wshxzt的北美生活

I've just come back from MLSS 2008. Generally speaking, it was quite good. The summer school invovled several topics, including Kernel methods and SVM, Sampling methods and MCMC, Reinforcement Learning, Graphical Models, Datamining, Vision, Game Theory and Online Learning. Videos are available at http://videolectures.net/mlss08au_kioloa/.

In my biased eyes, the most excellent talk was Sampling methods given by Nando de Freitas. He introduced rejection sampling, importance sampling and MCMC in the first day; and in the second day, he briefly introduced the model selection, MDP and POMDP. He gave many samples to make things easier to be understood. For me, the most interesting part is how to estimate the expection of features in graphical models. But I'm still wandering whether q(x) in MCMC could always be good enough that MCMC would be more efficient than exact inference. Anyway, this video lecture is quite worth seeing.

Another interesting talk was reinforcement learning given by Csaba Szepesvári from reinforcement learning school. I almost knew nothing about reinforcement learning before, so this talk was tough for me. However, i feel reinforce learning is a nature way that agents learn this world. If I knew more about the math theorems RL used, I would learn more in the lecture as I got stuck when so many symbols and theorems were presented in the begining of the lecture.

Avrim Blum's problems on game theory were quite interesting. He viewed reward of different actions taken by two players as a matrix, and tried to compute the expection of two players in different problems. And he discussed the existence of Nash equilibria. Though this is a little bit different from machine learning concerns, it is still worthwhile to think about these problems. The second part of his talk was about similarity functions that are not legal kernel functions (not semi-positive or even not symmetric), and use these functions to do clustering.

Other talks are also very good, but the topics of these talks are so pupolar that I guess almost everyone know the idea of them.

Finally, I'd thanks Marcus and all organizer for setting up this summer school and making everything possible. This was the greatest experience I ever had!