Seminar talks - ETH Zurich
As part of my master's studies at ETH Zurich I had to present somewhat recent research material in front of an audience. The slides of these talks are available here.Variational Methods for Graphical Models
A small presentation I prepared for the causality reading group. The presentation and summary are heavily based on the paper "An Introduction to Variational Methods for Graphical Models" by Jordan et al.
Feature Selection: Bayesian Support Vector Machines
A presentation which I've done for the course Feature Extraction. It is about a method how to combine Support Vector Machines and Bayesian Learning to do Feature Ranking/Selection. Which in fact boils down to defining a special trigonometric loss function and a regularizer which includes ARD (Automatic Relevance Determination) parameters.
Latent Dirichlet Allocation
My seminar talk, in December 2005, was about the paper "Latent Dirichlet Allocation", which is written by Blei, Ng and Jordan. Below you can find the slides of my talk.
LDA is a generative probabilistic model for discrete data, such as text corpora. It is an extension of the pLSI, proposed by Hoffman, and introduces the concept of a Dirichlet prior on the document level.
Keywords: LDA, topic, word, document, Dirichlet, mixture model, expectation maximization, variational inference.
- Slides (PDF, 1.1MB)
On Obfuscating point functions
My seminar talk, in April 2005, was about the paper "On Obfuscating point functions", which is written by Hoeteck Wee and was presented at Eurocrypt 2005. Below you can find the abstract and the slides of my talk.
In this week's lecture we will re-examine the question whether the obfuscation of a given program is (somewhat) feasible. But instead of trying to obfuscate the general class of programs, we will focus on programs that test for equality; so called point functions.
We will see that a non-standard one-way assumption enables us to construct a weak obfuscator for point functions. This result is astouning as it's the first known construction of obfuscators for a non-trivial family of functions under general computational assumptions.
- Slides (PDF, 539KB)