The Brain as a Computer

For hundreds of years we’ve had plenty of evidence that the brain is somehow the seat of the soul, but until the middle of the twentieth century it was all but unimaginable how this could possibly be true. The brain could be seen to be composed of lots of different, curiously shaped organs, in left and right pairs, given vivid names by early anatomists—the hippocampus (seahorse), the amygdala (almond), the wrinkly cortex (bark)—but what on earth did these parts do? They didn’t digest food or purify the blood, did they? Was the brain just an organ for cooling the blood, a sort of radiator, as Aristotle thought? The parts were connected by nerve fibers, so perhaps they communicated with each other somehow. Descartes suggested that some nerve fibers were like bell-pull wires—when you tugged at one end, something happened at the other end, but what, exactly? A bell ringing wouldn’t seem to get you any closer to understanding the brain as the mind, and nobody had any better ideas.1

Then along came Turing, building on a tradition that went back to Babbage and Pascal and Leibniz and others, with the suggestion that a brain could be composed of simple parts that were ultimately just mechanical (like a mousetrap, a ringing bell, a lock and key, a synapse), but if these parts were organized to interact together in clever ways, they might actually be able to do something intelligent on their own, without human intervention, and without any ghosts in the machine to guide them along. They might compute. Before Turing hit on his idea, “computer” was a job title; industry and government hired thousands of people to be computers, calculating tables for use in business, navigation, gunnery, and banking, for instance. Maybe, he surmised, the brain is itself like a (human) computer; maybe it processes information by slavishly following huge lists of very simple instructions (like Inc and Deb, shorthand for incrementing or decrementing a value and branching). Once the early theorists of cognitive science, Alan Turing and John von Neumann, the creator of cybernetics, Norbert Wiener, and the creator of information theory, Claude Shannon, as well as many others, articulated the idea, it could even seem obviously right—how could we not have seen it all along? Brains must take in information from the sense organs and process it, by somehow computing with it, until they have extracted the valuable ore of meaning, which, by further computing, they could then categorize and store for future use in guiding the behavior of the body, which provided the brain with its energy and protective home. The key innovation in Turing’s vision was that it eliminated the awkward element that was all too evident in earlier imaginings of information-processing: junctures that required a clerk or translator or librarian, that required, in short, some kind of understander to appreciate the import of the signals. Turing saw that in one sense this was inescapable: intelligent processes would always require choosing one course or another on the basis of the discrimination of some difference in the signal. But he could reduce this understanding to the barest minimum: conditional branching, the mindless process by which a device decides (sorta decides) to go left rather than right because it senses (sorta senses) 1 rather than 0, A rather than B, x rather than y. That, and arithmetic, were all you needed. You could then build up devices of any level of discernment by piling virtual machines on top of virtual machines on top of virtual machines—to put it anachronistically. This vision has been tempting for more than fifty years, but the details are not falling into place easily. If brains are computers, they aren’t much like the computers we use every day. We need to remind ourselves of the fundamental features of computers so that we can consider more biologically realistic alternatives to the businesslike architectures that are featured in our stereotypes.

This computational picture is a thinking tool for asking how meanings might reside in brains and other machines, and how evolution could produce such architectures without a Master Programmer or Intelligent Designer. The redhead example below explores what meaningful internal machinery might look like without storing a sentence that states its meaning.

---

1 Actually, Descartes’s great imagination got him at least close to some very good ideas. He imagined that the tugging wires might open tiny gates or pores, releasing floods of pent-up “animal spirits” (cerebrospinal fluid) that could then do some kind of hydraulic work; not a bad sketch of an amplifier! Even better, he saw that at least some intelligent (apt, appropriate) work could be done entirely mechanically by such a device: an automatic reflex. Pulling your foot from a hot fire when the heat tugged on a wire was his example: all just brain-machinery, no mind required!

V. MORE TOOLS ABOUT MEANING

28. A THING ABOUT REDHEADS

One tempting idea is that all the information we carry around in our brains—our beliefs, perceptions, memories, policies, and so on—is broken into sentence-like pieces that are filed away and ready to be awakened. Brain-writing cannot just install a false belief, and people can share a belief (about that murder in London, for instance) without—apparently—having to share a formula in Brainish. But what else could store information in the brain? We human beings can learn things “piecemeal,” so there must be some way of adding independent facts roughly one at a time.

Economists (and others) often like to point out that you cannot do just one thing. Doing “one thing” always has consequences. Similarly, the idea that you can learn just one thing is dubious. But to a first approximation you can. Consider a fact you may just have learned: there is a mammal called a pudu. Unless you paused to check on this, you probably can’t affirm anything else about pudus, aside from the fact that they nurse their young, have backbones, and are relatively rare (since otherwise you surely would have heard of them). It’s no mystery how you learned this: you read a sentence and believed it. But can animals or small children who do not yet have a language learn a single fact (like a fact expressed in a simple sentence) from some interesting bit of experience? The idea that knowledge or belief or learning has to be decomposable into sentence-sized gobbets is probably an illusion of anthropomorphism. We human beings encounter lots of declarative sentences, spoken and written, in the course of a day, and we thereby come to know all manner of facts (and believe a smattering of falsehoods). Some we store in libraries and archives, and some we store only in our brains. We seldom memorize the actual sentences, word for word, but when we salt away the gist of an encountered sentence, it must be—mustn’t it?—a matter of storing a sentence-like something-or-other, a formula in Brainish. If this isn’t so, what are the alternatives?

Suppose Pat says that Mike “has a thing about redheads.” What Pat means, roughly, is that Mike has a stereotype of a redhead that is rather derogatory and influences Mike’s expectations about, and interactions with, redheads. It’s not just that Mike is prejudiced against redheads, but that he has a rather idiosyncratic and particular thing about redheads. And Pat might be right—more right than he knew! It could turn out that Mike does have a thing, not an idea or a thought or a belief or an image or any of the other traditional things that furnish our conscious experience, but a bit of sub-personal cognitive machinery in his brain, that is about redheads in the sense that it systematically comes into play whenever the topic is redheads or a redhead, and adjusts various parameters of Mike’s cognitive machinery, making flattering hypotheses about redheads less likely to be entertained, or confirmed, making relatively aggressive behavior vis-à-vis redheads closer to implementation than otherwise it would be, and so on. Such a thing about redheads could be very complex in its operation or quite simple. The contribution of Mike’s thing about redheads could be perfectly determinate and also undeniably meaningful, and yet no expression of this meaning as a sentence-believed-true could be more than a mnemonic label for its role. That is, it might be impossible to characterize the role of this thing as a belief, weirdly specific or weirdly vague, that all redheads are F . . . (where we cash out “F” with whatever content seems to do the most justice to Mike’s attitude). Mike certainly has an attitude about redheads, but it isn’t any particular propositional attitude, to use the philosophical jargon. In other words, it defies categorization in the format

Mike believes that: for all x, if x is a redhead then . . . ,

no matter how deviously we pile on the exclusion clauses, qualifiers, probability operations, and other explicit adjusters of content. Philosophers (and other theorists) have often tried to “reduce” all cognitive states to information-bearing states—call them beliefs and desires—that can be expressed in such formulas, and while the tactic is a great way of providing a rough sketch of some person’s psychology (it’s the intentional stance, in effect), the idea of making it ultra-precise is hopeless. We can say, if we like, that various beliefs are implicit in the system. What this means is that the system is (currently) designed to operate “under the assumption” that the world’s redheads have such-and-such features. When computer programmers put a comment on the source code, telling everyone that this system relies on a defined set of assumptions, they know enough not to devote labor trying to render the propositions precise, because they appreciate that these are mnemonic labels for us observers, not anything the computer has to sorta read and sorta understand, and even for us observers the comments are not specifications of content that can be used the way a chemist uses formulae to describe molecules. Giving an intentional-stance interpretation of some sub-personal brain structure is like putting a comment on a few lines of code; when done well, it provides an illuminating label, not a translation into English or other natural language of some formula in Brainish that the brain is using in its information-processing. By missing this trick, some philosophers have created fantasy worlds of internal sentence-manipulating machinery, where it is imagined to make all the difference whether the content of a particular brain event is expressed using a disjunctive predicate (I see a boy-OR-girl) or with a predicate lacking logical structure (I see a child), for instance.

What is the aim of this intuition pump? It is simply an attempt to suggest that the familiar refrain in support of the language of thought—“What else could it be?”— might have a good answer that would take the wind out of the sails of those who find it obvious. I wish I could provide an ambitious, alternative computational architecture that triumphantly displayed a working alternative, but I can’t. At the time of this account, nobody could, but then almost nobody was trying, since the conviction is still widespread that the language of thought is “the only straw floating,” as somebody put it many years ago. My assessment at the time was that nobody in cognitive science had developed a working model of the language of thought either, or had even tried very hard. It’s a very, very difficult problem.1 I would like to encourage an open mind on this score.

---

1 An aside for the experts: At the time of this discussion, CYC (Lenat and Guha, 1990) was certainly the most impressive artificial intelligence (AI) implementation of something like a language of thought, an enCYClopedic, largely hand-coded database with a proprietary inference engine presiding over it. CYC had been under development for over a quarter of a century at the time, the work of many hands, but it achieves its power by being ruthlessly nonbiological and nonpsychological in its design. (See the excellent Wikipedia entry on CYC.) Mike’s thing about redheads almost certainly is not an axiomatized redhead-microtheory composed in Brainish and incorporated into a large database rather like CYC. We do not yet know how much can be done by a host of things about things of this ilk because we have not yet studied them directly, except in very simple models (such as the insectoid subsumption architectures of Rodney Brooks and his colleagues—see Brooks, 1987). One of the chief theoretical interests of Brooks’s Cog project (a humanoid robot—see Dennett, 1994b) was that it pushed these profoundly nonpropositional models of contentful structures into territory that is recognizable as human psychology.