The 2006 Machine Learning Summer School in Taipei, Taiwan ended on August 4, 2006. It has been a very exciting two weeks for a record crowd of 245 participants (including speakers and organizers) from 18 countries. We had a lineup of speakers that is hard to match up for other similar events (see our WIKI for more information). With this lineup, it is difficult for us as organizers to screw it up too bad. Also, since we have pretty good infrastructure for international meetings and experienced staff at NTUST and Academia Sinica, plus the reputation established by previous MLSS series, it was relatively easy for us to attract registrations and simply enjoyed this two-week long party of machine learning.
In the end of MLSS we distributed a survey form for participants to fill in. I will report what we found from this survey, together with the registration data and word-of-mouth from participants.
The first question is designed to find out how our participants learned about MLSS 2006 Taipei. Unfortunately, most of the participants learned about MLSS from their advisors and it is difficult for us to track how their advisors learned about MLSS. But it appears that posters at ICASSP (related but not directly related to ML) work better than at NIPS (directly related to ML).
Question 2 to 6 ask participants their demographical background. They are mostly graduate students working on one of the application fields of machine learning (e.g., bioinformatics, multimedia, NLP processing). Asked about why they attended MLSS, as expected, about 2/3 replied that they wanted to use ML and 1/3 replied that they wanted to do ML research. Most of participants attended all talks, which is consistent with our record. The attendance rate was kept at about 80 percent every day, a remarkable record for both speakers and participants. Asked about what makes it difficult for them to understand the talks, about half replied mathematics, about a quarter replied “no examples” and less than a quarter replied English. Finally, all talk topics were mentioned as being helpful by our participants, especially those talks that are of more introductory nature, such as graphical models by Sam Roweis, SVM by Chih-Jen Lin, and Boosting by Gunnar Ratsch, while talks with many theorems and proofs are less popular.
We let participants provide their suggestions to us. An issue that was brought up very often is that many wanted lecture notes to be distributed in advance, maybe a day before the talks, maybe it would be better if before MLSS. One of them suggested we put prerequisite math background on the Web so that they would have been better prepared (but that may scare away many people, not good for organizers…). A quick fix for this problem is to provide Web pointers to previous MLSS slides and video and urge registered participants to take a look at them in advance to prepare themselves before attending MLSS.
Another frequently brought-up issue is that they would like to hear more concrete examples, have a chance to do some exercises, and learn more about the applications of the algorithms and analysis results given in the talks. This is reasonable given that 2/3 of our participants’ goal was learning how to use ML. So Manfred decided to adapt to their needs “online” by modifying his talk slides over night (thanks!).
Then our participants would like organizers to design more activities to encourage interaction with speakers and among participants. I think we could have done a better job here. We could have let our speakers “expose” to participants more often than staying in a cozy VIP lounge. We could have also provided online and physical chat board for participants to expose their contact IDs.
We scheduled our talks mostly based on the time constraints given by the speakers. Roughly, it was like we made graphical models/learning reduction first, SVM and kernels next, then online/boosting/bounds, and finally clustering. It turned out that our speakers were so good that they covered and adapted to others related talks and made the entire program appear like a carefully designed coherent one. So most participants liked the program and only one complaint was about this part.
Putting all of the comments together, I think we have two clusters of participants that are hard to please at the same time. One cluster of participants is looking for new research topics on ML or trying to enhance their understanding of some advanced topics on ML. If MLSS is designed for them, speakers can present their latest or even ongoing research results. Speakers can take the chance to spread their idea in the hope that the idea can be further explored by more researchers. The other cluster of participants is new to ML. They might be trying to learn how to use ML to solve their problem, or just trying to have a better idea of what ML is all about. To them, speakers need to present more examples, show them applications, and present mature results. It is still possible to please both of them. We tried to balance their needs by plugging in research talks every afternoon. Again our speakers did a great job here and it seemed work pretty well. We also designed a graduate credit program to give registered students a preview and prerequisite math background. Unfortunately, due to long beauraucratic process, we could not announce it early enough to accommodate more students and the program did not accept international students. I think we could have done a better job helping our participants understand the nature of the summer school and be prepared.
Finally, on behalf of the steering committee, I would like to take this chance to thank Alex Smola, Bernhard Scholkoph and John Langford for their help to put together this excellent lineup of speakers and the positive examples they established in previous MLSS series for us to learn from.