Posted on

NIPS 2008 workshop on Kernel Learning

We’d like to invite hunch.net readers to participate in the NIPS 2008 workshop on kernel learning. While the main focus is on automatically learning kernels from data, we are also also looking at the broader questions of feature selection, multi-task learning and multi-view learning. There are no restrictions on the learning problem being addressed (regression, classification, etc), and both theoretical and applied work will be considered. The deadline for submissions is October 24.

More detail can be found here.

Corinna Cortes, Arthur Gretton, Gert Lanckriet, Mehryar Mohri, Afshin Rostamizadeh

Posted on

Who is Responsible for a Bad Review?

Although I’m greatly interested in machine learning, I think it must be admitted that there is a large amount of low quality logic being used in reviews. The problem is bad enough that sometimes I wonder if the Byzantine generals limit has been exceeded. For example, I’ve seen recent reviews where the given reasons for rejecting are:

  1. [NIPS] Theorem A is uninteresting because Theorem B is uninteresting.
  2. [UAI] When you learn by memorization, the problem addressed is trivial.
  3. [NIPS] The proof is in the appendix.
  4. [NIPS] This has been done before. (… but not giving any relevant citations)

Just for the record I want to point out what’s wrong with these reviews. A future world in which such reasons never come up again would be great, but I’m sure these errors will be committed many times more in the future.

  1. This is nonsense. A theorem should be evaluated based on it’s merits, rather than the merits of another theorem.
  2. Learning by memorization requires an exponentially larger sample complexity than many other common approaches that often work well. Consequently, what is possible under memorization does not have any substantial bearing on common practice or what might be useful in the future.
  3. Huh? Other people, thank you for putting the proof in the appendix, so the paper reads better. It seems absurd to base a decision on the placement of the content rather than the content.
  4. This is a red flag for a bogus review. Every time I’ve seen a review (as an author or a fellow reviewer) where such claims are made without a concrete citation, they are false. Often they are false even when concrete citations are given.

A softer version of (4) is when someone is cranky because their own paper wasn’t cited. This is understandable, but a more appropriate response seems to be pointing things out, and reviewing anyways. This avoids creating the extra work (for authors and reviewers) of yet another paper resubmission, and reasonable authors do take such suggestions into account.

NIPS figures fairly prominently here. While these are all instances in the last year, my experience after interacting with NIPS for almost a decade is that the average quality of reviews is particularly low there—in many instances reviewers clearly don’t read the papers before writing the review. Furthermore, such low quality reviews are often the deciding factor for the paper decision. Blaming the reviewer seems to be the easy solution for a bad review, but a bit more thought suggests other possibilities:

  1. Area Chair In some conferences an “area chair” or “senior PC” makes or effectively makes the decision on a paper. In general, I’m not a fan of activist area chairs, but when a reviewer isn’t thinking well, I think it is appropriate to step in. This rarely happens, because the easy choice is to simply accept the negative review. In my experience, many Area Chairs are eager to avoid any substantial controversy, and there is a general tendency to believe that something must be wrong with a paper that has a negative review, even if it isn’t what was actually pointed out.
  2. Program Chair In smaller conferences, Program Chairs play the same role as the area chair, so all of the above applies, except now you know the persons name explicitly making them easier to blame. This is a little bit too tempting, I think. For example, I know David McAllester understands that learning by memorization is a bogus reference point, and probably he was just too busy to really digest the reviews. However, a Program Chair is responsible for finding appropriate reviewers for papers, and doing so (or not) has a huge impact on whether a paper is accepted. Not surprisingly, if a paper about the sample complexity of learning is routed to people who have never seen a proof involving sample complexity before, the reviews tend to be spuriously negative (and the paper unread).
  3. Author A reviewer might blame an author, if it turns out later that the reasons given in the review for rejection were bogus. This isn’t absurd—writing a paper well is hard and it’s easy for small mistakes to be drastically misleading in technical content.
  4. Culture A conference has a culture associated with it that is driven by the people who keep coming back. If in this culture it is considered ok to do all the reviews on the last day, it’s unsurprising to see reviews lacking critical thought that could be written without reading the paper. Similarly, it’s unsurprising to see little critical thought at the area chair level, or in the routing of papers to reviewers. This answer is pretty convincing: it explains why low quality reviews keep happening year after year at a conference.

If you believe the Culture reason, then what’s needed is a change in the culture. The good news is that this is both possible and effective. There are other conferences where reviewers expect to spend several hours reviewing a paper. In my experience this year, it was true of COLT and for my corner of SODA. Effecting the change is simply a matter of community standards, and that is just a matter of leaders in the community leading.

Posted on

NIPS 2008 workshop on ‘Learning over Empirical Hypothesis Spaces’

This workshop asks for insights how far we may/can push the theoretical boundary of using data in the design of learning machines. Can we express our classification rule in terms of the sample, or do we have to stick to a core assumption of classical statistical learning theory, namely that the hypothesis space is to be defined independent from the sample? This workshop is particularly interested in – but not restricted to – the ‘luckiness framework’ and the recently introduced notion of ‘compatibility functions’ in a semi-supervised learning context (more information can be found at http://www.kuleuven.be/wehys).

Posted on

The SODA Program Committee

Claire asked me to be on the SODA program committee this year, which was quite a bit of work.

I had a relatively light load—merely 49 theory papers. Many of these papers were not on subjects that I was expert about, so (as is common for theory conferences) I found various reviewers that I trusted to help review the papers. I ended up reviewing about 1/3 personally. There were a couple instances where I ended up overruling a subreviewer whose logic seemed off, but otherwise I generally let their reviews stand.

There are some differences in standards for paper reviews between the machine learning and theory communities. In machine learning it is expected that a review be detailed, while in the theory community this is often not the case. Every paper given to me ended up with a review varying between somewhat and very detailed.

I’m sure not every author was happy with the outcome. While we did our best to make good decisions, they were difficult decisions to make. For example, if there is a well written paper on an interesting topic which analyzes a flawed abstraction of the topic, should it get in? I would rate this a ‘weak accept’.

Here are some observations/thoughts about the process (Several also appear in Claire’s report).

  1. Better feedback isn’t too hard. The real time sink in reviewing a theory paper is reading it. Leaving a few comments, even if just “I don’t like the model analyzed because it misses important feature X.” is relatively easy. My impression of the last COLT was that COLT had entirely switched from minimal author feedback to substantial author feedback. This year’s SODA was somewhere inbetween, depending on the PC member involved, which is a definite trend towards stronger comments for SODA.
  2. Normalization There were very substantial differences amongst the PC members in what fraction of papers they wanted to accept, and this leaked into the final decisions. Normalizing reviewer ratings is standard operating procedure at some machine learning conferences, so I helped with that. Even with that help, further efforts at normalization in the future seem like they could help, for example in getting the decision on the paper above right.
  3. Ordering There were various areas where we tried to order all the reasonable papers and make a decision based on the ordering. Where the papers are sufficiently related, I think this is very helpful, and the act even changed my opinion on some papers a bit by putting them in better context. Not everyone imposed the same ordering, because there are somewhat different tastes: Do you care about the techniques used? (A traditional theory concern) or about the quality of the result? (I’m more focused here.) Nevertheless, it helped reduce the noise. Incidentally, there is substantial theoretical evidence that decisions by ordering are more robust than decisions by absolute score producing an ordering.
  4. Writing quality I was surprised by the poor writing quality of some SODA papers—several were basically not readable without a thorough understanding of referenced papers, and a substantial ability to infer what was meant rather than what was said. Some of these papers were accepted, which would have been impossible in a conference with double-blind reviewing.
  5. PC size The tradition in theory conferences is to have a relatively small program committee. I don’t see much advantage to this for SODA. The program committe is small enough and SODA is broad enough that it seems dubious to claim that every PC member is an expert on the subject of all of their papers. Also, (frankly) the highest quality reviews from my batch of papers weren’t written by me, but rather by reviewers that I picked who had the time to really grind through all the nitty-gritty of the paper. It’s easy to imagine that a larger PC would improve reviewing quality by avoiding overload.
Posted on

How do we get weak action dependence for learning with partial observations?

This post is about contextual bandit problems where, repeatedly:

  1. The world chooses features x and rewards for each action r1,…,rk then announces the features x (but not the rewards).
  2. A policy chooses an action a.
  3. The world announces the reward ra

The goal in these situations is to learn a policy which maximizes ra in expectation efficiently. I’m thinking about all situations which fit the above setting, whether they are drawn IID or adversarially from round to round and whether they involve past logged data or rapidly learning via interaction.

One common drawback of all algorithms for solving this setting, is that they have a poor dependence on the number of actions. For example if k is the number of actions, EXP4 (page 66) has a dependence on k0.5, epoch-greedy (and the simpler epsilon greedy) have a dependence on k1/3, and the offset tree has a dependence on k-1. These results aren’t directly comparable because different things are being analyzed. The fact that all analyses have poor dependence on k is troublesome. The lower bounds in the EXP4 paper and the Offset Tree paper demonstrate that this isn’t a matter of lazy proof writing or a poor choice of algorithms: it’s essential to the nature of the problem.

In supervised learning, it’s typical to get no dependence or very weak dependence on the number of actions/choices/labels. For example, if we do empirical risk minimization over a finite hypothesis space H, the dependence is at most ln |H| using an Occam’s Razor bound. Similarly, the PECOC algorithm (page 12) has dependence bounded by a constant. This kind of dependence is great for the feasibility of machine learning: it means that we can hope to tackle seemingly difficult problems.

Why is there such a large contrast between these settings? At the level of this discussion, they differ only in step 3, where for supervised learning, all of the rewards are revealed instead of just one.

One of the intuitions you develop after working with supervised learning is that holistic information is often better. As an example, given a choice between labeling the same point multiple times (perhaps revealing and correcting noise) or labeling other points once, an algorithm with labels other points typically exists and typically yields as good or better performance in theory and in practice. This appears untrue when we have only partial observations.

For example, consider the following problem(*): “Find an action with average reward greater than 0.5 with probability at least 0.99” and consider two algorithms:

  1. Sample actions at random until we can prove (via Hoeffding bounds) that one of them has large reward.
  2. Pick an action at random, sample it 100 times, and if we can prove (via a Hoeffding bound) that it has large average reward return it, otherwise pick another action randomly and repeat.

When there are 1010 actions and 109 of them have average reward 0.6, it’s easy to prove that algorithm 2 is much better than algorithm 1.

Lower bounds for the partial observation settings imply that more tractable algorithms only exist under additional assumptions. Two papers which do this without context features are:

  1. Robert Kleinberg, Aleksandrs Slivkins, and Eli Upfal. Multi-armed bandit problems in metric spaces, STOC 2008. Here the idea is that you have access to a covering oracle on the actions where actions with similar average rewards cover each other.
  2. Deepak Agarwal, , and Deepayan Chakrabati, Multi-armed Bandit Problems with Dependent Arms, ICML 2007. Here the idea is that the values of actions are generated recursively, preserving structure through the recursion.

Basic questions: Are there other kinds of natural structure which allows a good dependence on the total number of actions? Can these kinds of structures be extended to the setting with features? (Which seems essential for real applications.)

(*) Developed in discussion with Yisong Yue and Bobby Kleinberg.