Thursday, July 29, 2010

Trying to make regression chapter comprehensible

I'm now going back over the regression chapter trying to make it a little more comprehensible. I've swapped the main application to Agarwal and Triggs body pose estimation. Every chapter can't be about faces.

I'm not super happy with the notation yet. There's a lot of maths and it could still be quite a lot clearer. Should get onto non-linear regression tomorrow and hopefully polish off the relevance vector machine on Saturday morning.

Tuesday, July 27, 2010


I'm still working on the regression chapter. I've been writing about relevance vector regression. It's taken me a while to figure it all out and even now, I think the notation is going to have to be substantially overhauled.

Thursday, July 22, 2010


I've been sketching out the regression chapter. I'm going to cover linear regression,
Bayesian linear regression (i.e. GPs), kernel methods and the relevance vector machine. I'm still hunting around for some good applications to illustrate this all with.

Monday, July 19, 2010


I've been away from U of T for some time so I haven't updated the online copy until today. There have been a lot of changes in the mean time. I've finished the first full draft of the chapter on complex probability distributions (EM Algorithm, factor analysis, MoG and the like) and I'm now working on the chapters on regression and classification. I've also sorted out a whole bunch more mistakes from the first few chapters.

Wednesday, July 7, 2010

Generative models

Still working on the generative models chapter - made a lot of small updates to the sections on t-distributions and factor analysis. I've managed to cut out about a page of irrelevant information along the way. Less is more...

Tuesday, July 6, 2010


Back from vacation... and back to working on the generative models chapter. Spent most of today trying to make the section on Mixtures of Gaussians somewhat more comprehensible. I think its gradually getting better. I'm slightly frustrated by the fact that I can't really explain the EM algorithm properly without ruining the flow of the chapter. The result is that the description of the fitting algorithm for mixtures of Gaussians is a bit unsatisfying. But I think the alternative is worse...