sign in sign up
<<
Home , John Backus, Oliver Selfridge

—

OCTOBER 1996 IEEE 1E3(1:PIE:FIT INTELLIGENT SYSTEMS THEIIR AIPIPILICATICIMS

4

t 11

411

*

1.11

• :* • • ' , •• • • ...a . • 41 me gr

I.

y g7e • ft • •

,. .,,,. • et' "' r: 4 , . ir,,,,,,• • . * i • - -...... " -.". -.. • I.:. -i' , • - -.et.-.„ • ,Aiff,Lr A ...... ' - • . ..445. • • i--.- Ilig - . • .4$* ' 4$ :04 - '.!. ,,_ -'•I ,' . , _ , ,. r ''.." ••1 Illi .- - ...... * . .4-- 1, lif - ...K.•.- -i; .•f . • ., • - .: - ... - .,.:,-- .-.. .4." '• .1. ....•. . / 4').1° :- ..-•.* * - .... • *Oat r• . •., :.P. r -•: -' t .- .,‘41. • -.4 ‘. v i.... • •• I% .•.' . ---/ - I ii- :•:....:It...2:4:2:bits.t...... _:,,...., 41::;, /( d - #-

40, , - .- . .4. e . I rr- r r»• r- .71 ' • •11 , - y • orl • -• • '. • r orp 4; -

" •

• 4 e' ••• • • • . FM TRIP A conversation with Al pignem Oliver Selfridge Using a knowledge-based architecture for a design automation application Facing Web content EE distribution challenges CO14PUTER SOCIETY 50yEARsoFsERvi( E • 1946-1996 0 What makes a compelling 4J10 THE INSTITUTE OF ELECTRICAL AND ELECTRONICS ENGINEERS, INC. empirical evaluation? powers of simple processing units but also Oliver Selfridge— in the physiology of computing in the brain. in from the start What was the next stagefbr you? In 1951, I transferred from Fort Mon- Peter Selfridge mouth in New Jersey to MIT Lincoln Labs [email protected] (Fort Monmouth had knuckled under to Senator Joe McCarthy, but Lincoln Labs had not, and therefore I moved). By another Driven by his curiosity about the nature I was studying mathematics under Norbert stroke of luck, I bumped into Marvin Min- of learning, Oliver Selfridge has spent over Wiener. By luck, of which I've had a great sky, who was fresh from a Doctorate at a half century enmeshed in the most excit- deal in my life, I was introduced to Walter Princeton and was a Fellow at Harvard. He ing developments in , Pitts, who was working with Warren was a very bright mathematician and communications, and . A McCulloch on a topic they called theoreti- thinker in general—his thesis was on a par- participant at the original conference at cal neurophysiology. I had studied logic, ticular model of, again, neural nets. Marvin Dartmouth in 1956(and at the Western and through Walter, Warren, and Norbert worked for me one summer and then got a Joint Computer Conference in Los Angeles got introduced to neural nets at that time. I job in the mathematics department at MIT, the year before, which he considers the true went to the Pacific at the end of World War working next to RLE,the Research Labora- start of Al), Selfridge formed working rela- II with the US Navy and came back to tory for Electronics, where we met often tionships and cemented friendships with graduate school, again at MIT. Norbert was with Warren, Walter, and the rest of the AI's founding members—John McCarthy, then writing Cybernetics,1 and Walter and I early community.(MIT's Radiation Labora- , and , among were helping him with various aspects of it. tory played a crucial role during World War others—as he went on to become a true AI As I studied mathematics(my original II. It invented and built all of the really pow- pioneer himself. field) and interacted with Norbert, Warren, erful radar systems in the latter part of the Before retiring in 1993 after 10 years as and Walter, I began to be interested in the war. It evolved into RLE, which is celebrat- Chief Scientist at GTE Laboratories, Com- specific processing that neural nets could ing its 50th anniversary this fall.) puter and Information Systems Lab, he do and even more interested in the general At that point, we were using the term served as a member of the National Secu- properties of learning. artificial intelligence and thinking about rity Agency's Advisory Board for 20 years, At this point McCulloch and Pitts had many of the general ideas about intelli- chairing its Data Processing Panel for the written the first two Al papers (although it gence that are still around today—the last 15 of those. He also served on various wasn't called that).2 The first showed that a tough cognitive processing that people did, advisory panels to the White House, as neural net could work out certain kinds of perception, games like chess. We were try- well as on the peer review committee for problems, such as pattern recognition in the ing to tie in the modeling that could be the National Institute of Health (NIH), general cognitive sense, and the second dis- done with neural nets to these more obvi- directed Project MAC and the Cambridge cussed acquisition of patterns(how we ous intelligent behaviors of people. My Project at MIT's Lincoln Labs, and was know "universals"). These two works fol- stress was then, as it still is, to understand Staff Scientist at Bolt, Beranek, and lowed all the glorious mathematics that how learning takes place as the primary Neuman (BBN). He continues to write and Turing and &Mel had done in the twenties source of intelligence in people. speak on and Al, and and thirties about computability and Turing especially on self-improving systems. machines. This mathematics was, of course, This was before computers were generally This profound, long-term familiarity with the beginning of a formal description of available—true? Was it before Von Neu- both the philosophical underpinnings of Al what computability meant. Johnny Von mann had invented the programmable com- and the practical application of AI technol- Neumann visited us at MIT occasionally, so puter? Ifso, how did you work—on paper? ogy places him in a most advantageous again by pure luck, before the age of twenty, We were essentially doing mathematical position to comment on the growth of the I had been introduced to McCulloch, Pitts, and algorithmic modeling on paper. This field in this 40th anniversary year of the Wiener, and Von Neumann. was in 1953, so we knew all about Von Dartmouth Conference. Peter Selfridge, Neumann's work—he had visited us and Oliver's son and an Al luminary in his own Your original background is mathematics, understood what we were up to. And, while right(AT&T Laboratories—Research, pre- but many ofthese ideas begin to verge into we weren't writing Al programs yet, we had viously part of AT&T Bell Laboratories, biology. Was this accidental? had some experience with early computers. member ofIEEE Expert's Editorial Board), I went through MIT, and so I got a very I had done some programming on World- recently asked his father to assess Al's fair liberal arts training, including things wind, a very old computer put together by progress since its earliest days. such as biology. One of my roommates was Jerry Wiesner at MIT to get money out of Jerry Lettvin, a super neurophysiologist, the Air Force (which it did superbly well). Peter Selfridge: How did you become inter- and McCulloch had also been a neurophysi- Lincoln Labs was a follow-on to that. ested in Al? ologist of some renown.3 When one looks at So, while Marvin and I were having Oliver Selfridge: It was at MIT, a long the early work on neural nets, you find these interesting thoughts, computer tech- time before the Dartmouth Conference, and yourself interested not just in the computing nology was exploding in the basement of

OCTOBER 1996 15 interested—Wes Clark and and so yes, we were in our late twenties. Belmont Farley at Lincoln There was a great deal of excitement. Labs come to mind. Bel- mont Farley wanted to be a Can you now talk about the Dartmouth wet neurophysiologist, and Conference, which people think ofas being so he was interested in the start ofAl? neural nets, too. By luck, at The Dartmouth Conference was funded one point I made a trip out by the Rockefeller Foundation, as I men- to the Rand Corporation— tioned. It started, I think, on August 6„ 1956, this was 1954—in Santa and lasted roughly four weeks. In that Monica, California. Rand sense, it wasn't a conference like today—it was smaller then; I was went on much longer, and was much more talking to Willis Ware and a loosely structured. Many people were young, smart psychologist invited and dropped by. I was not there the from Carnegie—it wasn't whole time. We would meet and give talks Carnegie Mellon then, it and argue, all those wonderful things. Many was Carnegie Institute of came who shared our goals and excitement: Technology—named Allen Roland Silver, Art Samuel, Leon Harmon. Newell. I talked about pat- There were not many people doing research tern recognition. And Allen in computing, but there was lots of interest and I hit it off, and this was among those who were. , who really the trigger. I mean he invented , turned up there. Oliver Selfridge, 1996. was ready for it. He got the And it was a John McCarthy show, so to whole idea, he got turned speak. At this stage, he was already begin- on, and he went back to ning to think about symbol processing. Building C at Lincoln Labs where they had Carnegie and turned Herb Simon on. Those Alan Newell really pushed us all into think- the Memory Test computer, all vacuum two went on to become, well, Newell and ing symbol processing rather than bit pro- tubes, and then the Transistor Test com- Simon of CMU,extraordinarily influential cessing—well, not quite rather than, but as puter(TXO). Jay Forrester was at MIT, and thinkers over the years. well as bit processing. Al needed bit pro- he came out and perfected the donut- Marvin and I were talking and writing cessing for, say, sensory integration for shaped core memories, and I remember about these things in general. What did we example, but if you wanted to do higher seeing the first million-bit memory. Now a think the problems were?—visual pattern levels, you wanted to look at what it took to million-bit memory is 64K words or 128K recognition, making speech, understanding process symbols. bytes, which,if you had it on your finger- speech. Later,planning played a bigger nail now, you would hardly be able to see role, but not then very much. I still had Was John McCarthy on the West Coast at it. This million-bit memory occupied, with learning very much in mind then, as I still this point? its frame, about one cubic yard. The power have. Marvin and I essentially organized No, John was at MIT with Marvin. I supply, built with vacuum tubes, took what to me is more the beginning of Al don't know John's years at MIT, but they another cubic yard. Marvin and I specu- than the Dartmouth Conference, which was must be something like 1954 to '58 or '59. lated about whether commercial memory the Western Joint Computer Conference, After the conference, John was striding might ever be as cheap as a dollar a bit! held in Los Angeles in 1955, the year around his and Marvin's office inventing This was only 40 years ago. before the Dartmouth conference. We got LISP."CAR, CDR, CDR, CAR" echoing The programming was usually done on funded by, I think, the ACM and the Insti- around. LISP, of course, became the pre- paper tape in machine code, essentially. And tute of Engineers, which I believe ferred Al for re- at about this time,IBM came out with the turned into the IEEE. searchers, probably to this day. 650. The 650 didn't read a tape, but instead, you programmed the machine with wires Was everybody there? So was there the same sense you had the and plugs, much like the old switchboards. There weren't many "everybodys," but year before, that this was the beginning of Marvin, I, Belmont Farley and Wes Clark, something very exciting? Back to the early Al community. Who was and Allen Newell gave the four papers. It It was tremendously exciting, both from involved at that time? was a great thrill, realizing that this was a the point of view of computation and from There were numbers of us thinking quite first, that this was going to be something— the point of view of beginning to under- hard about Al. John McCarthy had met which it is. stand what a mind is. Marvin at MIT. John was the one who got How does a mind do its computation? the funding for the Dartmouth Conference You were all in your late twenties? This is a tremendously antique question. from John Morrison who worked at the Marvin is a couple years younger than I By antique, I mean that it goes back 5,000 Rockefeller Foundation. There were num- (I was sure that he would catch up, but he years at least, when people were looking at bers of other people also beginning to be never has). Allen was about Marvin's age, the clockwork of the heavens. Clockwork

16 IEEE EXPERT is literally a copy of the rotation of the tion (ideas such as planning ahead for CPU This was the first true hands-on, widespread heavens, which is why up to a few decades processing or disk access); those were hands-on equipment. We were also lucky ago clocks all had hands that went derived directly from interaction with Al because a bunch of people from Lincoln around—a model of the Big Dipper going and the Al techniques. Labs, where I was working (many of them around the pole star. You see lots of discus- my friends) went off and started Digital sions early on about how to make a robot or Was the Turing Test discussed? Equipment Corporation. DEC was building a man model out of clockwork. Oh, it was indeed—thought of and dis- computers that were essentially run by But here we were, beginning to get at cussed, as was the Turing machine. This users in those days. Before that, in the these issues in computer software, where was, after all, the breakthrough model of fifties, you ran a program by writing punch we had a chance of being able to find out computation. It became then, and still is cards, which you then submitted to the desk the basics, to see what was do-able, and to probably, a very powerful educational of the computation center. The cards would get at the secrets about the nature of man. device—then and now, in advanced compu- be scanned and often sent them back with tation classes, homework assignments an error message such as "it doesn't parse" Did you all have the feeling that you were would often be to program a Turing or it's no good for some other reason. If you part ofa grand tradition, spanning machine. Back then it was recognized that were lucky, the deck ran after only a few millennia? there were interesting questions about which iterations and you got your results back, but Marvin, John, Allen, and I, and many of computational models were equivalent to perhaps days or weeks later. the others, had considerable backgrounds, Turing machines. At the time of the Dart- Well, when Digital built the PDP-1, you and we had studied philosophy. Not only mouth conference, there were certain mathe- sat at the console and you wrote your pro- that, but we also knew McCulloch, who matical games called Post tag systems. gram yourself and you could try it. This was a philosopher of some depth. Warren was a startling revelation. In fact, the first was far more on the emotional side of the time-sharing system was done on the PDP- folks doing AT, and Norbert of Cybernetics 1. McCarthy at that point was spending the was an emotional baby who had the math Because we understand summer at Bolt, Beranek, and Neuman down cold. Wonderful, contrasting intel- (BBN), with its Number 1 and Number 2 lects to work with. Also at MIT then was them so well, people don't PDP-ls. BBN was an acoustical research Giorgio di Santillana, a Historian of Sci- think of today's chess center, but Licklider was there, .C.R. ence, an exciting and visionary philoso- programs as Al. Back then, Licklider out of the Harvard Acoustic Lab. pher, a special hero of mine. It was very the kind of look-ahead Marvin and I both knew him very well. I strange, as I said, that I had pure luck all procedures and evaluations had met him in the forties. He had been my life with these things. But the convic- working on acoustics, but he was also won- tion that this was very important, I still we used were definitely derfully interested in computation, as a hold today—it is! We are facing the same part of Al. way to go by itself and for Al. He went questions that have been around for 5,000 down to Washington in the early sixties to years, questions about the nature of Marvin showed then that these systems run the Information Processing Technology mankind. What is man, that thou art mind- could be interpreted as Turing machines. Office, one of the major offices of ARPA, ful of him?(Psalm 8, verse 4) That is, a tag system can compute anything a the Advanced Research Project Agency, for computer can. That's not practically useful, the Air Force. And Licklider gave some Back to the Dartmouth Conference: what but it gives a deep sort of foundation to money to his old firm BBN to do some- visions were discussed? Robots? understanding the basis of computation. thing interesting, like time sharing. He We were really thinking about robots, played an absolutely critical role in funding talking about robots. Most of the problems Was this part ofthe beginning ofcomplex- Al work in those days, and he has been were much harder than we thought. I re- ity theory? largely forgotten. member a year or two after the Dartmouth Yes, but it wasn't thought of that way. Conference telling quite a good chess Complexity theory itself took another 12 or When was time sharing first tried? player (I believe it was Hugh Alexander, 15 years before it had its name as such. McCarthy hacked time sharing at BBN one-time chess champion of England) that in 1961 or 1962 on a PDP-1 with a total of my machine would beat him in five years. What was the next step, now that comput- 16K of memory, using not the old model 33 Well, it took 35 years, but now the world ers were beginning to be available? Teletype, which came later, but the Tele- chess championship is clearly within reach. There were lots of things discussed, yet type that used a paper tape punch—click, Because we understand them so well, peo- we were handicapped by the inadequacies clack, clack, clack. John had three users ple don't think of today's chess programs of computers. Some of the early stuff on simultaneously—they weren't able to do as AI. However, back then, the kind of neural nets worked appallingly slowly. Mar- much, but to do something. look-ahead procedures and evaluations vin and I recognized that computational This work continued at Project MAC at used were definitely part of Al. Ordinary capabilities had to be improved. Seven MIT, which was set up again by Licklider, computation didn't use those notions about years after the Dartmouth Conference, who gave ARPA money to MIT in the early planning, look-ahead, and tree searches. again owing to a piece of wonderful luck, I days. I should also pay credit to Professor There wasn't anything like it in computa- was part of Project MAC,started at MIT. Interview continues on p. 84

OCTOBER 1996 17 Chances are you missed some of these.

COMMIT

••••,*•• •• NALYSIS AND INTELLIGENCE

aTIORS *OM 999999 VISUALIZATION A COMPUTES GRAF

-••• ••— - k

Now s your chance to catch up.

ler IL V Read the articles you missed. Find publishedirescriptions of new products. Search for references or articles relevant to your research or current projects. Computer Society '95 is a fully searchable CD-ROM containing all 16 magazines and transactions published by your society in 1995. ecial Member Price: $89.95. Catak #SW00049.

Order from the 50 YFA RS OF SF 12y1CF IEEE Computer Society Online Catalog COMPUTER New! on our secure web site: SOCIETY0 WWW.computer.org/cspress or call 1-800-CS-BOOKS