Oral-History:Walter Karplus
About Walter Karplus
Walter Karplus (1927-2001) was born in Vienna, Austria. After becoming a U.S. citizen he served in the U.S. Navy and was educated at Cornell University (B.E.E., 1949), the University of California, Berkeley (M.S., 1951), and the University of California, Los Angeles (Ph.D., 1955). Karplus became a UCLA faculty member in 1955; in the 1970s, he was instrumental in the university’s establishment of its computer science department. His administrative service at UCLA included terms as department chair and interim dean. Professor Karplus was a Life Fellow of the IEEE and served as President of the IEEE Neural Networks Council (1995-1996), promoting this group’s 2001 evolution into the Neural Networks Society.
Conducted as part of a project on the National Science Foundation's support for computer science, this telephone interview explores the roles of NSF funding in Karplus’ research. Karplus received frequent NSF grants between 1965 and 1972; by the 1980s he was relying on industrial rather than NSF funding. In 1961, Karplus began using computers to solve partial differential equations. With NSF sponsorship and industrial collaborations, he applied this approach to model air pollution, water resources, space vehicles, power distribution systems, and nuclear reactors. Partial Differential Equation Language (PDEL), the software package he assembled for equation solving, was used by national laboratories, private companies, and universities. During the later years of his NSF sponsorship, Karplus focused on system identification problems, or the modeling of real-world systems, such as underground water reservoirs, using their physical properties.
In this interview, Karplus details his combination of analog and digital technologies to design efficient networks. Although Karplus ended his analog work in the late 1960s, his analog subroutines concept influenced 1990s models of artificial neural (brain) networks. He considers the NSF’s goal of funding research, as defined against development, and explains the influence of this policy on his own balance of computer science with engineering and aerospace industry collaborations. He describes the NSF application process and explains his personal transition from NSF to industrial funding. At the conclusion of the interview, Karplus summarizes the impact of his research, particularly in the area of hybrid digital and analog computing.
About the Interview
WALTER KARPLUS: An Interview Conducted by Andrew Goldstein, IEEE History Center, 26 July 1991
Interview #114 for the IEEE History Center, The Institute of Electrical and Electronics EngineeringEngineers, Inc.
Copyright Statement
This manuscript is being made available for research purposes only. All literary rights in the manuscript, including the right to publish, are reserved to the IEEE History Center. No part of the manuscript may be quoted for publication without the written permission of the Director of IEEE History Center.
Request for permission to quote for publication should be addressed to the IEEE History Center Oral History Program, IEEE History Center, 445 Hoes Lane, Piscataway, NJ 08854 USA or ieee-history@ieee.org. It should include identification of the specific passages to be quoted, anticipated use of the passages, and identification of the user.
It is recommended that this oral history be cited as follows:
Walter Karplus, an oral history conducted in 1991 by Andrew Goldstein, IEEE History Center, Piscataway, NJ, USA.
Interview
INTERVIEW: Walter Karplus
INTERVIEWER: Andrew Goldstein
DATE: 26 July 1991
PLACE: Telephone Interview
Research and NSF funding summary, 1960s-1970s
Karplus:
Hello.
Goldstein:
Hello. Are you free now?
Karplus:
Yes. This is a good time.
Goldstein:
Let me just remind you of our project. We're working on a history of the National Science Foundation. We want part of that history to contain a discussion of some of the research that was done under NSF support. And looking over the grant lists we saw your name and we thought-- I would like to discuss the work you were doing under the NSF. You received grants regularly from 1965 through 1972.
Karplus:
Yes.
Goldstein:
Can you just give me, beginning with a general description of the work you were doing?
Karplus:
My work in general, and I guess my NSF support, started in 1961. This support continued in a continuous sequence of about twelve years in the general area of using computers for the solution of partial differential equations, in particular the mathematical models that characterize physical systems. Over the years I worked with air pollution, water resources, and a variety of other things. But they all had the common denominator that they were fundamental equations of classical physics. But the solution of these equations holds particular computational difficulties.
Goldstein:
I see. So, it was numerical analysis work?
Karplus:
Yes. It's at the interface of numerical analysis and computer engineering because the problems I chose were always too computationally intensive. They were too time consuming for conventional computers. We were looking for ultimate schemes. Back when I started, I was just coming out of my analog days. I had started off in analog computing and introduced the notion of using analog subroutines in digital computer programs. Thus, at certain points in the computation, the digital computer would assign certain time consuming tasks to a special purpose network that I had built.
Analog and digital computing
Goldstein:
So the electronics were analog electronics?
Karplus:
It was called the DSDT system (for Discrete Space, Discrete Time). It was used in a number of ways. And, just recently people have begun referring to it again.
Goldstein:
And then the results from the analog segment would be digitized and fed back into the computer for later?
Karplus:
That's right. A large network of resistors receiving input in the form of voltage and then currents, relaxed almost instantaneously to what amounts to the solution of the Kirchhoff’s Law equations that govern the system. If the network is correctly set up those equations are analogous, that is they're similar term by term to the digital model that they're trying to solve. So, you have essentially a network that does the job much more rapidly.
Goldstein:
I see. So you would design circuits that simulated the behavior of the equations that you were interested in?
Karplus:
Yes. Although, in fact, the circuits were general purpose in the sense that a large class of equations could be handled by the same network.
Goldstein:
So would there be a different network for each problem? Or did every problem require several different networks?
Karplus:
The problems of physics fall into broad categories. Elliptic, parabolic and hyperbolic equations. For instance, parabolic equations are heat transfer problems. They also characterize diffusion of a pollutant in the atmosphere. Almost every area of physics has some parabolic equations.
Goldstein:
Okay.
Karplus:
Well, anyway, that was the early work. Gradually I evolved into other aspects of modeling and simulating distributed parameter systems characterized by partial differential equations. For a while, I was interested in techniques of solving these problems by pure digital techniques and designed a variety of simulation languages. Those are software packages to facilitate the solution of certain problems, particularly those of interest to engineers, more accessible or user-friendly. Like FORTRAN makes it easier to solve mathematical problems. So using my language you could, in a very compact form, describe the equations in the systems.
Applications; aerospace industry
Goldstein:
Now had you moved on to a different set of problems? Or were these largely the same problems that you had?
Karplus:
When I say largely the same, those are the problems of mathematical physics.
Goldstein:
The solution of parabolic or elliptical equations?
Karplus:
They cover all the application areas.
Goldstein:
So there weren't specific problems that were motivating your research? Because they had general application?
Karplus:
Right. Although I always worked on applications in interesting areas.
Goldstein:
What areas were these? And what suggested them to you?
Karplus:
Well, in Los Angeles we're at the heart of the aerospace industry, and I had close contact with aerospace companies. So I was frequently stimulated by interactions with them. Often, my students were part-time employees of companies, like Hughes, TRW, Lockheed and the like. The same is true, to some extent, of NASA. I worked a long time in the cockpit flight simulator field.
Goldstein:
Did you receive any funding from the corporations whose problems you were working on?
Karplus:
Well, I had several contracts with NASA. The corporations not so much. IBM gave us a lot of hardware. That was important. I don't recall the aerospace companies directly providing money.
NSF funding
Goldstein:
I know the NSF was supplying money. Were you also receiving money from the Department of Defense or the AEC?
Karplus:
No. NASA.
Goldstein:
Just NASA?
Karplus:
Well, there was also the Office of Water Resources Research. And a few others. It was just by coincidence that I happened not to go after DOD sponsorship.
Goldstein:
I see. Was NSF money relatively easy to come by? Did you have a tough time persuading them to sponsor your research?
Karplus:
Well, if you get on the escalator, you do what you have to stay on. I worked hard to get the initial grants. And then I ran with the pack, essentially. So the renewals kept coming. Each would be for two or three years but, each one was a separate, peer-reviewed proposal.
Goldstein:
I see. Perhaps in the early days they specified hardware, whereas in the later days it was your time or the time of research assistants or graduate students? Is that accurate?
Karplus:
It was always a major part.
Goldstein:
What?
Karplus:
Being people minded.
Goldstein:
So even in the beginning when you were working more with hardware that was a minor component?
Karplus:
Well, maybe not minor. But hardware you could always somehow get a good deal from companies.
Software solving partial differential equations, PDEL
Goldstein:
I see. You said that you assembled several software packages for the solution of these partial differential equations. What were the names of the packages? How widely were they distributed?
Karplus:
Let's see. One was called PDEL, Partial Differential Equation Language. We sent it to a few people and then a small software company here took over its distribution. And I know it also went to several national laboratories, some private companies and of course, universities.
Collaborations in mathematics and applications
Goldstein:
Who were some of your collaborators during the years?
Karplus:
Do you mean the names of professors that worked with me?
Goldstein:
Yes. Did you work closely with people who were more mathematically oriented for solution methodologies for these?
Karplus:
From time to time, yes. I worked with people in the math department. But I think my primary collaboration was with people in application areas, like civil engineers interested in underground water reservoirs.
Goldstein:
I see.
Karplus:
And, aerospace engineers.
Goldstein:
I was talking not long ago with David Young, who seems to have been doing related work. He was working on iterative solutions for the partial differential equations that derive from oil recovery and reservoirs.
Karplus:
He's at the University of Texas, right?
Goldstein:
Yes.
Karplus:
He's much more of a mathematician than I am. He's an applied mathematician. I'm really more of an engineer, who uses math.
NSF definitions of research
Goldstein:
I see. As you were working on the software packages, did you feel you were still in engineering? Had you felt that you had been drifting over towards computer science, or software design?
Karplus:
With National Science Foundation sponsorship, almost from the beginning there were problems about "What is research?" and "Is engineering research?" And the NSF Program monitors were always under pressure to justify that what they were doing was really research rather than a sort of development that industry should be doing.
Goldstein:
I see.
Karplus:
So I always try to wear the mantle of the computer sciences, even as I worked in the application areas
Analog approach and artificial neural networks
Goldstein:
You say there's recently been a revival of interest in your approach, in your analog approach. Can you site some examples?
Karplus:
Well, one of the hot areas in computer science is now artificial neural networks, networks that are modeled on the brain. Now these, in reality, are all analog. At least today all the neural networks are analog. So, we're back to having analog networks. So, the concept, that I introduced, of analog subroutines was sort of revived. A couple of people I know referenced my work and then called me up about it.
Goldstein:
All right. Who, for instance?
Karplus:
John Caulfield, in Alabama.
Goldstein:
Right. Back in the '60s, what caused you to start to wind down the analog work and focus more directly on software packages? Did you feel that your simulations were better?
Karplus:
No. I just saw how the wind was blowing. An important part of our function here is to educate students, to send them out in the world with a bag of tricks to help them get jobs, and do well. Companies were looking for people trained in more all-digital simulation rather than in analog or hybrid simulation. So while I never really lost my fascination for analog, I moved on.
Research applications
Goldstein:
You mentioned a few specific applications; you said air pollution was one. Can you name any others? Things you said perhaps for the aerospace industry?
Karplus:
I mentioned flight simulation. It was something that I did for many years. I worked on nuclear reactor simulation a lot. I looked at the thermo-hydraulic transients in the nuclear reactors, meaning the cooling water temperature and pressure.
Goldstein:
Now those results, I would think, would be of great interest to either utilities or the AEC?
Karplus:
Yes.
Goldstein:
When you would work on a system, what would become of it? Would you develop something and sell it to a client?
Karplus:
No. Usually my output was technical papers.
Goldstein:
I see.
Karplus:
I don't recall ever trying to sell something, although I moonlight as a consultant. But that's not really the research that gets sold. It's the expertise that I've learned and acquired during my research.
Goldstein:
Okay.
Karplus:
There's also electrical power networks, power distribution systems. In the Apollo program there was training for the dynamics of space vehicles, and the control of space vehicles. There was an application there. I did mention aquifers, underground water reservoirs?
Goldstein:
Yes.
Karplus:
That was a major application. You know, I could if you like, I could mail you my curriculum vitae, which has over a hundred papers.
Goldstein:
Oh, that might be useful. I'll be expecting it. You may have touched on this just before, but these specific problems, were you aware that they demanded treatment? Or did you consult with someone who brought them to your attention?
Karplus:
Well, I circulated around quite a bit at conferences and in the industry. So one way or another they came to me.
Goldstein:
Were you attracted to a problem because of its urgency, because an idea of how to attack it occurred to you?
Karplus:
Yes. (I hate to sound like I had a solution looking for a problem.) But for the techniques that I developed, there are some other phases that we haven't talked about yet, that lend themselves to the treatment of certain problems and not others. My ears were always open for good applications for the techniques that we were working on.
System identification problems; underground water reservoirs
Goldstein:
When you say other phases we haven't talked about, what do you mean?
Karplus:
Okay. After the simulation languages, for a number of years, I guess, until the end of my support from NSF, I was interested in the system identification problems. Which is, how do you make a model of a system that exists in the real world? In particular how do you derive the parameters? For example, if you have an underground water reservoir, and you would like to predict how the water level will rise or fall over the years, you need to describe the physical properties of the ground, of the geologic strata that make up that reservoir.
Goldstein:
Right.
Karplus:
But that information is not directly accessible to you, as a rule. So you must infer those from observations of the system. You may watch the water level. Let's say you observe fifty wells for several years. And from those observations you try to infer what is a probable mathematical model of the system.
Goldstein:
I see.
Karplus:
It's called an Inverse Problem. It's a tough problem that does not have a unique solution. And some of the techniques that we developed were particularly applicable for that.
Goldstein:
Would you would test the different models?
Karplus:
You might say that. Its called validation of the model. You would use the model to predict something that was not used in designing the model. And in that way, you would test it.
Goldstein:
So these were actual physical reservoirs that you were trying to model. Can you think of the locations of some of them?
Karplus:
In the Los Angeles area. We had nine aquifers within the Los Angeles city limits. And Los Angeles is a very arid place. So our water control district is always interested in these models, as were others from all over the world.
Digital computing; NSF application process
Goldstein:
And this was strictly digital, this work?
Karplus:
Yes.
Goldstein:
I see.
Karplus:
I've done nothing analog, since the late '60s.
Goldstein:
Now you said that one of your motivations for switching from analog to digital work was sensitivity to the needs of the graduate students. I wonder also whether the NSF encouraged that switch. Were they more interested in supporting work in digital computing?
Karplus:
Well, I never had the feeling that they were pushing me in any direction. It was really more of what the reviewers would say. You would send a proposal to NSF, they would select five reviewers and each reviewers gives a grade. And if the average grade is above a certain threshold then you would get money. So if you go too far afield, you run afoul of this review process.
Goldstein:
Right.
Karplus:
Now, to some extent, the project director at NSF has your fate in his hands by the way he picks the reviewers. But just to answer your question, I did not get the impression that NSF was pushing me in one way or another. But there was a general push to accent science and research as distinguished from overly practical applications.
Goldstein:
Right. One other question I'm wondering about, these aquifers in LA or elsewhere in the world, you would create the models but, do you have any sense of what became of your results.
Karplus:
Really, my output was methodology rather than the specific numbers. I provided tools that other people then used to do what they wanted to do.
Goldstein:
So you weren't working closely with the utilities or water managers?
Karplus:
No. Only in the sense that I tried to make my models applicable and relevant.
Transition from NSF to industrial funding, 1970s-1980s
Goldstein:
I see. And then you say that phase took you through to the end of your NSF support. Did you change research topics or find an alternate funding source?
Karplus:
Yes, there were a number of funding sources, but I guess I got older and older and my contemporaries were all moving elsewhere and the available money became smaller and smaller, and the criticism more and more arch.
Goldstein:
I see.
Karplus:
And, in fact, at just about that time the computer science department, of which I was chairman, got a very big NSF grant, in the multi-million dollars.
Goldstein:
For facilities?
Karplus:
Yes, it was for facilities. But that also made a lot of other things happen. So I was not that hungry.
Goldstein:
So you ceased applying?
Karplus:
Right. And then starting with the '80s, industrial money became easier to get. And the state of California developed a number of programs. So, all in all, I guess it became less and less interesting for me to get financial increments from NSF.
Research significance; hybrid computing
Goldstein:
- Audio File
- MP3 Audio
(114 - karplus - clip 1.mp3)
I see. If I could ask you to describe the impact of your work, how would you do that? Or perhaps even its forebears, the work that you built on. I'd like to, when I write about your research, try to locate it in a continuum of development.
Karplus:
I would say that I was one of the pioneers of hybrid computing, the interaction of analog and digital computing, and maybe one of the three, maybe four, most influential people in that area. I would say that the digital simulation languages attracted comment, but did not have all that much impact. The method for system identification that I've described, and in fact we're still working in that general area now, using neural nets, have stimulated a lot of other people to do things. Now that's as far as methodology is concerned. As far as applications to specific areas are concerned, I think I've had a impact on a number of projects in the aerospace industry within NASA.
Goldstein:
I see. When you say you were one of the three or four pioneers and leading figures in digital analog hybridization, who else worked in that area?
Karplus:
Granino Korn, George Bekey, and Robert Howe.
Goldstein:
And where are they? Do you know?
Karplus:
Howe is at University of Michigan. Bekey is at University of Southern California. Korn is at University of Arizona.
Goldstein:
Thank you. This has been very helpful. And I look forward to getting your cv. I want to offer you an opportunity to characterize your work for the NSF, or offer any other statements you feel would be useful.
Karplus:
I would love to read over what you write before you submit it.
Goldstein:
Oh, certainly. Well, thank you for talking to me.
Karplus:
All right.
Goldstein:
Goodnight.
[End of interview]
