these.

  The blueprint attitude evidently goes with the modern ontological stance that presumes a knowable and cognitively disposable world, and Ashby's thoughts here on going beyond design in a world of mechanisms evolving quasi-organically once more make the point that ontology makes a difference, now at the level of engineering method. We can come back to this point in later chapters.

  _ _ _ _ _

  Ashby never reached the shores of Lake Chad, but one feature of DAMS's performance did become important to his thinking: a behavior called "habituation." In his only published discussion of DAMS, after a discussion of DAMS itself, Ashby turns to a theoretical argument, soon to appear in Design for a Brain, that he claims is generally applicable to any "self-switching network, cortex or D. A. M. S. or other, . . . no matter in what random pattern the parts are joined together and no matter in what state its 'memories' have been left by previous activities." This argument has two parts: first, that a system like DAMS will naturally split itself up into subsystems that "tend to be many and small rather than few and large"; and second, that such a system becomes habituated to a repeated stimulus, inamsuch as "it will tend to set its switches so that it is less, rather than more, disturbed by it. "Then Ashby returns to his machines, noting first that the latter effect had been demonstrated on the homeostat, where, indeed, it is true almost by definition: the first application of any stimulus was liable to provoke a large response—the tripping of the unselectors—while once the homeostat had found an equilibrium configuration, its response to the same stimulus would be small: a damped oscillation returning to the equilibrium state. By 1951, Ashby could also remark that this property "is already showing on the partly-constructed D. A. M. S." (1951, 4, 5; Ashby's italics).

  Ashby regarded habituation in his machines as support for his general approach to the brain. "In the cerebral cortex this phenomenon [of diminishing response to a stimulus] has long been known as 'habituation.' It is in fact not restricted to the cerebral cortex but can be observed in every tissue that is ca pable of learning. Humphrey considers it to be the most fundamental form of learning" (1951, 5). But, as Ashby put it in Design for a Brain ,"The nature of habituation has been obscure, and no explanation has yet received general approval. The results of this chapter suggest that it is simply a consequence of the organism's ultra-stability, a by-product of its method of adaptation" (1952, 152).33 The significance of this observation is that Ashby had gone beyond the simple mimicry of adaptation to a novel result—discovering the go of a phenomenon that hitherto remained mysterious.34 And in his journals, Ashby took this line of thought still
further. Reflecting on DAMS on 22 May 1952 (p. 3829), he arrived at an analysis of "dis-inhibition" (he writes it in quotes): "The intervention of a second stimulus will, in fact, restore the -response to its original size. This is a most powerful support to my theory. All other theories, as far as I know, have to postulate some special mechanism simply to get dis-inhibition."35

  If DAMS never reached the promised land and Ashby never quite reached Lake Chad, then, certainly the DAMS project led to this one substantive result: an understanding of habituation and how it could be undone in ultrastable machines. We can come back to this result when we return to Ashby's psychiatric concerns.

  _ _ _ _ _

  I can add something on the social basis of Ashby's research in the DAMS era and its relation to the trajectory of his research. In the early 1950s, Pierre de Latil visited the leading cyberneticians of the day, including Walter as well as Ashby, and wrote up a report on the state of play as a book, Thinking by Machine: A Study of Cybernetics, which appeared in French in 1953 and in English in 1956, translated by Frederick Golla's daughter, Yolande. De Latil recorded that "Ashby already considers that the present DAMS machine is too simple and is planning another with even more complex action. Unfortunately, its construction would be an extremely complex undertaking and is not to be envisaged for the present" (de Latil 1956, 310). I do not know where the money came from for the first versions of DAMS, but evidently cost became a problem as Ashby began to aim at larger versions of it. On an ill-starred Friday the 13th in September 1957, Ashby noted to himself, "As the RMPA [Royal Medico-Psychological Association] are coming to B. H. [Barnwood House] in May 1960 I have decided to get on with making a DAMS for the occasion, doing as well as I can on the money available. By building to a shoddiness that no commercial builder would consider, I can probably do it for far less than a commercial firm would estimate it at." Clearly, by this time Ashby's hobby was turning into a habit he could ill afford and remained a hobby only for lack of institutional support.36 That his work on DAMS had lapsed for some time by 1957 is evident in the continuation of the note: "In addition, my theoretical grasp is slowly getting bogged down for lack of real contact with real things. And the deadline of May 1960 will force me to develop the practical & immediate" (p. 5747).