While engineers and programmers labor to infringe the patent, what are other scientists learning about nature's model of a brain? Many workers in the field of artificial intelligence have been unconcerned how closely their handiwork resembled the real McCoy, as long as it worked. For the dream I am offering, however, this is a crucial matter.
Not only is the human mind the only working model we've got of the thing we're trying to imitate; beyond that, the facts about it must cooperate, or the larger objective will be impossible. There will be no artificial immortality if it turns out that mind is untransplantable, inseparable from the squishy stuff inside a cranium. Neither is there much hope if all the minds we can build work too differently from the kind nature supplies. Whatever process we may imagine "reading out" the one into the other, it clearly requires that there exist some systematic mapping between them.
Brain research is not the dynamo that computer development is. There are tough obstacles. For every little thing that investigators learn about a brain, they have to cut up a bunch of dogs and cats. The parts are unbelievably small and tangled, and if you're looking at one, the odds are that it's no longer functioning. And then there's the institutional thing: the profit motive isn't enlisted. Medical research limps along on charitable contributions and handouts from the government.
For all that, there is motion. What is interesting is that much of the motion just now is inspired by insights from information science, son of digital computer. This is true not only in brain research proper, but in all the fields surrounding it.
In mathematical biology, for example, the two big thrills are automata theory and control systems theory. Automata theory (as adapted by biologists) makes up abstract models of "neural nets," suggested by what biologists know of actual nervous systems. Mathematical techniques are then applied to deduce what nervous systems could do, if they were indeed like the models. The accent here is not on theorizing what is machine-computable, but on testing the reasonableness of the models. (If a model has shortcomings that animal brains don't share, then the model needs revision -- and the nature of the shortcomings may suggest the revisions to make.)
Computer insiders may be inclined to put this down, knowing from experience that automata theory is the perfect escape from productive labor. But remember that biologists have precious little opportunity to program or engineer the objects of their study. "Hands on experience" means a blood bath. Taking canine and feline sensitivities into account, it's just as well that biologists check their models indirectly, where possible. The hope is that automata theory, properly cooked, can serve as a sort of analytical simulation.
The other hot item in mathematical biology is control systems theory. This deals with the massive systems of differential equations that turn up when feedback of information (usually in the form of electrical or physical quantities) is used to control processes in the real world. The related mathematics are in a state of runaway progress -- partly because this is an easy direction for progress to take, and partly because the governments of the U.S. and the U.S.S.R. are interested in guiding missiles down each other's moonbeams.
Biologists see control systems wherever they look: subconscious functions like the regulation of breathing and the control of posture, visual tracking mechanisms, and (most important) much of what happens to signals knocking about in a neural net.
The extent to which the computer analogy has grabbed bi-sci imaginations may perhaps be seen best in the 1968-1969 report of the Biological Computer Laboratory at the University of Illinois.(4) This group works full-time on the study of living organisms as information-processing systems. They are into everything from mathematical modeling to the study of motor control in salamanders. If some of their work on pattern recognition, language processing, etc. differs from what goes on in a computer manufacturer's lab, it is only in being more concerned to see whether nature is doing it the same way.
Similar things are afoot in psychology. An "information processing" school of thought, radiating out from Carnegie-Mellon University, has rapidly gained a place alongside such earlier approaches as Freudianism, behaviorism, etc. Besides the usual day's work -- mulling over last night's dreams, cataloging optical illusions, measuring response times, etc. --, a psych student may now program an imitation of a thought process.
A computer program, the thinking goes, is one way to state a theory about a mental process. For certain types of subject, it's a handy language. It beats mathematical notation for describing procedures and filling in details; it can cope, as statistics can't, with the individual case; and it has at least a small tendency to demand clarification of things that ordinary language would gloss over. (This last claim must be taken with a grain of salt; confusion can be expressed in any language.) Also, if the theory is very complicated, the program can be run on a computer, and unexpected consequences brought to light.
Psychologists of this persuasion have been prominent contributors to the lore of artificial intelligence itself, right from Day One. The "General Problem Solver" of Newell, Shaw, and Simon is considered a classic in the field. Generations of grad students at Carnegie Tech cut their eye teeth tinkering with this program, which originally aspired to take on all problems. (Eventually it settled for studying the plight of an information system when stuck with such expectations.)
Another classic is Feigenbaum's "discrimination tree," which attempts to explain certain types of learning and forgetting. Current thesis literature is full of imitations and variations. And then there is Reitman's "ARGUS" program, which explores the issue of centralized versus decentralized control in a thinking mechanism.
Stanley L. Jaki, who has written a book to prove that artificial intelligence is impossible, derides all these developments. Comparisons between the mind and the computer, says he, are a passing intellectual fashion -- and he hopes that it will pass sooner rather than later, because it demotes mind to the level of mechanism. (The argument is essentially definitional: he accepts as "mechanism" only what is mindless.) He points out that every new gizmo from the granddaddy clock to the governor on a steam engine has been touted in its day as a model of the human mind.(5)
Well, to be sure: scientists have to have hypotheses. They can't just let the facts wash over them like psychedelic colors; they have to bring structured guesses to the facts and try them on for size. And of course their guesses are suggested by phenomena they already understand, and especially by what they've recently learned to understand. That's where the action is. The process leads to some bloopers, but what other process is there? It will continue until the mystery is solved.
Meanwhile, let's argue the "information processing" analogy on its own merits. Its point of departure, whether in biology or psychology, is the reasonable observation that a nervous system (brain included) does, after all, process information. That has implications, whether the system acts on its own or is used as an instrument.
We've been playing around for twenty years, now, with some information-processing mechanisms of our own. They may be paltry things in comparison with a mind, but then, precisely because our minds are nimble, the minds deduce from the artifacts some general conclusions about what it takes to process information. We know some things to look for.
First there is the question of data representation. Information comes at the organism in the form of light waves, vibrations in the air, pressures, chemical particles, etc. As we find none of these duplicated in the brain, it's obvious that information has to be encoded. As a matter of fact, it is likely that each scrap of information is encoded at least twice: once for transmission along the nerve cells up to the brain, and then again into a code suitable for storage. (Not just any code will do; at each stage, the code has to be convenient in connection with the physical components that are used.)
That brings up the second question: storage. How is information retained? How is it called out again? Apart from the physical questions implied here (which are mysterious enough), what is the logical organization? How is the information arranged to make related pieces relate? How is it worked up into structures and patterns? Is there a hierarchy of storage devices (temporary vs. permanent, for instance)?
Computer experience suggests that there may be layers and layers of logical organization, each working on the results passed up from the layer below it. In computers -- to simplify a little --, available components, together with certain ways of hooking them up and certain ways of looking at the voltages which result, provide engineers with a repertoire of ANDs, ORs, NOTs (or just "NANDS" and "NORS"), timings, and routings. From these, the engineers design a mechanism with a repertoire of operating codes, data units, storage addresses, number types, etc.
From there the programmers take over. Blissfully unaware of the way a mess of voltage differentials got transformed into the fiction of a logical "NAND", or of how that fiction in turn got transformed into the fiction of a "CLEAR AND ADD", they work with the repertoire produced by the engineers as if this were the machine. That is all the machine they ever see.
The programmers in turn invent higher-level fictions for each other. The FORTRAN programmer, working with mathematical functions and replacement equations, may never hear of "CLEAR AND ADD"; things like that take place below his threshold of visibility.
Then there is the question of an operating system. What program masterminds the operation? How does it choose what to attend to next? What are the programming characteristics of peripheral devices (like legs)? How much concurrency is possible? How are interruptions processed (and how many can be stacked up)? What are the provisions for error recovery? What imposes purpose and coherence on the system? How does it sustain a line of thought? How does it generate new business for itself (assuming our intuition is correct in telling us that it does so) -- or, if it is in reality interrupt-driven, what is the pushdown mechanism that accounts for long-term effects?
None of these questions is exactly new. They are all just re-phrasals of questions that science has been asking about the mind all along. In research, however, the way you put a question has a lot to do with the possibility of answering it. The "information processing" slant on things is a gold mine of suggestions; it narrows the search for plausible theories tremendously. Everything it suggests is based on practical experience -- and there will be plenty more where that came from.
No sensible person would claim to know that this key will unlock the mystery on which so many other keys have failed (or to know the opposite, either). Still, on the face of it, it's a sensible person's bet. This is no "granddaddy clock" analogy -- we are dealing with machines which process information, a function to be explained also in nervous systems. The comparison is plainly to the point.
Before departing this topic, it might be useful to sketch what is already known of the nervous system as an information system. Much of the most solid knowledge, of course, antedates the "information processing" point of view. (Those of us steeped in the parvenu science of computers are somewhat startled, on picking up a review of "the literature" in a field like neurology, to find that it starts off with still-definitive results published, say, in 1850 or 1910.) The "information processing" analogy, however, puts everything in a new light.
Firstly, the physiologists have a relatively filled-in picture of the components and how they work. They can tell you about a nerve cell, and how the sodium ions bang at the door to get in, and the chain of consequences by which this causes a current to ripple along the cell. Since we are not at this point designing improvements, but merely trying to grasp how the mechanism works, this information is not yet central to our purpose. It is rather as if we knew all about the chemical reactions in baking clay, and couldn't figure out how brick walls are made. Still, facts about thresholds, refractory periods, inhibition, etc., do guide our thinking somewhat, and will come in handy when we get to the point of spelling out details.
Meanwhile, anatomists have provided us with a fairly comprehensive wiring diagram. That's no mean achievement; following a chain of nerve cells is like disentangling one strand in a wadded-up ball of spider silk. In fact, it's even worse: functionally, the smallest strand that can be followed is often more like a transatlantic cable than one family's extension. Biologists are obliged to play mean tricks, like damaging one end of a nerve and doing an autopsy later to see where the damage has spread. (The stringy part gets thicker as it decays, and stains differently.)
A wiring diagram is useful information, but if you've ever looked inside a computer, you can appreciate that there's a lot it doesn't tell you -- even if somebody does drop a hint that the blamed thing is processing payroll.
It's the levels of explanation in between that are missing: all that stuff that the engineer learns in his courses on logical design, and that the programmer digs out of the programming manuals. We have barely scratched the surface of such questions as data representation, logical organization, and system integration.
Even where wiring is concerned, the trail typically vanishes just where it gets interesting. By dint of laborious experiments and revealing calamities (various forms of brain damage), neurologists have learned the gross distribution of functions between various "black boxes" comprising the system. They know, for example, that the cerebellum is necessary for motor coordination. But if you look any closer, the parts get too small to follow. You can listen to Sir John Eccles, a leading authority on the cerebellum, and come away with little more than this: that there are several types of cell, there are millions of each, they interlock like the cast in one of those pornographic stage scenes, and each does its own thing.
I was informed by Warren McCulloch of MIT, a couple years ago, that we do know at least the form of the code used in transmitting certain sensory information: a frequency code in one case, variation in the lengths of bursts in another. He said nothing, however, of the code used for storage; apparently that is a vast blank.
A little bit of this, a little bit of that ... Experiments on cats have shown that you can cut away more and more of the higher nervous centers, and quite complicated motor functions will still get performed (if a little bit of juice is applied to appropriate nerve endings). With nothing left but the stem of the brain, the remains of the cat can be made to run on a treadmill. Having lost its sense of balance, it has to be supported in a sling, but the paws step along in good time. That at least confirms the existence of "peripheral units" with sophisticated repertoires of their own. Not that anyone would doubt it who has ever walked down the street while absorbed in a daydream.
Perceptual experiments have established beyond reasonable doubt that preprocessing does take place somewhere between the eyes and whatever it is in our brain that is first "aware" of seeing something. We see patterns, not dots, and can be fooled in any number of ways by misleading diagrams; our preprocessing units make (spurious) sense of them before passing them up to our consciousness. The detailed workings of these preprocessors, however, are utterly foggy.
An expert in any of these fields could extend this summary at book length. The upshot, however, would have to be the same: that there are appalling gaps in our knowledge, but that the hints we do get all encourage the "information processing" analogy. It is precisely to the hazy areas that this analogy brings new questions and guesses. We are going to see a booming business in the investigation of such matters. With just a little bit of luck, we should soon solve some problems that have baffled everyone for thousands of years. And if we make it a matter of deliberate policy, we just might know what we need to know in time to take the advantage I have suggested of robots.
|
|
Pattern Recognizer Meets Theorem Prover: A Robot Gets its Head Together. |
 |