Tuesday, January 14, 2020

ITCS 2020, Reflections

I've spent Sunday/Monday at ITCS, or Innovations in Theoretical Computer Science, where I am giving a talk on this paper on Scheduling with Prediction and the Price of Misprediction (LIPIcs page) (which is one of several recent works on Algorithms with Predictions (powerpoint slides)).

I'm told it's 10 years since ITCS started as a conference, and I was one of the skeptics that really did not think it was a good idea 10 years ago.  So as I'm sitting in the sessions, what do I think now?  What are the pros and cons of ITCS?

On the negative side, ITCS is not very big.  It is just over 100 people registered, so it's like a big workshop/small conference size.  (And considering that it's usually held in central places with lots of students, those numbers are buffered by locals.)  Somehow, it's scheduled the 2nd week of January, right after SODA, which seems problematic, and certainly (if it's kept that way) may keep it from getting any larger.  The number of "senior people" around at any one time seemed generally small, a problem for things like "Graduating Bits" (see below).  As ITCS, at least 10 years ago, was supposed to build up to another "premier" TCS conference, focused on "innovative" work, the attendance seems a bit disappointing.

On the neutral side, the conference is single-session, and to make that work, talks this year are limited to 12 minutes.  Your mileage may vary on whether you think this is good or bad;  it seems to work.  (My take:  I slightly prefer parallel sessions, because it means there's more times where there's something I want to see, and that's more important than the times where there are two talks at the same time I want to see.  And 12 minutes is certainly workable but maybe a few minutes short.  But again, that's just my opinion.)  This year (and possibly going forward), some papers were accepted without a full talk -- instead they have a 3-minute talk and a poster in a poster session.  Again, it's not clear to me if this is good or bad (though more paper acceptances makes me lean to the good side), but it seemed to work fine and people were happy with it.  (Such papers are considered full publications and treated the same in the proceedings.)

On the positive side, the conference seems well run.  It's held in a university building, so no expensive hotel costs;  instead they're generous with food and keep registration costs reasonably low.  They use LIPIcs, which provides a good system and approach for publishing papers at low cost.  (Note, I was until recently part of the LIPIcs editorial board, so I'm biased there.)  They seem to be covering their expenses from what I understand.  The business meeting was blessedly short.  They're recording all the talks.  They're doing interesting things like "Graduating Bits" for people who are job-looking (where people graduating or coming out of a postdoc give short talks about their work).

In terms of content, it seems really good.  I've seen several good talks and interesting papers.  While I'm not sure how to quantify whether ITCS work is more "innovative" than the work at other major TCS conferences, I do actually think they are noticeably more open at ITCS than other conferences about accepting papers based on the paper's underlying idea rather than on "technical mastery".

My thoughts 10 years ago were that ITCS was not a great outcome for the community, and that instead the community should push for:

1)  Aiming to do better about opening up the criteria for paper acceptance, including weighing innovation/practical relevance in reviewing papers at FOCS/STOC/SODA.
2)  Increasing the number of papers accepted to these conferences, as too many good papers were being rejected.

Viewed under this lens, ITCS could, I think, be viewed as a success.  The theory community seems unwilling to expand conferences by accepting more papers.  (I note that while the STOC theory-fest has changed and expanded STOC, it hasn't really seemed to increase attendance, and while the number of accepted papers has increased slightly, it hasn't kept pace with the growth in the field.)  ITCS provides another venue for high-quality theory papers, thereby increasing the number of such papers published each year within the theory community, and I think it is generally viewed as a high-quality conference.  And, as I mentioned, ITCS seems at least somewhat more flexible in its criteria for what is an acceptable paper.  ITCS has, I think, in these regards been a benefit for the theory community.

However, despite the success of ITCS, I think it's a band-aid on structural issues in the theory community.  While these issues are complex and many-sided, just comparing with the growth and excitement in the AI community is a little depressing.  Indeed, what I see is AI absorbing significant parts of the theory community;  lots of theory papers now end up at NeurIPS, AAAI, ICML, or AISTATS, because the theory community doesn't seem to have room for them, or doesn't judge them as sufficiently significant.  I view this as a problem for the theory community, the result of the problems I saw 10 years ago for which I didn't think ITCS was the right response.  (Though perhaps it was/is not really viewed as a problem by others;  all of CS, including theory, seems to continue growing;  theory just seems to me to be growing less than it could or should.)

To conclude, my thanks and congratulations to those many people that have organized and maintained ITCS over the years;  I appreciate your work, as I think the whole community does. 

By the way, I thought it never snowed in Seattle, so I'm confused; what is all the cold white stuff outside?


Grigory Yaroslavtsev said...

This goes back to many years ago, but just colocate SODA and ITCS, imo: http://grigory.us/blog/stoc-focs-proposal-colocate/

Michael Mitzenmacher said...

I agree in spirit, but I can imagine logistically it's very problematic.

ITCS clearly wants its independence -- for some years it was under ACM/SIGACT, but left. Co-locating with SODA means either putting the conference under the SIAM umbrella, or reaching a level of cooperation that is probably difficult to maintain each year.

For example, SODA (being quite large, and with several co-located activities already) is generally held at a major hotel. ITCS could go this route, but it could raise costs, and it would lose other aspects of its independent feel. ITCS could try to be "right after" SODA in the same city, but SODA is not always held someplace with a willing university and set of local arrangements people willing to host.

My take is co-locating ITCS and SODA (either at the same time, or one right after the other) makes a lot of sense. But my sense is also the organizers of ITCS are not willing to give up the independence required to make that happen on an every-year basis. It's a non-trivial tradeoff.

Jakob Nordstrom said...

The comparison to AI is very interesting, and is something I have been thinking about for a while.

Although I have not been around for long enough to actually know, it seems to me that the theory community has chosen to keep STOC/FOCS (almost) the same size or has let the conferences grow only very slowly, although the area as such has expanded a lot. This has meant that it has got too hard even for very good papers to get accepted, and that entire subareas have more or less left to form their own high-profile conferences.

It is not clear to me that it would be fair to say that visiting STOC of FOCS these days gives a representative sample of all the exciting work going on in theory. Even more importantly, the attendance numbers show that these are not natural must-go-to conference to hear about research and network with people, and sadly the STOC Theory Fest (which is fantastic) does not (yet) seem to have changed this.

The AI community has chosen a very different, big-tent, approach, where AAAI and IJCAI still accept pretty much all (sub)areas. This is a bit like having FCRC twice a year, and means you really meet (almost) all people when going there. I think this is a better way to keep the community together, and that TCS has been hurting itself a bit by splintering off into subcommunities (though, to be sure, AI has other problems, such as the conferences getting very, very big, and the level of noise in AAAI/IJCAI reviews and acceptance decisions these days is a very real problem, in a way that I do not think is the case for theory).