I thought the introduction (section 5.1) to this chapter was good. They provided a good general idea of what statistical machine learning is. I particularly thought that they covered the ideas of overfitting and underfitting well. I think the introduction could have been made a little bit better with a few more concrete examples (rather than just seeing the problem of learning some function).

Also, I found the more technical sections slightly less approachable. I think the authors could have done a better job of defining the vocabulary and terms and, again, added some more less technical examples and explanations (althought this is just my bias since I lean towards applications rather than pure theory :)

Chapter 5 of Vapnick's book is much easier to read and understand after having had exposure to the previous chapters and perhaps more formal training (i.e, a class) on the subject. Even though the notion of shattering and related ideas have already been talked about in the seminar, the precise mathematical framework for discussing these ideas is needed before this chapter can be tackled.

I am very much interested in seeing a link between the theory in this chapter with the learning algorithms that apply it, like support vector machines for example. Some other questions I have are what are the theories that compete with Structural Risk minimization, and why do SVMs employ that as opposed to some other theory.

From a technical and conceptual point of view, I cannot discern between VC entropy and annealed entropy. Hopefully all of these things will be addressed tomorrow.

Perhaps this is easier to read afting having read the previous chapters - it's a bit difficult without. They do a passable job of explaining things up until the union bound, but i'm still having trouble with that. I hope that's on the agenda for the presentation.