Edited by author 10-22-2001 09:40 PM
First, I'd like to make a couple of specific comments about the paper(s). I found the slight tweak in the M2 version where the algorithm outputs a simple set very interesting. Intuitively, this seems like a good idea and experimentally, it seems to work better. However, I did find the concept of pseudo-loss a bit confusing (hopefully, this confusion will be resolved in the presentation :).

I also have a few general comments about boosting in general that the paper somewhat hinted at. One of the amazing things about AdaBoost (and boosting in general) is that it is applicable in such a broad range of applications. And to boosting's credit, the performance is usually impressive.

Another interesting thing that is somewhat discussed in the paper is that the contributions of boosting are still a bit of a mystery. I think more and more research is helping to isolate the exact effects of boosting, but I still don't think the whole picture is perfectly understood.