It's a lucky cause that can hitch a ride with the computer industry. This one can: the industry can scarcely budge without advancing the Day of the Robot. Artificial intelligence and things robotic are only at the edge of the industry's vision. The things that are front and center, however, lead it down much the required road, and it is fueled by buckets of gold.
If the industry did no more than its most obvious thing -- making bigger and better computers --, that would be a powerful boost. A common complaint of programs that try to understand English, recognize the objects in a scene, or put two and two together (in the figurative sense) is that they run out of time and space when they've barely gotten started.
Take the case of Ross Quillian's "semantic memory" program. Quillian has a scheme for "teaching" a computer the "meanings" of words. At least, he "teaches" it the connections among words, and how they relate; when so programmed, the computer plays a good verbal game of making analogies, explaining new phrases, etc.
Alas, though, all that information about connections and relations -- which are the essence of the scheme -- takes up storage at an explosive rate. Let me cite the fine print from a 1966 report.(1) As of that time, Quillian and his associates had painstakingly encoded "almost all of the 850 words of Basic English." (That's not many words. Try understanding a foreigner when you have learned just that much of his vocabulary.) But, says the report, "Space limitations have so far required that definitions of no more than twenty of these words be used to constitute a model memory during a given series of word comparisons." In other words, the computer could imitate a human being who knew twenty words! (Even so, the program was impressive.)
Well, these barriers will fall. Nobody who has worked with computers over the last ten years would doubt it. It is hard to give anyone else an adequate idea of the pace which has been kept -- but let me try. There is no single way to measure the capacity of a computer; it's not a single-purpose machine. Its effectiveness depends on what you're trying to do with it, and how fast you need it done. (If you could afford to wait a trillion years, a given capacity for storing information might suffice for a problem otherwise considered impracticable.) Still, one can look at different aspects of speed and storage capacity.
Only ten years ago, a programmer was lucky to get his hands on, say, a large IBM 650. (At least, I felt lucky. "Rich kids" who worked for the government might have done better, but the 650 was respectable for its time.) If today the same programmer works with a large System 360 Model 91, then he is getting
or some mixture of the above, and
That only begins to tell the story, for it deals only with the "central memory" of the system. The computer can also store millions upon millions of characters in places where it takes longer to reach them and work with them. Some of what we now consider "external storage" devices are as fast as what we used to consider the main memory. All this in ten years, or twelve, if you consider the IBM 650 to have been slightly dated in 1960.
The pace is not slackening. Every computer manufacturer and every school of electrical engineering has a horde of people working on new components, on new representations of information, and on new organizations of the hardware.
Andrew Marvel wrote to his coy mistress that, if only he had time enough,
An hundred years should go to praise
Thine eyes, and on thy forehead gaze;
Two hundred to adore each breast;
But thirty thousand to the rest;
An age at least to every part
And the last age should show your heart.
It is the measure of something or other that where the computer is concerned, whole tribes of men do devote a man-age to every part. However, they do their devotions as a team, and the man-age is over in a matter of months. New computer "generations" are now coming along at intervals of about eight to ten years.
Besides advances in sheer size and speed, we may expect long strides in parallel processing -- that is, the harnessing of several computers (or processing units) to work simultaneously on the same problem. This has long been of keen interest to artificial intelligence types, because it looks useful for both pattern recognition and associative processes.
Consider the human eye: images fall on the retinal cells at the back of it and are reconstituted from so many perceived spots into the patterns that register with your consciousness. How is it done? Sorry, that's nature's little secret. Somehow or other the jungle of nerves behind the retina make a little computation, and presto! You become aware of a pattern.
Whatever the nerve cells are doing, though, it's a reasonable guess that they do it as a massed chorus -- not one after the other. Pattern recognition researchers have long felt frustrated by the need, in present computers, to examine one spot after the other. It takes forever, and probably focuses attention on all the wrong problems.
Similarly, consider what happens when you associate words with their meanings. If we are in the Yankee stadium, the batter swings and misses, and you hear me mumble something about a "second strike", armaments negotiations won't even cross your mind. Neither will baseball flash through your mind if we are at a noisy dinner party, the people in our corner are discussing ABMs, and all you catch of what I say is "second strike."
The fact that the wrong interpretation never even crosses your mind is strongly suggestive of parallel processing. It would be more in the style of a sequential computer to go through the possibilities one by one, rejecting and discarding wrong ones which happen to come first in its stored list.
In both pattern recognition and associative processes, the very fact of simultaneity may be a critical part of nature's trick, especially if the nervous system uses a time-dependent code. (Such things can in principle be simulated by serial computations, but they then assume a form that makes them hard to see through. Moreover, it becomes uneconomical to fish around for the key relationships.)
Again, relief is in sight: there is currently a great burst of interest in parallel processing. Artificial intelligence, which stands to benefit, is not the chief motivation; people are interested because parallel processing looks like one technique among others to get the next big increment in speed and size.
Parallel processing in turn will get a boost from the construction of computer networks. Lately, everyone is building a computer network. Parallel processing isn't what's on their mind (let alone artificial intelligence). They're hooking the computers together so that when one is busy, it can delegate some of the work to a neighbor; or so that several installations can share an uncommon piece of equipment that none of them could justify alone.
That's what people have in mind -- but an immediate by-product will be the possibility of treating the whole network as one big machine, and experimenting with parallel computation as never before.
On quite another front, we may expect beneficial side-effects from certain trends in the way computers are used. Two hot items of particular interest are time-sharing and management information systems.
In time-sharing, a bunch of users sit at terminals (which you can think of as typewriters hooked to the computer by phone), and enter requests or information line by line. The computer services a line of input from one user, and then, instead of sitting back and filing its fingernails while he gets around to the next line -- as it generally could --, it services somebody else's next line. When the stars are favorable, the computer gets back to each user as fast as he can finish the next line, so that it feels to each user as if he had the machine all to himself.
In management information systems, managers at various levels can ask the computer for the latest facts and figures on matters that concern them. The computer has a lot of information at its fingertips, just by virtue of the fact that it processes records on everything that moves -- payroll, inventory, production, customers' bills, purchases from suppliers, and all that makes up a day in the life of a business executive. Moreover, it is capable of processing masses of raw data into summary forms that a manager can digest in reasonable time.
For our purposes, the interesting point is that both types of system put an enormous premium on developing a convenient "man-machine interface." That is, the market for this sort of thing will never amount to much, a long as it takes a professional programmer to play.
The older, batch processing computer systems encouraged the user to leave his problem with a programmer. At most, he might take the trouble to learn a special-purpose language like FORTRAN or RPG, spell out his desires on coding forms, and let the keypunching staff take it from there.
In time-sharing, the ultimate user is encouraged to sit around and communicate directly with the computer. Half the sales appeal lies in the fact that the computer can react a little more "intelligently." It can prompt him what comes next, it can point out certain types of errors as they occur, it can help with erasures and revisions, it can try a program part-way.
Similarly, in management information systems, the ultimate user is encouraged to key his questions directly into the computer. If he has to hire a specialist to phrase the questions, the information system isn't buying him much.
In short, the heat is on to make dialog with computers easy for amateurs. Computer manufacturers, software houses, and time-sharing agencies are responding in a variety of ways, all of which are healthy for artificial intelligence.
One approach is to make the computer understand plain English. That, for the time being, is too ambitious. However, it can be made responsive to a limited vocabulary that is likely to come up in a particular line of work. It can even be made to live with loose grammar, incomplete sentences, and other idiosyncrasies of ordinary speech.(2)
Artificial languages, of course, are nothing new to the computer world; manufacturers have been telling us for years that COBOL is "English-like" (provided you're the kind of English speaker who sees objects lined up to the right and calls them "synchronized right"). What will be new about languages coming down the road (especially inquiry languages) is that they will require less precision. They will rely more on context, and less on grammar, to establish the meaning. This, after all, is how a human being operates. If scramble sentence this I, catch still you'll on.
Another approach is to arrange the data filed in a computer system for versatile types of access. Traditionally, each file has been specially organized for one purpose. The payroll file, for example, is organized to answer efficiently a handful of questions: how much should we pay each employee this week, what deductions must be made, what are his year-to-date taxes and gross income? If somebody should come at the file with a request for average overtime hours of pipefitters, the whole thing would have to be scanned to collect the relevant cases.
The lure of management information systems is that all that information sitting around in computer files can be tapped for management decision-making. The information is only as good, however, as the computer's ability to collect, rearrange, and summarize it on the spur of the moment, in a relevant way. To offer much of a service, the system must not only take in the meaning of all the most likely questions; what's more, it must tackle head-on the problem of "associative memory." Just like any program that ever aspired to artificial intelligence, it desperately needs a method of organizing information for quick recall in the light of whatever question may be asked. The "artificial intelligentsia" have been knocking their heads against this problem for years, and will be deeply grateful for the insights that are developed.
Yet another approach is to substitute handier means of communication in place of the typewriter. TV displays, light pens which can "write" back into the TV screen, and "scratch pad" devices such as the RAND Tablet are already operational. (Usage is not yet widespread because the cost is still high; presumably, improvements in technology and manufacturing techniques will overcome this handicap.) Other media, such as film-scanning devices, are available, but rather expensive and complicated to make use of.
Because the payoff is near-term and the usefulness is obvious, there is intensive effort in graphics devices and image-processing. The science of pattern recognition will get a free ride.
A more remote possibility is voice recognition by computers. There is work in this area, though the experts seem to agree that success will take a while. The potential usefulness, at least, is clear. Particularly in information systems, the prospective customer is a business executive. Everybody knows that the higher you look in the hierarchy of executives, the less literate a class of people you are dealing with. Typically, a business executive makes his living by listening and by flapping his jaw; the written word is not his thing. If he does have to resort to the written word, he shows his discomfort by using a Dictaphone.(3)
To say that all these capabilities we've been discussing are strongly motivated is of course no guarantee that they will be accomplished. To say that they are getting the money treatment is not to say that money can buy just anything. (It is indicative, however, of opinions held by people who have been shrewd in the past.) Still, nobody who is in touch with the field, and who has a feel for the pace and direction of things, can have much doubt: in the 1970's, many problems that used to be "artificial intelligence" concerns will be tackled and solved as mundane business ventures.
|
|
A Word In Edgewise #6, February, 1971: |
| Smarts and the Computer: Can Programming Shade Into Intelligence? |
 |
 |