Machine Learning (Theory)

12/4/2006

Structural Problems in NIPS Decision Making

This is a very difficult post to write, because it is about a perenially touchy subject. Nevertheless, it is an important one which needs to be thought about carefully.

There are a few things which should be understood:

  1. The system is changing and responsive. We-the-authors are we-the-reviewers, we-the-PC, and even we-the-NIPS-board. NIPS has implemented ‘secondary program chairs’, ‘author response’, and ‘double blind reviewing’ in the last few years to help with the decision process, and more changes may happen in the future.
  2. Agreement creates a perception of correctness. When any PC meets and makes a group decision about a paper, there is a strong tendency for the reinforcement inherent in a group decision to create the perception of correctness. For the many people who have been on the NIPS PC it’s reasonable to entertain a healthy skepticism in the face of this reinforcing certainty.
  3. This post is about structural problems. What problems arise because of the structure of the process? The post is not about individual people, because this is unlikely to be fruitful.

Although the subject is nominally about NIPS (which I have experience with as an author, reviewer, and PC member), the points may apply elsewhere.

For those that don’t know, it’s worth reviewing how the NIPS process currently works. Temporally, it looks like the following:

  1. PC chair is appointed.
  2. PC chair picks PC committee to cover many different areas. NIPS is notably diverse.
  3. PC committee members pick reviewers for their areas.
  4. Authors submit blinded papers.
  5. Papers are assigned to two PC committee members, the “primary” and the “secondary”.
  6. Reviewers bid for papers within their areas which they want and don’t want to review.
  7. Reviewers are assigned papers based on bid plus coverage.
  8. Reviewers review papers.
  9. Authors respond to blinded reviews.
  10. Reviewers discuss and rate papers.
  11. PC members digest author/reviewer interaction (and sometimes the paper) into an impression.
  12. PC members meet physically at the PC meeting.
  13. PC members present all papers that they believe are worth considering to other PC members and a decision is made.

Naturally, there are many details left out of this long list.

Here is my attempt to describe the problems I’ve seen:

  1. Attention deficit disorder. The attention paid to individual accept/reject decisions is (and structurally must be) small. There are several effects which drive this:
    1. The people on the NIPS PC are typically busy and time constrained.
    2. The number of papers assigned to individual PC members is large—perhaps 40 to 80, plus a similar number assigned as a secondary.
    3. Many of the people have traveled a very long ways to reach the PC meeting. Jetlag is common, and often significantly effects your ability to think carefully.
    4. The meeting itself is 2 days long. The average time spent on any decision must be less than 5 minutes, and everyone knows this. The implicit encouragement to digest a paper down to its most simple description is significant. No one on the PC has seen the paper except for the primary and the secondary (if you are lucky) PC members, so decisions are made quickly based upon relatively little information. (This is better than it sounds in most cases because effectively the decision was made by the primary PC member beforehand.)
  2. Artificial scarcity. NIPS is a single track conference with 3 levels of acceptance “Accept for an oral presentation”, “Accept for a poster with a spotlight”, and “Accept as a poster only”. It’s fairly difficult to justify a paper as “of broad interest”, which is ideal for an oral presentation. Will a neuroscientist really pay attention to this learning theory paper? Is this dimensionality reduction algorithm going to interest someone in learning theory? It’s substantially easier to justify a paper as “possibly of interest to a number of people”, which is about right for poster spotlight. Since the number of spotlights and the number of orals is similar, two effects occur: papers which are about right for spotlights become orals, and many reasonable spotlights aren’t spotlights because they don’t fit.
  3. The Veto Effect. If someone on the PC has a strong dislike for your paper, there is a very good chance for reject. This is true even when attention is explicitly payed by the PC chair to avoiding the veto problem. It’s even true when your paper has the strongest reviews in the area (no joke!). There are several fundamental problems here:
    1. People, especially in person, do not generally want to be confrontational. Consequently, if someone who is rarely confrontational speaks strongly against a paper, it’s rare2 for an alternate voice to be heard.
    2. It is easy to instill “fear, uncertainty, and doubt” in people. Was this paper covering the same material as some other paper no one knows? Are the assumptions criticizable? This problem is greatly exaggerated by attention deficit disorder.

It is easy to complain about these problems and substantially harder to fix them. (There is previous discussion on this.) Here is my best attempt to imagine fixes.

  1. Attention Deficit Disorder. The fundamental problem here is that papers aren’t getting the attention that they deserve by the final decision maker. Several changes might help, but nothing is going to be a silver bullet here.
    1. Author responsibility. Unfortunately, some authors abuse the system by submitting papers which should not be submitted. Much of this has to do with inexperience—many authors are first time paper writers. For these authors, some better effort educating people about what is an appropriate paper is good. This year, an effort was made to do this, and followups may be helpful. For a small fraction of papers, authors intentionally skate the edge of what is reasonable. Should an ICML paper with 30% different content be submitted to NIPS? This small fraction takes more time than their fraction indicates and (frankly) isn’t always caught. Some form of “shame list” may be an appropriate way to deal with this, although much caution would have to be exercised.
    2. Many of the problems here are unremovable artifacts of a physically present PC meeting. Going to a virtualized process would eliminate these problems (and introduce others). Any such decision would have to be carefully considered, but it is not impossible—there are plenty of succesful conference committees which never meet physically.
    3. The PC meeting can be run a bit differently.
      1. Bob Williamson and I managed to go through our secondary assignments and make independent decisions, then reconcile. In contrast, for most papers, the secondary PC member was inoperative at the PC meeting. This made some difference, and it’s easy to imagine that systematically having this reconciliation be a part of the PC meeting is helpful. The reconciliation step does not take very long and is parallelizable.
      2. Not making a decision at the PC meeting could be a real option for a small number of troublesome papers. There is perhaps a week-long timegap between the PC meeting and the release of the decisions during which decisions could be double checked. This option must only be used rarely, and never as a means for excluding interested PC members from the decision.
      3. Information can be more widely shared. I don’t see any real advantage to limiting the knowledge of papers not in your area to “title+authors”. At the PC meeting itself, it would be helpful to have all of the papers available to all of the members.
  2. Artificial Scarcity. My understanding is that the makers of NIPS purposefully preferred a single track conference, and it’s hard to argue with the success NIPS has enjoyed. Nevertheless, it seems notable that the NIPS workshops (which are excessively multitracked) are more succesful than the NIPS conference by some measures. Going to a two-track or partially two-track format would ease some of the decision making.

    Even working within the single track format, it’s not clear that the ratio between orals and spotlights is right. Spotlights take about 1/10th the time that an oral presentation takes, and yet only 1/10th or so of the overall time is allocated to spotlight presentations. Losing one oral presentation (out of about 20) would yield a
    significant increase in the number of spotlights, and it’s easy to imagine this would be beneficial to attendees while easing decision making.

  3. The Veto Effect. The veto effect is hard to deal with, and it’s only relevant to a small number of decisions. Nevertheless it’s important because some of the best papers are controversial at the time they are published. The are two ways I can imagine for dealing with the veto effect: (1) allowing author feedback (2) devolving power from the PC to the reviewers. Allowing author feedback would have to be coupled with delayed decision making. Eliminating the power of the PC to reject very highly rated papers is also controversial, but may be worth considering.
2 Comments to “Structural Problems in NIPS Decision Making”
  1. Alina Beygelzimer says:

    There is a relevant post
    here.

  2. Elad Yom-Tov says:

    ” … Much of this has to do with inexperience.”

    SIGIR has a system whereby authors can apply for “mentors”, relatively senior people in the IR community who help them improve their articles before they are submitted to the conference. Maybe NIPS should follow their example.
    See http://www.sigir2007.org/mentoring.html

Sorry, the comment form is closed at this time.

Powered by WordPress