Sunday, June 24, 2007

Traveling: International Symposium on Information Theory

I am a big believer that traditional information theory and computer science theory are becoming much more intertwined. Perhaps, someday, the boundaries will be so fuzzy that we won't really distinguish them. Since I don't think that day is here, I'll describe the International Symposium on Information Theory -- a conference, it seems, most CS people don't know about -- where I'm headed this week.

ISIT is the major information theory conference of the year, by which I mean it's huge. I mean, ridiculously huge. Eight parallel sessions huge. I don't know the actual numbers, but I'd guess attendance is in the large hundreds, nearing one thousand, maybe more. I'll ask the IEEE people when I get there.

We don't have a conference like this in CS theory. If we get 300 people to a conference, it's a big deal. We prefer to split things up into smaller, more coherent units. There are pros and cons to both approaches. The obvious benefits of the big conference is that it gets everyone together.

With smaller conferences, CS theory conferences ostensibly have a higher quality bar. As you might expect at such a big conference, with acceptance rates of 50% or higher, the paper quality varies dramatically. You tend to see a lot more preliminary ideas or work that's not as fleshed out as a STOC/FOCS/SODA paper would be -- of course, in ISIT, you only have 5 pages. On the other hand, you'll also get breakthroughs presented here, like Amin Shokrollahi's Raptor Codes.

I am still a bit of an outsider at ISIT, as it is not my regular crowd, although every year I seem to know more faces. So you could view it as simply self-interest that I encourage more CS people to submit and attend. If you're doing anything related to codes (network coding, list decoding, low-density parity-check coding -- all topics many CS theory people are working on), you probably already know about ISIT -- there are lots of session on coding, naturally. But they also have sessions in cryptography, quantum information theory, compression, sequences and complexity, and network information theory. From that list of topics, you might think it's a CS theory conference! But that's the point -- the two areas are increasingly close, and more interaction, including CS people attending ISIT, would help both sides.

I hope to have interesting news to report.


Anonymous said...

I'm at ISIT right now (as Mike already knows) and was fortunate enough to sit next to a well established information theorist on the flight to Europe, which gave me plenty of time to ask him tons of questions about ISIT and the information theory community as a whole.

ISIT seems to be a great example of what happens when an entire worldwide community decides to have a full-fledged gathering every year, varying the location to even out the travel load, with the explicit goal of driving up attendance. As a result, acceptance rates are deliberately high (one of my reviews essentially said that the paper should be accepted merely because it was non-obvious and appeared technically correct; never mind that it's actually good, or so I think). Also, it looks like they really take the limit of two presentations per author seriously, and not just to generate money, but to really get everyone they can, including students. I even heard that elements of the "old guard" occasionally push for a "one presentation per author" rule. Furthermore, papers are short: 5 pages max, and the limit was 1 a few years ago. Once again, the primary purpose of the conference seems to be to get everyone together and try to expose everyone to as much of the breadth of the field as possible. The really technical stuff should just be put in a preprint somewhere and wait for a journal. (Remember those? Other fields use them to make sure that ideas are completely developed and that papers are thoroughly reviewed for correctness and often revised for, get this, clarity.)

Maybe this is overly naive of me, but I'm also getting the impression that the theoreticians and practitioners get along better in information theory than in computer science. But I'll wait for Mike to give his take on the theory/practice divide in CS and related fields before I throw in my (less informed, but probably more inflammatory) 2 cents.

Anonymous said...

a similar huge conference i liked a lot is ISMP -- the intl. symp. on math. programming, which covers almost all aspects of OR. last year's was my first, and it didn't hurt that it was in rio ..


GASARCH said...

This points to a problem we have in CS or CS theory---
we have lost the ability to
have an INFORMAL
lighly-refereed conference.
COLT was originally a workshop
where most things got in.
NOw its highly competitive.
I think RANDOM was similar.
A low-pressure place to exchange
information and ideas is
a good thing. DAGSTUHL
does that, but thats
by-invite-only. A similar
problem is that diff fields
have diff conferences.
FCRC is trying to bring
people together, with
some sucess. But it has
its problems too.

Arvind Narayanan said...

I have friends who go to ISIT. I believe the number is more than a thousand.

Suresh Venkatasubramanian said...

two comments:

* it's not that we are unable to have lightly reviewed conferences (some would argue only half in jest that STOC/FOCS are already such conferences); it's that (at least in my experience) people are unwilling to imagine journals as the primary publication source, throwing up all kinds of objections that make little sense when viewed in relation to other disciplines.

* On the more technical point of fusing IT and theoryCS, one area that I find personally interesting is the use of info-theoretic methods in data mining. There, the idea is that you use the metaphorical power of compression and signal processing to analyse data in a relatively model-independent way.