The Cybernetic Brain

by Andrew Pickering

  Figure 6.4.Schematic of the cybernetic factory. Source: Beer 1994a, 192.

  Beer envisaged the T-machine as something like Pitts and McCulloch's scanning device (Pitts and McCulloch 1947, discussed in chap. 3) updated in the light of more recent neurophysiological research. The "senses" of the T-machine would be numerical inputs representing the state of the factory's environmment (supplies and orders, finance) and its internal state (stocks, performance measures, etc.). The function of the T-machine was "scansion, grouping and pattern recognition" (Beer 1962a, 173). It would, that is, turn atomistic raw data into a meaningful output, in much the same way as the human brain picks out "universals" from our sensory data. The V-machine was conceived essentially as a T-machine running backward. Its inputs would be framed in the language of T-machine outputs; its outputs would be instructions to the motor organs of the plant—directing production operations and flows, ordering stock, or whatever.

  Between the T- and V-machines lay, yes, the U-machine. The U-machine was to be "some form of Ashbean ultrastable machine" (Beer 1962a, 189)—a homeostat, the brain artifact of the firm. The job of the U-machine was continually to reconfigure itself in search of a stable and mutually satisfactory relationship between the firm and its environment. The U-machine was thus the organ that would enable the factory to cope with an always fluctuating and changing, never definitively knowable environment. It was the organ that could take the automatic factory to a level of consciousness beyond that of a spinal dog. Figure 6.5 summed up Beer's abstract presentation, accompanied by the words "The temptation to make the outline look like a coronal section of the living brain was irresistible and I apologize to cerebra everywhere for such insolence" (197).9

  The second major section of Beer's essay was a progress report on how far he had gone toward realizing a cybernetic factory at the Templeborough Rolling Mills, a division of United Steel engaged in the manufacture of steel rods.10 This can help us think more concretely about the cybernetic factory, and here we need to refer to figure 6.6. The top level of the diagram represents various material systems relating to the flow of steel within the plant and their interconnections: the "Supplying system" feeds the "Input stocking system" which feeds the "Producing system." and so on. The next level down, "Sensations," is the most important. Nineteen "sensations" are shown in the diagram, running from "a. tons bought" to "s. tons requested." Each of these sensations should be understood as taking the form of numerical data relating to aspects of the plant or its environment—the current state of production, the profit and loss account, the balance sheet, as shown in lower levels of the figures. The "sensation" aspect of this diagram relates to the T-machine of Beer's formal discussion, and his claim was to have sufficiently simulated a T-machine to make it clear that an automatic one could be built. The grouping of data into nineteen categories, for example, entailed "a large number of decisions . . . which, ultimately, the brain artefact itself is intended to take by its multiple multiplexing techniques. The research team in the field has, however, taken these decisions on an informed basis, by operational research methods" (Beer 1962a, 202).

  Figure 6.5.The cybernetic factory as brain. Painting by Stafford Beer. The T, U, and V machines are labeled on the smaller painting in the bottom left. Source: Beer 1994a, 198, fig. 3.

  Figure 6.6.The steel mill as cybernetic factory. Source: Beer 1994a, 200–201, fig. 4.

  The "sensations," then, were to be considered inputs to the T-machine, and further numerical transformations were supposed to correspond to the functioning of "the T-Machine proper" (Beer 1962a, 203). These transformations, derived in practice from OR studies, first recombined the nineteen sensations into twelve "functions"—six referring primarily to the company and six to its environment. The functions all depended on ratios of expected behavior to actual behavior of precisely the form of the indices developed in Beer's earlier OR work, discussed above. "This last point," Beer wrote (204–5),

  is important, since it incorporates in this exemplification the essential "black box" treatment of unknowns and imponderables common to all cybernetic machines. For a model of performance in any field may be inadequate: predictions and judgements based upon it will be effectual only insofar as the model isadequate. But in exceedingly complex and probabilistic systems no analytic model can possibly be adequate. The answer to this paradox, which I have used successfully for 10 years, is to load the raw predictions of any analytic model with a continuous feedback measuring its own efficiency as a predictor. In this way, everything that went unrecognized in the analytic work, everything that proved too subtle to handle, even the errors incurred in making calculations, is "black boxed" into an unanalyseable weighting which is error-correcting.

  Here, then, we have an example of one way in which Beer's cybernetics tried to handle the unknown—a predictor that reviewed its own performance in the name of predicting better.11

  The values of the twelve parameters were measured daily in the steel mill and "were plotted on boards in an Operations Room for the benefit of management, as a by-product of this research" (Beer 1962a, 205). A plot of a year's readings is shown in figure 6.7, which Beer referred to as an encephalogram (205). He was reaching here for a suggestive connection between his work in management and brain science à la Grey Walter, referring to emergent periodicities in the data and noting that the "encephalographer finds this structural component of information (the brain rhythm) of more importance than either its amplitude or voltage" (182). This tempting idea seems to have proved a red herring, alas; I am not aware of any subsequent development of it, by Beer or anyone else. Several other, readily automatable statistical and mathematical transformations of these data then followed, and the work of the T-machine, as simulated at Templeborough, was said to be complete. Given that "the T-Machine was said to be set-theoretically equivalent to a V-Machine," the problem of constructing the latter could be said to have been shown to be soluble, too (208). But figure 6.4 also shows the intervention of the U-machine, the homeostatic brain, into the life of the cybernetic factory: what about that?

  Figure 6.7.An EEG of the firm. Source: Beer 1994a, 206, fig. 5.

  The outputs of the simulated T-machine in successive time steps were recorded at Templeborough as a "generalized gestalt memory" indicated in the lower left and right of figure 6.6, the left portion relating to inner states of the factory, the right to its environment. These memories could be thought of defining "two phase spaces in which the company and the environment can respectively operate." And the U-machine was intended to search for a set of "preferred states" within this space via a "mutually vetoing system by which the homeostatic loop in the diagram continues to operate until both company and environmental points in phase-space (representing vectors of functions) lie in the appropriate preferred states set" (Beer 1962a, 208).12 This notion of mutual or reciprocal vetoing was very important in Beer's work (and Pask's), so I want to digress briefly here to explain it.

  The idea of mutual vetoing came directly from Ashby's cybernetics, and here Beer, like Bateson and Pask, took the symmetric fork in the road. Imagine an interconnected setup of just two of Ashby's homeostats, both of which are free to reconfigure themselves. Suppose homeostat 1 finds itself in an unstable situation in which its essential variable goes out of whack. In that case, its relay trips, and its uniselector moves to a new setting, changing the resistance of its circuit. Here one can say that homeostat 2—with its own internal parameters that define the transformation between its input from and output to homeostat 1—has vetoed the prior configuration of homeostat 1, kicking it into a new condition. And likewise, of course, when homeostat 2 finds itself out of equilibrium and changes to a new state, we can say that homeostat 1 has vetoed the first configuration of homeostat 2. Eventually, however, this reconfiguration will come to an end, when both homeostats achieve equilibrium at once, in a condition in which the essential variables of both remain within limits in their mutual interactions. And this equilibrium, we can then say, i
s the upshot of a reciprocal vetoing: it is the condition that obtains when the vetoing stops and each machine finds a state of dynamic equilibrium relative to the other's parameters.

  This is enough, I think, to unravel the above quotation from Beer. One can think of the U-machine and the firm's environment as two reciprocally vetoing homeostats, and the U-machine itself attempts to find a relation between its inputs from the T-machine and its outputs to the V-machine that will keep some essential variable standing for the "health" of the company within limits. Beer never reached the stage of defining exactly what that essential variable should be at this stage in his work. For the sake of concreteness, we could imagine it as a measure of profitability, though Beer proposed interestingly different measures in subsequent projects that we can review below.

  It was clear enough, then, what the U-machine should do, though in 1960 Beer still had no clear vision of how it should be made, and at Templeborough "management itself," meaning the actual human managers of the plant, "plays the role of the U-Machine" (Beer 1962a, 208). The state of the art was thus that by that date a cybernetic factory had been simulated, though not actually built. Beer was confident that he could construct automated versions of the T-Machine, as the factory's sensory organ, and the V-machine, as its motor-organ equivalent. Neither of these had actually been constructed, but their working parts had been simulated by OR studies and data collection and transformation procedures. The U-machine, which figured out the desirable place for the factory to sit in the factory-environment phase space, continued to be purely human, simulated by the managers who would review the "gestalt memory" generated by the T-machine and figure out how to translate that into action via the inputs to the V-machine. The U-machine, then, was the key (209):

  As far as the construction of cybernetic machinery is concerned, it is clear that the first component to transcend the status of mere exemplification must be the U-Machine. For exemplifications of T- and V-input are already available, and can be fed to a U-Machine in parallel with their equivalent reporting to management. . . . Having succeeded in operating the cybernetic U-Machine, the research will turn to constructing cybernetic T- and V-Machines. . . . After this, management would be free for the first time in history to manage, not the company in the language of the organism, but the T-U-V(R) control assembly in a metalanguage.

  But what was the U-machine to be? Beer ended his talk at Allerton Park with the words "Before long a decision will be taken as to which fabric to use in the first attempt to build a U-Machine in actual hardware (or colloid, or protein)" (212). Colloid or protein?

  Biological Computing

  Beer's thinking about the U-machine was informed by some strikingly imaginative work that he and Pask engaged in in the 1950s and early 1960s, both separately and together—work that continued Ashby's goal of a synthetic brain but with an original twist. Ashby had built an adaptive electromagnetic device, the homeostat, which he argued illuminated the go of the adaptive brain. Following his lead, Beer and Pask realized that the world is, in effect, already full of such brains. Any adaptive biological system is precisely an adaptive brain in this sense. This does not get one any further in understanding how the human brain, say, works, but it is an observation one might be able to exploit in practice. Instead of trying to build a superhomeostat to function as the U-machine—and Beer must have known in the mid-1950s that Ashby's DAMS project was not getting far—one could simply try to enroll some naturally occurring adaptive system as the U-machine. And during the second half of the 1950s, Beer had accordingly embarked on "an almost unbounded survey of naturally occurring systems in search of materials for the construction of cybernetic machines" (Beer 1959, 162). The idea was to find some lively system that could be induced to engage in a process of reciprocal vetoing with another lively system such as a factory, so that each would eventually settle down in some agreeable sector of its environment (now including each other).

  In 1962 Beer published a brief and, alas, terminal report on the state of the art, which makes fairly mind-boggling reading (Beer 1962b), and we can glance at some of the systems he discussed there to get a flavor of this work. The list begins with quasi-organic electrochemical systems that Beer called "fungoids," which he had worked on both alone and in collaboration with Pask. This was perhaps the aspect of the project that went furthest, but one has to assume Pask took the lead here, since he published several papers in this area in the late 1950s and early 1960s, so I postpone discussion of these systems to the next chapter. Then follows Beer's successful attempt to use positive and negative feedback to train young children (presumably his own) to solve simultaneous equations without teaching them the relevant mathematics—to turn the children into a performative (rather than cognitive) mathematical machine. Beer then moves on to discuss various thought experiments involving animals (1962b, 28–29):

  Some effort was made to devise a "mouse" language which would enable mice to play this game—with cheese as a reward function. . . . In this way I was led to consider various kinds of animal, and various kinds of language (by which I mean intercommunicating boxes, ladders, see-saws, cages connected by pulleys and so forth). Rats and pigeons have both been studied for their learning abilities. . . . The Machina Speculatrix of Grey Walter might also be considered (with apologies to the organic molecule). . . . However no actual machines were built. . . . By the same token, bees, ants, termites, have all been systematically considered as components of self-organizing systems, and various "brainstorming" machines have been designed by both Pask and myself. But again none has been made.

  Figure 6.8.The Euglena homeostat. Square, Euglena culture, with tropism displayed as shown; solid diamond, stimulus; circle, sensory receptor; hatched triangle, inhibiting influence, and, open triangle, stimulating influence, of a's sensation on b's stimulus. Source: Beer 1994a, 30, fig. 2.

  Beer had, however, devoted most of his own efforts to systems composed from simpler organisms: colonies of Daphnia, a freshwater crustacean (Pask had considered Aedes aegypti, the larva of the yellow fever mosquito), of Euglena protozoa, and an entire pond ecosystem. The key question with all three systems was how to interest these biological entities in us, how to couple them to our concerns, how to make a U-machine that would respond to and care about the state of the cybernetic factory. And this coupling was where Beer's attempts foundered (1962b, 29):

  Many experiments were made with [Daphnia].Iron filings were included with dead leaves in the tank of Daphnia, which ingested sufficient of the former to respond to a magnetic field. Attempts were made to feed inputs to the colony of Daphnia by transducing environmental variables into electromagnets, while the outputs were the consequential changes in the electrical characteristics of the phase space produced by the adaptive behaviour of the colony. . . . However, there were many experimental problems. The most serious of these was the collapse of any incipient organization—apparently due to the steadily increasing suspension of tiny permanent magnets in the water.

  Euglena are sensitive to light (and other disturbances) in interesting ways, and Beer sought to achieve optical couplings to a tank full of them "using a point source of light as the stimulus, and a photocell [to measure the absorption of light by the colony] as the sensory receptor" (fig. 6.8).

  However, the culturing difficulties proved enormous. Euglena showed a distressing tendency to lie doggo, and attempts to isolate a more motile strain failed. So pure cultures were difficult to handle. Moreover, they are not, perhaps, ecologically stable systems. Dr. Gilbert, who had been trying to improve the Euglena cultures, suggested a potent thought. Why not use an entire ecological system, such as a pond? . . . Accordingly, over the past year, I have been conducting experiments with a large tank or pond. The contents of the tank were randomly sampled from ponds in Derbyshire and Surrey. Currently there are a few of the usual creatures visible to the naked eye (Hydra, Cyclops, Daphnia, and a leech); microscopically there is the expected multitude of micro-organisms. [The coupling is via
light sources and photocells, as in the Euglena experiments.] . . . The state of this research at the moment is that I tinker with this tank from time to time in the middle of the night. (Beer 1962b, 31–32)

  Clearly, however, Beer failed to enroll the pond ecosystem, too, as a U-machine. The cybernetic factory never got beyond the simulation stage; we do not live in a world where production is run by Daphnia and leeches, and Beer's 1962 status report proved to be a requiem for this work. I now want to comment on it ontologically and sociologically, before moving on to later phases in Beer's career in management.

  Ontology and Design

  The sheer oddity of trying to use a pond to manage a factory dramatizes the point that ontology makes a difference. If one imagines the world as populated by a multiplicity of interacting exceedingly complex systems, as modelled by Ashby's homeostats, then one just might come up with this idea. It follows on from what has gone before, though even then some sort of creative leap is required. In contrast, it is hard to see how one would ever come to think this way from a modern technoscientific perspective. One would think instead of trying to program a computer to do the job of management, but that is a very different approach, in ways that are worth pondering.