Machine Learning

PhD course, 2013

News

• 130317: The project proposal presentations will take place on Friday 10-12 in Signalen. Recall that the project proposals are due on March 20. Details regarding the projects are available here.
• 130317: The exam period has been prolonged with one extra week. The final deadline is now April 26, 2013.
• 130225: Congratulations to Manon and Jens who shared first place in SVN competition.
• 130213: Do not forget the SVM-classification competition for next lecture. Send you solution and a clear presentation of your classification result on test data by e-mail before next lecture. The task is available here.
• 130208: An illustration of back propagation is available here.
• 130131: BoB just pointed me to Dilbert's take on Machine Learning. Have a look here.
• 130130: During the lecture today I made a misstake in deriving the least squares classifier. The easiest solution is indeed what Hanna suggested, which is simply to add the trace I forgot and make use of a different trace rule instead. The correct derivation is available here.
• 130124: During the lecture yesterday we spoke briefly about l_p-regularized least squares. We mainly talked about the convex problems arising from p=1 and p=2, but there was also a very good question concerning the possible use of for example the l_0 norm. Here it is important to note that p=1 is a convex problem, p<1 leads to nonconvex problems. An obvious strategy is to start with norm p=1 and then initialize the nonconvex problem for the p<1 norm in the solution of the p=1 norm. Click here for some results clearly indicating that it can be interesting to consider making use of l_p-norms with p<1.
• 130124: During lecture 2 yesterday I mentioned a new book on the interplay between optimization and machine learning. Here is a link to the book as promised.
• 130117: During the lecture yesterday I mentioned how to identify a linear dynamical system using conjugate priors. The details are available here (the Gibbs sampler is explained during lecture 10)

  Adrian Wills, Thomas B. Schön, Fredrik Lindsten and Brett Ninness, Estimation of Linear Systems using a Gibbs Sampler. Proceedings of the 16th IFAC Symposium on System Identification (SYSID), Brussels, Belgium, July 2012. [pdf]

• 130114: The project page has now been updated.
• 121204: Fredrik Lindsten has agreed to deliver a bonus lecture on Bayesian nonparametric (BNP) models. This lecture will provide an introduction to an interesting, timely and powerful class of models.
• 121128: The 2013 home page is now available.

Informationsansvarig: Thomas Schön
Senast uppdaterad: 2013-04-02