Sunday, October 29, 2006

The Generalist

As you readers know, my colleague Janos Simon went to FOCS last week. Janos wasn't an author, he wasn't on the program or organizing committees, he wasn't an invited speaker, he didn't sing in the band and he doesn't live within driving distance of Berkeley. Janos Simon went to FOCS not because he had an official reason, but rather to keep up with the latest results in theory. FOCS had low attendance this year (about 220) because it doesn't attract many Janos Simons, people who travel to FOCS because they want to, not because they feel they have to.

Janos Simon is part of a dying breed, a generalist in theoretical computer science. Janos goes to most STOC and FOCS conference and few specialized conferences like SODA and Complexity. In three decades of research, Janos has studied complexity, algorithms (distributed, parallel, randomized), VLSI and networks well as having interests in nearly every area of theory. It's impossible to arrange parallel sessions at FOCS in way that won't cause a conflict for Janos.

Back in the 70's a graduate student could keep up with the latest research in all of TCS and produce significant results in his or her career in many different areas. Today we seem to push out very specialized researchers focused not just on some subarea like algorithms or complexity but even in more specialized areas like streaming algorithms or PCPs.

Such specialization has its risks, for example getting so caught up in an area that you lose track of the motivation for studying the topic in the first place. Also we lose some sense of community when theorists in different areas lose interest in following each others work. But can a young Janos Simon survive today as a generalist or are we doomed to further specialization as students search for successful research agendas?


  1. But can a young Janos Simon survive today as a generalist or are we doomed to further specialization as students search for successful research agendas?

    I think this is a great point. It is an unfortunate consequence of the fact that a student, even one specializing in an area such as streaming algorithms or PCPs, must pick up so much baggage before he can make significant contributions in his/her own area. This makes it harder even for an interested student to pick up other subfields. After all, the student wants to graduate in a decent amount of time!
    Wonder how Mathematics as a discipline has dealt with this. After all, it is much older a discipline than Computer Science and presumably a grad student has to pick up even more baggage. Wonder whether Math grads also suffer from the same problem.

  2. I think this is a problem specific to TCS, not CS in general. If you would like to be a generalist, do research in an applied field of CS.

  3. This is a very interesting and provocative post. So let me respond by saying its just wrong, wrong, wrong, wrong. (I'd put it more bluntly, but I try to watch my language on the net.)

    Maybe I'm out of touch. But let's look at recent theory graduates like Bobby Kleinberg or Shuchi Chawla. Want to take a look at their home pages and explain to me how they're not general enough? Even when I look at, say, a Scott Aaronson, who seems a trifle obsessed with quantum (not that that's a bad thing, Scott!), I see his strong connections to the rest of complexity, and learning, and so on.

    Then there are theorists like, say, me (and Muthu and Moses and Piotr and Ashish and Karger and Kleinberg and the list goes on...) who, besides being theorists, also travel in other circles, such as networking/databases/information theory/AI/computational biology etc. Some of us are just so general our interests have gasp actually expanded beyond just "theory"!

    Really, the idea that TCS people, including the young ones, are not sufficiently general is just so preposterous, I feel duty-bound to correct this misstatement.

    If you want to figure out low attendance at FOCS, I'd consider alternative explanations. One obvious one is funding. I can only afford to go to so many conferences a year; when I don't have a paper, I have to be very choosy. Given what I've seen of the general funding situation, I'd guess many people have to be mighty choosy about conferences where they don't have a paper.

    OK, I'm kind of lying, I could swing FOCS monetarily, but I can't really explain to my wife and two kids why I'm traveling cross country for a conference I don't have a really good reason to go to. (I'm not sure, but it may be that I haven't been to a non-local conference where I didn't have a paper or was speaking since my first child was born.) I don't know how many others are in the same boat.

    I have other suspicions about issues related to low attendance -- perhaps some locals just dropped by for a day without signing up; perhaps the format should be changed to accept more papers and have more parallel sessions; or perhaps it even has something to do with the content (which is a discussion for another day).

  4. That's a very biased comment. Existence of a few "generalists" doesn't change the fact that many students and faculty are very specialized.

  5. One thing they should/could do to get attendance up is to have a special price for postdocs, closer to the student price. Postdocs (like students) should especially go to the conference to keep on top of what is going on, but with a fee of $455, it is too expensive. When you are an academic postdoc, you do not have your own funds. You are similar to a student in that you have to ask someone else for money, so asking for that kind of fee on top of the other costs is just too much if you don't have a paper.

  6. 1. postdocs should definitely get student rates (or at least be able to complain to the local chair, and then get student rates)

    2. I don't think of the people on mike's list as generalists in the sense that Lance intended. it's almost like responding by saying "I know people who build canoes, collect Latvian stamps, and go to all the OR conferences!" Generalists in Lance's sense are people like Arora and Feige and U. Vazirani.

  7. We were worried about the same issue when I was in graduate school-- ``In the seventies, it was so easy to learn all the complexity there was, but now we need to get so much background to learn an area before doing research, we can't hope to master many areas''. I think we were wrong then (and especially if you look back at the careers of the people I was worrying with), and I think it's wrong now to think that a successful research agenda requires over-specializing. (I also see absolutely no sign that ToC people are more specialized than experimentalists. ) However, it's always been tough to be a
    Renaissance researcher.

    Again, when I was in grad school, I heard some wise advice from Dick Karp. He said that to do exciting research on important problems that other people also have tried and failed to solve, you need to identify an ``edge'', some skills you have that even very talented other researchers don't. Specializing is one way to get an ``edge'', and your ``edge'' often leads to a focus that may look like over-specializing to others, especially as a grad student when your resume has 6 papers, 4 of which are related. But other edges are things like the ability to formulate new problems, a knowledge of an applications area, or mastery of a certain area of mathematics that might have many applications.

    I've a taxonomy of successful researchers into three groups. Issues researchers center their work around a set of questions, such as ``Which theorems are hard to prove?'' or ``What does it mean for an algorithm to learn?''. They'll consider many formulations of these questions, and each answer leads to further questions, possibly requiring new techniques.
    Technique-focused researchers have a ``bag-of-tricks'' and try to find the problems that they know how to solve with their ``bag'', no matter what area they arise in. They love to add new techniques to their bag-of-tricks, and solving new problems is just a way of proving the utility of a technique. Finally, problem-focused researchers fall in love with problems that are beautiful, and work hard on those problems that appeal to their particular aesthetic.

    All of these approaches can be more or less narrow. The secret is to pick a middle-ground that's what you can handle. Also, I've found that when I get too focused, I need to pick a project that has absolutely nothing to do with anything I've done before, just to get me out of a rut.

    Finally, I want to mention that when you are advising students, you need to be broader than when you are working on your own work. You need to give each student some area where they can shine independent from each other and from you. That means you'll need to stretch beyond your current interests in some way.

    So I think current graduate students will always look more specialized than older researchers, but this is independent of the state of the field, and more reflective of the phases of a research career.

    Russell Impagliazzo

  8. As a graduate student, I find it very hard to generalize. Not because of any pressure on me, but because learning topic after topic after topic without going really deeply into any one topic is very tiring and unsatisfying (even just two is difficult). I am trying to focus on one area now and to use that topic as a lens to continually include more of theory into my field of view. However, this all takes time (much to my frustration).

    When students specialize, it may not be that they are becoming specialists, but that they are at the beginning of becoming generalists. After three decades, it would seem natural to look at more than one topic; but for three years, it may be more fruitful to just focus on one, very gradually expanding topic.

    Any thoughts on this? Will this strategy work?

  9. I was going through one google video(educational) in which the person says that 20th century was the century of specialists but the 21st century is going to be for those people who can work in many areas. Moreover, the access to information on net (specifically videos) will make people to encroach into new research areas very easily. Infact some time back I blogged on how to encroach into new research field.

  10. This is a very interesting and provocative post. So let me respond by saying its just wrong, wrong, wrong, wrong.

    I can't see what your argument is (besides giving few counter examples, which does not constitute real evidence).

    It is only natural that as a discipline becomes more mature, things would become more complicated, and more technically demanding, which yields more specialized sub-areas.

    I see this as a blessing; It means that TCS has to offer deep and profound questions with real mathematical challenge.

    I wouldn't want to work in a field where every tiny result is a `profound' one.

  11. Overall, TCS is both deeper and broader than it was 25 years ago (25 years ago only a select few knew/used Chernoff bounds, there was no learning theory, quantum computing, cryptography was in its infancy, there was little game theory and no mechanism design.) However, despite all the changes, it is still remarkable to me how low the barriers are to entry into different subfields of TCS. Often it is essentially a matter of learning one or two new techniques or ideas.

    We all tend to concentrate on things at which we are successful and this soemtimes causes our research papers to be narrowly focused, but that should not limit what we read. We should be omnivores as readers. Reading broadly within the field is a very important and valuable habit to get into, particularly beginning as a graduate student.
    While it is important to become an expert at something, there is plenty of time in a graduate career for more than that - we all need to take breaks from our research problems and that adds up to a lot of time.

    The very fact that barriers to entry are so low make this outside reading a good strategy. You may find some low-hanging fruit elsewhere or see connections that others have missed. There is a recurring phenomenon in our field in which a solid grad student working slowly in a narrow area for their PhD suddenly seems to blossom into a star as a postdoc after having been exposed to other areas and research problems.

    Conferences like FOCS and STOC are great precisely because they not only provide the latest and greatest for the experts in their respective subfields of TCS, they also provide a veritable smorgasbord that makes it easy to be an omnivore.

    BTW: Lance's post was inspired in part by FOCS attendance this year but I don't think that one should base too much on this number, especially in making decisions about parallel sessions or acceptance rates. This year's FOCS with parallel sessions was at least as well attended as the most recent previous FOCS in the Bay Area. Last year's FOCS in Pittsburgh had over 300 attendees (a high for several years) with single (not parallel) sessions and a lower acceptance rate than this year.

  12. Lance's post was inspired in part by FOCS attendance this year but I don't think that one should base too much on this number

    Let's not get into another one of this pointeless arguments of "FOCS is fine", "no it isn't", "yes it is". Let's just move on and not beat this dead horse any further.

  13. Even when I look at, say, a Scott Aaronson, who seems a trifle obsessed with quantum (not that that's a bad thing, Scott!), I see his strong connections to the rest of complexity, and learning, and so on.

    As a current graduate student who is rather less prolific than Scott, I'd suggest that he does not exemplify the typical graduate experience.

    Breadth is indeed an increasing problem for students, but I don't know what we can do about it besides thank God we're not in mathematics...

  14. Maybe FOCS attendance is low because people realize it's much better to publish in journals.

  15. As a grad student in the northeast, I would love to go to FOCS, but my advisor, and theorists in general, just don't have much funding. I go to TCC every year and that's about it (though my interests are more general than cryptography; I've been to FOCS once; and someday I'll make it to Crypto.) But otherwise I feel, as a grad student, generally more isolated from the community than when I was an undergrad!

    I guess I'll just have to try to get a paper accepted there.

    In the meantime, I'm glad to read all the blog posts here, and on Luca Trevisan's blog, about FOCS, though I wish I was there. Are there other theory blogs available now? These are great, and don't require grant money.

  16. It seems as though specialists will automatically spring up in response to the existence of unsolved problems in technical theories. But just reading the papers of a good generalist won't rub off so easily, because the generalized insight is not necessarily in explicit form.
    To have a healthy environment for developing generalists probably requires something more--good survey papers for one, and survey papers organized along each of the axes Russell mentioned. And of course, surveys should be published--half the time it's what one is looking for--and notable survey writers like Goldreich and Trevisan should be applauded.

    To the previous poster I humbly submit my own complexity-focused blog, with the caveat that, as a first-year grad student, it's a journeyman effort. But my tastes have to a fair degree been formed thru Lance & Luca's work, so maybe some of the appeal will carry over.

    I urge any and all to help contribute to a more vibrant web life for TCS, as a direct antidote to the relative inaccessibility of conferences.

  17. I must say that I am in full agreement with Prof. Impagliazzo's comment. I am a graduate student currently in an applied area but aching to study much more of TCS.