Edited by author 10-23-2001 03:24 AM
It said that in the paper on page 6 that pseudo-loss (AdaBoost.M2) and error (AdaBoost.M1) were identical when applying them on two-class problems. It also said that on page 3 that the main disadvantage of AdaBoost.M1 was that it is unable to handle weak hypotheses with error greater than 1/2. My question is what improvement can you make in the two class case when the weak hypotheses have error greater than 1/2. You would not be able to use pseudo-loss, or AdaBoost.M2, since it would just be identical to error, or AdaBoost.M1. I suppose in practice it is typical for weak hypotheses to be less than 1/2 in the two class case, making it unnecessary to think about that problem anyway.

I think that they maybe should omitted the second part of the paper, or the part where they talk about the performance of using a nearest-neighbor classifier on an OCR problem. I agree with Junwen that they should have expanded on this idea to include a discussion about how it could be used to improve SVMs and other related algorithms, even though SVMs at the time were not very well-known and understood. But I do think they should have limited the paper just to their first discussion and analysis. So perhaps this paper should be divided into two separate papers.

What interested me about the paper was the differentiation of plausiblity and probability. I would be curious to hear an explanation of this distinction in tomorrow's presentation. Overall, I thought the paper was good and the idea of pseudo-loss is neat.