Interview with Paul Roossin

Matthew: I met you socially, yet we had a conversation about rheology, microscopy and even
jazz in that first discussion. It made me want to be your friend and work with you. The way you
talk about these subjects and other science topics, it is almost as if you feel the disciplines
are easy, or at least accessible to anyone. Is this the case, and have you always had that
confidence?

Paul:  I started reading science and technology books at an early age, and it became clear to me that all the scientific areas of study are inter-related. Learning about physics helps one learn about chemistry, which in turn makes it easier to learn biology.  It’s a connected web, so that once you are familiar with the main supporting filaments, it’s not too difficult to add new connections.  I feel that the basic concepts of each scientific discipline are indeed accessible to anyone. It does takes a commitment to proceed with the next step – learning how to think analytically and logically – but it is well worth it, for this is a skill that brings deep emotions of joy, satisfaction, and humility.

Matthew: I am especially interested in your work in machine translation at IBM, and how your
views on AI influence you now. I recently read that Larry Page would not hire a head of AI for
Google because he felt that AI should be incorporated throughout the entire company, not a
department. Do you think that this should be true for most companies, including Nanotronics?
How might Nanotronics incorporate some of your ideas on the topic?

Paul:  My views on AI were shaped by my colleagues at IBM Research. The prevailing AI paradigm when I joined the lab centered on the notion that humans are experts at specific and different tasks.  In order to recreate human intelligence in a machine, one needs to codify the knowledge about a specific task and to give the computer logical procedures which mimic human deductive thinking.  By the way, the word “expert” as used here might mean chess-master or medical diagnostician, but could also mean being competent at tasks that all humans do well, such as understanding language, segmenting a visual scene into its components, or recognizing faces. My colleagues convinced me that this technique would never work past a toy, demonstration program.  To make a machine seem convincingly intelligent, it is necessary to supply it with tons of tagged training data.  We used to say, “there’s no data like more data.”  The training data is then used to tune statistical models that may have many millions of parameters.  This is a very different view from what was happening in the 1980s, and it is only because these methods (pioneered at IBM) worked so much better than what had come previously that other groups started to take notice.  Now, everyone who does large vocabulary speech recognition or machine translation uses this information theoretical-approach.  It is catching on across the other subject domains of AI as well.

We are still very far from a general AI methodology, however –  that is, one that doesn’t need an enormous amount of model customization and huge quantities of specific training data to get a new computer programming task to behave intelligently.  If a large company like Google wants to promote intelligent programs across the board, then it makes sense for them to delocalize the process.  For instance, the map-reduce methods Google developed to handle huge amounts of data, are a company-wide approach which encourages and supports statistical methods of AI.  But, for a company like Nanotronics, where we have a small number of very targeted tasks we are attempting to solve, I think it is fine to create a custom solution to the most pressing problems.  Of course, I would like to see these problems solved using statistical modeling, if the training data were available.

Matthew: I was in a research lab recently talking with a prominent cancer scientist. About
30 minutes into that discussion, you entered the room, sat down on the sofa and started
contributing. I picture you going around New York doing this all of the time. Is my image of your
scientific lifestyle accurate?

Paul:  Sort of.  I do a lot less of that now than I used to.  At IBM Research, and at Rockefeller University before that, I loved popping in on my friends’ labs, asking what they were up to, and then start to brainstorm with them.  At Rockefeller, I was friendly with Joshua Lederberg, who was president of the institution, and a great molecular biologist.  One day he showed me a few large folders brimming with telegrams and hand-written letters.  They had arrived shortly after one of Josh’s predecessors closed the small dining room and built the current large cafeteria.  Apparently, all the doctoral students, post-docs, and faculty were required to eat at the small dining room four days a week, with attendance taken, and the one rule was that you have to sit next to new people as often as possible.  When the large new cafeteria was opened, that rule along with mandatory attendance eliminated. The letters in Josh’s folders, from famous scientists the world over, came pouring in. They all said the same thing: Please don’t change the R.U. dining hall, or you will destroy the greatest scientific resource on the planet.

Josh’s point was that good and exciting things happen when thinkers from different disciplines get together and chat casually.  My friend John Cocke at IBM used to say the same thing; in fact, his official job at IBM for many years was just to roam the halls and brainstorm with researchers, programmers and project managers.  Here’s someone who invented RISC architecture as well as half of compiler theory, and he felt his best contribution was to facilitate conversations within the IBM community.

These people and others instilled in me a deep love for diving into topics and trying to come up with new ways of looking at problems.  If you are focused in a narrow area, you end up thinking along the party lines after a decade or two; it’s easier for someone not inculcated in the conventional understanding of a topic to bring innovation and lateral thinking to the issues.  In any case, I derive more pleasure from knowledge breadth than from depth, which is why I like and feel comfortable to drop in on most any scientific discussion and add my two cents.

Matthew: When I think of innovation and science I basically think of two different models from
the 19th century. I think of Michael Faraday who learned to understand electromagnetism,
but had no idea about an application for it, and I think about Thomas Edison who applied
electromagnetism but did not understand the theory behind it. Do you consider yourself
following either of these paradigms?

Paul: Uhh, I don’t think I would ever dare consider having my name mentioned in the same breath as either of these two great discoverers.  But as to the basic research versus engineering dichotomy you refer to, I believe that the similarities between the two tasks far exceed their differences.  To me, the important thing, the essence that unifies these superficially disparate activities, is that they both rely upon the scientific method for advancement.  Both Faraday and Edison tried to achieve their goals by having a falsifiable hypothesis, testing it experimentally, incrementally refining that hypothesis on the basis of the experiments’ results, and iterating this process.  Occasionally, it pays (usually after a series of failures) to throw the whole thing out the window and start anew with a completely different hypothesis.  This is science; this is the scientific method.

Try this game:  Look around.  Consider the objects in your environment.  Think about when each was invented, when its components were invented, and what it took for that object to have been conceived, designed, and manufactured.

This is a game I like to play while daydreaming.  What I notice is that almost every non-living thing in our environment is a new invention, one that couldn’t have been there in that form more than a hundred or two hundred years ago.  Most things couldn’t have been made even fifty years ago.  This is not the way that humans have lived over the millennia; change was glacial, and your life was pretty much the same as your parents and their parents.  Not so any longer.  And this is because of the rise of modern science.  This one invention, the scientific method, is responsible for almost everything you see and do in life.  It is, to me, the most profound invention of all times, and it is one of process, not of material.

So, Faraday was after fundamental understandings of nature, and Edison wanted to build practical devices.  I see that as more of a personality quirk than a different approach to the world.  Neither is more important or more essential.  Both are necessary.  Both are fun.

Matthew: Can you explain your basic views on science education? While you seem to be an
outspoken proponent of certain ideas for grammar school and high school, I feel that you might
have some skepticism about higher education. Am I right?

Paul:  I have profound skepticism about education as practiced in the States.  We live in a technological society, and thus as a group we need to decide to promote or to limit such things as genetic engineering, medical procedures, alternative energy sources, and weapons.  It is very hard for an individual to have a clear opinion on these matters if he doesn’t understand the scientific fundamentals that underlie such areas.

The problem is that science is not taught in a way that fosters passion and a joy of asking why and how.  It’s fabulous if you have been bitten by the how-does-it-work bug and are self-motivated, but, if not, science education feels like an endless parade of memorizing disconnected trivia and gobbledygook.

Part of the problem is that many K-12 science and math teachers don’t have the bug themselves, so they don’t know how to teach a love for the process.  All they know is how to teach a bookload of facts.  (Many can’t even do that, and so, as a group, teachers are loathe to take qualifying tests to prove them competent.)  I’ve noticed that once students get turned off to science or math, usually around grade seven or eight, it’s extremely difficult to get them to turn back on.  We’ve lost them forever.  It’s not that they don’t like science; it’s that they don’t like science class.  As has been demonstrated repeatedly, a strong research and technology sector makes for a strong economy.  Other countries see K-12 science education as a top priority; I’m very puzzled why the US doesn’t.

As to college education, that is a very big topic, but let me just say that I think there is something broken with a system that says you can’t get hired unless you have a college degree, but to get one means (for most people) to go heavily into debt, and end up not learning all that much during those four years.  The OECD publishes their rankings of educational success by country every three years, and the results from their 2009 report does not paint a rosy picture for the U.S.  I’m mystified why this doesn’t cause a bigger public outcry. I do think there are some innovative solutions that can be brought to bear on this problem, which I’d be happy to tell you or any readers of this interview about over a beer.  On me.