A Cascade of Homunculi

20. A CASCADE OF HOMUNCULI

In the millennia-old quest to understand the mind, theorists have often succumbed to the temptation to imagine an inner agent, a little man—homunculus, in Latin—who sits in the control room in the brain and does all the clever work. If you think of the human nervous system as, say, a huge telephone switchboard network (as thinkers liked to do as recently as the 1950s and 1960s), you have the problem of the telephone operator at the center: Is his or her mind composed of a smaller telephone network with its own central operator, whose mind in turn is composed of . . . ? Any theory that posits such a central homunculus is doomed by the prospect of an infinite regress.

But maybe the mistake isn’t postulating a homunculus, but postulating a central homunculus. In my first book, Content and Consciousness (1969), I made a big blunder, tempted by a wisecrack I couldn’t resist. I wrote,

The “little man in the brain”, Ryle’s “ghost in the machine”, is a notorious non-solution to the problems of mind, and although it is not entirely out of the question that the “brain writing” analogy will have some useful application, it does appear merely to replace the little man in the brain with a committee. [p. 87]

And what, exactly, is wrong with a committee? (Aha—my implied reductio is called out!) I eventually came to realize (in Brainstorms, 1978a) that the idea of replacing the little man in the brain with a committee is not such a bad one; it is, I think, one of the fundamental good ideas of cognitive science. This was the classic strategy of GOFAI (good old-fashioned artificial intelligence; Haugeland, 1985) that came to be known as homuncular functionalism:

The AI programmer begins with an intentionally characterized problem, and thus frankly views the computer anthropomorphically: if he solves the problem he will say he has designed a computer that can [e.g.,] understand questions in English. His first and highest level of design breaks the computer down into subsystems, each of which is given intentionally characterized tasks; he composes a flow chart of evaluators, rememberers, discriminators, overseers and the like. These are homunculi with a vengeance. . . . Each homunculus in turn is analyzed into smaller homunculi, but, more important, into less clever homunculi. When the level is reached where the homunculi are no more than adders and subtractors, by the time they need only the intelligence to pick the larger of two numbers when directed to, they have been reduced to functionaries who can be replaced by a machine. [p. 80]

The particular virtue of this strategy is that it pulled the rug out from under the infinite regress objection. According to homuncular functionalism the ominous infinite regress can be sidestepped, replaced by a finite regress that terminates, as just noted, in operators whose task is so dull they can be replaced by machines. The key insight was breaking up all the work we imagined being done by a central operator and distributing it around to lesser, stupider agents whose work was distributed in turn, and so forth.

This was a fine advance, but an unwanted artifact of the top-down approach of classic GOFAI was its rigid bureaucratic efficiency! We may have been able to imagine getting rid of the central king or the CEO, but we still had an army of midlevel executives reporting to an array of vice presidents (whose interactions constituted the highest level of the system) and ordering their subordinates into action, who in turn would call up more menial clerical workers, and so forth. This hyper-efficient organization, with no wasted motion, no feather-bedding, no insubordination, was largely dictated by the fact that the large computers on which the early AI models were developed were tiny and slow by today’s standards, and people wanted results fast. If you needed to impress the funding agency, your AI had better not take hours to answer a simple question. It should be businesslike in its execution. Besides, writing thousands of lines of code is a lot of work, and if you have succeeded in breaking down the target task—answering questions about moon rocks or diagnosing kidney diseases or playing chess, for instance—into a manageable set of subtasks that you can see how to program and then integrate into a workable system, you can have your “proof of concept”1 at a feasibly low cost in terms of time and money.

Notice that computers have always been designed to keep needs and job performance almost entirely independent. Down in the hardware, the electric power is doled out evenhandedly and abundantly; no circuit risks starving. At the software level, a benevolent scheduler system doles out machine cycles to whatever process has the highest priority, and although there may be a bidding mechanism of one sort or another that determines which processes get priority, this is an orderly queue, not a struggle for life. As Marx would have it, “From each according to his abilities, to each according to his needs.” The computer scientist Eric Baum has aptly dubbed this hierarchy “politburo” control. Probably a dim appreciation of this fact underlies the common folk intuition that a computer could never care about anything. Not because it is made out of the wrong materials—why should silicon be any less suitable a substrate for caring than carbon?—but because its internal economy has no built-in risks or opportunities, so it doesn’t have to care.

Neurons are not like this. The general run of the cells that comprise our bodies are probably just willing slaves—rather like the selfless, sterile worker ants in a colony, doing stereotypic jobs and living out their lives in a relatively noncompetitive (Marxist) environment. But brain cells—I now think—must compete vigorously in a marketplace. For what? What could a neuron want? The energy and raw materials to continue to thrive—just like its unicellular eukaryote ancestors and more distant cousins, the bacteria and archaea. Neurons are biological robots of sorts; they are certainly not conscious in any rich sense. Remember, they are eukaryotic cells, akin to yeast cells or fungi. If individual neurons are conscious, then so is athlete’s foot! But neurons, like their mindless unicellular cousins, are highly competent agents in a life-or-death struggle, not in the environment between your toes, but in the demanding environment of the brain, where the victories go to the cells that can network more effectively, contribute to more influential trends at the virtual machine levels where large-scale human purposes and urges are discernible. Many of the subsystems in the nervous system are organized as opponent processes, engaged in a tug-of-war between two sub-subsystems, each trying to have its own way. (Our emotions, for instance, are well seen as rival storms displacing each other as best they can, thwarting each other or collaborating against yet another storm.) I now think, then, that the opponent-process dynamics of emotions, and the roles they play in controlling our minds, are underpinned by an economy of neurochemistry that harnesses the competitive talents of individual neurons. (Note that the claim is that neurons are still good team-players within the larger economy, unlike the more radically selfish agents, cancer cells. I recall Nobel laureate biologist François Jacob’s dictum that the dream of every cell is to become two cells; neurons vie to stay active and to be influential, but do not dream of multiplying.) In this view, intelligent control of an animal’s behavior is still a computational process—in the same way that a transaction in the stock market is a computational process—but the neurons are selfish neurons, as neuroscientist Sebastian Seung (2007) has said, striving to maximize their intake of the different currencies of reward we have found in the brain. And what do neurons buy with their dopamine, serotonin, or oxytocin? They are purchasing greater influence in the networks in which they participate, and hence greater security. (The fact that mules are sterile doesn’t stop them from fending for themselves, and neurons can similarly be moved by self-protective instincts they inherited ultimately from their reproducing ancestors.)

So bottom-up, neuroscientifically inspired, homuncular functionalism is looking better and better as a model of how the brain works, since the more chaotic and competitive “computational architectures” it generates are more plausible from a biological point of view: we can begin to discern the developmental processes that could build and revise these architectures, starting in the embryo and continuing into adulthood, and also see how they could have evolved from simpler nervous systems, themselves teams of less accomplished homunculi, which sorta perceive, signal, and remember.2

---

1 Proof of concept is a useful but dangerous bit of engineering jargon: asked to design a very sophisticated and complicated widget, you design a very simple widget—you solve what is nicely called a toy problem—and when it works on its stripped-down task, you declare that you have “proof of concept” and now it’s “just” a matter of scaling up to the fully competent widget that was requested, which will require only more time and money; the hard, conceptual problem has been “solved.” Sometimes this works just as advertised.

2 Not everybody likes this idea. The book Philosophical Foundations of Neuroscience (2003), the collaborative effort of a neuroscientist, Max Bennett, and a philosopher, P. M. S. Hacker, heaps scorn on it. For a detailed rebuttal (and their rebuttal to my rebuttal), see Bennett et al., 2009.