Patrick's Blog

I uploaded a graph cut MEX wrapper to the software section of this website.

I’m currently in Boston and will stay here till the end of the week. Next week, I’ll be attending the Snowbird Learning Workshop and AISTATS in Clearwater Beach, Florida. If you want to meet somewhere on the way, drop me a line!

I just released the first publicly available version of stroll, a library for structured output learning. stroll amongst other training algorithms also includes the spanning tree algorithms that I’ll be presenting in two weeks at AISTATS 09 in Clearwater, Florida. To learn more about the software, see the project page.

Today, I gave a small presentation to the first year students here at ETH to tell them why I personally think that learning about math in the first years is important. I decided to upload the slides, you can find them here; if this is helpful in order to get at least one person interested in calculus, algebra etc. than my goal is achieved! The slides are in German.

I really like books with great illustrations (great here means that the font should be the same as in the text etc, probably the best example of such a book is the “The Art of Computer Programming” by D. Knuth). I thus hardly ever felt happy with drawing things in a vector graphics program and always chose to program/typeset the graphics in some sort of programming language. In the past years I’ve played around with different packages/programs for doing this; the list includes:

• Matlab. I like it a lot for prototyping, but the plots usually look crappy. Same is true for the open source alternative octave.
• Gnuplot. Very powerful but the standard outputs usually look really bad.
• PSTricks. Some very impressive examples, but sucks when used in combination with PDF.
• PyX. From the website: “PyX is a Python package for the creation of PostScript and PDF files.” While I liked the Python environment and the power it gives you, it lead to problems when I wanted to include the plots in presentations: the fonts don’t agree or it’s harder to draw edges to other elements on the slides. It just lacked the integration with LaTeX. Note: it’s possible to work around some of the issues like for example the font problem, but usually it feels quite hacky.
• PGF/TikZ. As it’s from the same author who wrote the excellent Beamer package for presentations it integrates nicely with LaTeX and Beamer. It’s supposed to work with both DVI/PS and PDF. They recently released version 2 and by looking and the documentation you can immediately realize the sheer power this package has, especially for drawing general illustrations. It became my tool of choice about a year ago for drawing graphical models and such like, I however didn’t like the plot environment so much (it’s hardly existing at all).

This weekend I then finally played around with the Gnuplot TikZ terminal which addresses my last issue with PGF/TikZ: plotting functions and data. With this tool you can use gnuplot to create the actual plotting scripts and also to create the illustrations. However, instead of producing for example a PDF, the terminal outputs a .tex file that can then be included in the LaTeX document. The actual drawing in the document is then done by PGF/TikZ. The results look truly awesome. The installation is still a little bit messy as it needs the yet unreleased gnuplot 4.3 CVS.

A small presentation (together with a really short summary of the topic) is now available from my webpage under the section variational methods.

My semester thesis in Machine Learning is finally finished! The main part of the semesterthesis deals with the question, how to validate protein database searches. Beside the whole database search we evaluated several validation methods: hypergeometric p-value, Linear Discriminant Analysis and Gaussian Mixture Model.

See here for more details. There you can download the project documentation and the slides of the presentation.

The slides of my talk about Bayesian Support Vector Machines are now online: slides, handout.

To learn more about the topic, without actually reading the slides, you might find this useful

The slides of my talk about the latent Dirichlet allocation, a paper written by Blei, Ng and Jordan are now online, see here.