I just got back from SIAM CSE 17 in Atlanta where I attended the first 4 days (Monday -Thursday). This has been the fifth SIAM CSE conference that I have contributed/participated in (starting in 2009) out of the nine that have taken place thus far (if I am not mistaken).
I must say that I have always enjoyed them as one can get exposed to a very wide range of applications of mathematical modeling without needing to be an expert in the underlying physics. It has always been inspiring to see different perspectives and to be able to transfer tools from one application to another. Another plus, particularly in the last few times, has been the significant proportion of presentations on "Uncertainty Quantification" (which is the general area I also work on), and this despite the SIAM UQ dedicated conference. This trend continued this year and UQ was again one of the top themes of the conference.
On most days there was more than 30 (!) parallel sessions happening (simultaneously). This seemed a bit more than previous times and it was always possible to find at least one talk (most often 3 or 4 in my case) to go to. As a result I missed a lot of talks that I would have liked to see. Having so many options is in many ways a plus. Nevertheless (this is where the criticism starts), I cannot say that it was easy to get a good idea of what each talk was about from the abstract (there isn't that much one can say in ~100 words) and not all talks presented necessarily something new (perhaps our groups' talks didn't either). I have always been envious of the system followed by the Machine Learning (ML) community in conferences such as NIPS, ICML etc. where the number of sessions is much fewer and, more importantly, the papers presented have undergone a (blind) peer-review with approx. 20%-25% being ultimately accepted. The submissions are 8-pages long (last I checked at least) which ensures a good balance in terms of richness of content and easing the burden for reviewers. I do not need to tell you that the impact of these papers is generally much higher than journal papers. Furthermore, the speed by which new research is disseminated is much greater than I have typically experienced in our community and the journals that we publish. It is not uncommon that some of the NIPS papers for example have a few dozens references only a few months after they have appeared.
One might say that such a system is too restrictive, but in my mind the (long-term) benefits outweigh the disadvantages. It is not unlikely that the first time such a system is applied to a SIAM CSE conference, the number of registered participants would decrease, but I have no doubt that in the medium to longer run our conferences would be as popular (for both academia and industry) as the aforementioned ML ones are nowadays. I would therefore strongly support any initiative in moving in that direction and adjusting the ML model to the special characteristics of the SIAM CSE community. Perhaps our submissions should be 10 or 15 pages, perhaps we should have 3 or 5 parallel sessions instead of one, but I am certain that such a move would greatly amplify the impact of our contributions, and of the SIAM-CSE community to the scientific landscape.
-S. Koutsourelakis (p.s.koutsourelakis@tum.de)