Doktoranden-/Diplomandenseminar SS08</h2>

This is an archive of the talks in the summer term 2008. The current schedule can be found here.

Monday, 22.09.08, 14.00, FR6046

Martin Slawski, LMU Munich

Feature selection and regularization for linear hypotheses formed of structured predictors


Linear hypotheses still are - despite (or possibly just because of) their simplicity - a standard instrument for handling regression- and classification problems. In situations characterized by an unfavourable ratio of sample size to the number of predictors/features, as occurring, e.g., in signal regression (Frank and Friedman [3]) or in high-throughput biological data, inference for the model parameters is performed subject to constraints reflecting prior assumptions. A common instance thereof is sparsity, which is realized by imposing an L1-constraint Tibshirani [6]. Various kinds of prior knowledge about the association structure of the predictors is represented in terms of a graph. Its Laplacian (Chung [2]) generates the class of discrete Tikhonov-Arsenin regularizers (Smola and Kondor [5]) used for semi-supervised learning on labeled graphs (Zhou et al. [7], Belkin et al. [1]). Unifying both concepts yields a regularizer, whose practical use is demonstrated in the classification of accelerometer data, feature extraction for handwritten digit recognition and pathway reconstruction in the spirit of Li and Li [4].


[1] M. Belkin, I. Matveeva, and P. Nyogi. Regularization and semi-supervised learning on large graphs. 17th Annual Conference on Computational Learning Theory, 2004.

[2] F. Chung. Spectral Graph Theory. AMS Publications, 1997.

[3] I. Frank and J. Friedman. A statistical view of some chemometrics regression tools (with discussion). Technometrics, 35, 1993.

[4] C. Li and H. Li. Network-constrained regularization and variable selection for analysis of genomic data. Bioinformatics, 24, 2008.

[5] A. Smola and R. Kondor. Kernels and regularization on graphs. 16th Annual Conference on Learning Theory, 2003.

[6] R. Tibshirani. Regression shrinkage and variable selection via the lasso. Journal of the Royal Statistical Society Series B, 1996.

[7] D. Zhou, J. Huang, and B. Schölkopf. Learning from labeled directed and unlabeled data on a directed graph. 22nd International Conference on Machine Learning, 2005.

Tuesday, 16.09.08, FR 6046

Ryota Tomioka, Tokyo Tech

Conic optimization and sparse component learning


Lasso, group lasso, and trace norm (dual spectral norm) regularizations are useful tools for discovering discriminative components in the inputs through principled convex optimization problems. In this talk, I show that these three regularizers can be all considered as "trace regularization" in different cones, namely positive orthant, second order cone, and positive semidefinite cone, respectively. Using the "trace regularization" view, I develop an efficient subgradient based L-BFGS algorithm for group lasso problem with general differentiable loss function; the algorithm is an extension of Andrew & Gao 2007 ICML, in which they dealt with the lasso regularization and it can be further extended to the dual spectral norm regularization. The algorithm is applied to a multi-class classification problem in the context of P300 speller system which has 2368 input dimensions and 15300 training instances.

Wednesday, 23.07.08 11.00 FR6046

Kai Joshua Miller

Power law changes and their coupling to low frequency phase in electrocorticography


I will briefly review the hypothesis and validation of functional changes in a cortical spectral power law with local activity in human cortex. Using a PCA based method on electrocorticographic recordings in humans, we were able to decouple this power law behavior from the classic alpha and beta rhythms, revealing its presence at low frequencies. The projection of the dynamic spectrum to this power law, which we denote "chi" is able to capture the dynamics of specific finger movement in specific electrodes. We examined the relationship of changes in the amplitude of this power law to the phase of intrinsic low frequency rhythms. We will then demonstrate that chi couples to the phase of the beta rhythm (so called phase-amplitude coupling – PAC). During periods of movement, this PAC is less pronounced than during periods of rest.

We show how a simple, small-scale, model of synaptic organization may provide intuition for the large scale phase-amplitude correlation we report. In this model: 1) chi reflects asynchronous summation of a large number of cortico-cortical inputs between pyramidal neurons. 2) Synchronous, sub-cortical or distant cortical, projections to pyramidal neurons are reflected in the beta rhythms. The beta rhythm constrains local computation, and this is the basis for the phase-amplitude coupling. During local computation, the amplitude of chi goes up, beta goes down, and the PAC decreases.

Wednesday, 16.07.08 14.00 FR6046

Christian Gehl and Patrick Duessel canceled

Wednesday, 09.07.08 14.00 FR6046

Tammo Krüger

Intelligent Intrusion Prevention


Anomaly detection techniques have been shown to be suitable for network intrusion detection: based on an automatically generated model of normality incoming traffic can be seperated into anormal and normal traffic. The majority of proposed system, which are based on anomaly detection, focus on the network level and therefore lack the option of realtime reactive response mechanism.

Based on anomaly detection techniques developed in the MIND project the successor project !ReMIND aims among other things at using these methods for intelligent adaptive defense. The !ReMIND project proposes two intrusion prevention mechanisms based on known and mature technologies: packet filters and application level firewalls. By incorporating adaptable and self-learning methods from machine learning the static and hand-crafted rules of these systems can be made more reactive and intelligent responses to anomalous behaviour can be triggered. Furthermore by monitoring the system state valuable information can be gathered and a feedback loop to the learning system can be established. By intertwining learning methods with static intrusion prevention mechanims and host based sensors all parts of such a hybrid systems can take advantage of this mixture, which will be outlined in this talk.

Wednesday, 02.07.08 14.00 FR6046

Alexander Zien

Tutorial: multiple kernel learning (MKL) (replacement of Doktorandenseminar, of course optional)

It will be rather tutorial style, so especially if you're not familiar with the topics this may be a good opportunity to get into them and to get your questions answered.

Friday, 04.07.08 14.00 FIRST Adlershof

Alexander Zien

Tutorial: semi-supervised learning (SSL) (optional)

It will be rather tutorial style, and may even have some practical exercises (if I manage to prepare them in time and if you're interested in doing them; please bring your laptops). So especially if you're not familiar with the topics this may be a good opportunity to get into them and to get your questions answered.

Wednesday, 25.06.08 14.00 FR6046

Gabriele Schweikert

Gene-finding across organisms


In this talk I will give a brief overview over the projects that I worked on during my Ph.D. time. Specifically I will discuss two tasks in more detail: First I will introduce mGene, a novel, purely discriminative gene finding system. Here we have combined state-of-the-art predictions of functional sequence elements using SVMs with a label sequence learning algorithm to predict full gene structures. The superiority of our method has been demonstrated in a genome annotation competition, where we outperformed 15 other systems.

In the second part of the talk I will cover the problem of transferring information between organisms in the context of genomic sequence annotation. While this domain adaptation scenario is very common in computational biology, it has so far attracted little attention in this field. I will present an empirical analysis of several domain adaptation algorithms and examine the difference between the methods depending on the distance between the two distributions at hand and the amount of available training data.

Thursday, 26.06.08 14.00 FR6046

Tobias Lang

Learning to Think


Nowadays, intelligent robots are not even able to prepare a cup of tea. The major handicap to perform higher-level tasks is the lack of adequate knowledge representation and reasoning formalisms. In my research over the next years, I'm looking for learnable world models that apply to noisy domains with many unknown objects. Based on these models, I'm investigating flexible and efficient action planning algorithms which enable the robot to achieve its goals, such as preparing a cup of tea. In my talk, I'll discuss the essential requirements for reasoning and I'll present the ideas I'm currently exploring.

Thursday, June 12.06.08 10.00 AND 15.00 FR6046

10.00: Dr. Kai Schreiber, Neuroscientist at Rutgers University Newark, USA

What is the ideal stimulus for stereo vision?


The visual system reconstructs depth from subtle differences in the two retinal images. The simplified description of binocular retinal geometry, states that because the eyes are separated horizontally, retinal disparities are horizontal, too. I present experimental evidence showing that this can be no longer true when the eyes move. Through the concept of the extended point Horopter, the empirical two dimensional retinal correspondence pattern obtained in our experiments reveals information both about the surface of the ideal stimulus for stereo vision, and about its ideal location in head centered space. An old argument made by Helmholtz about the adaptive value of certain specifics of retinal correspondence is revisited, and its specifics are updated in light of our new experimental findings.

15.00: Dr. Christian Hesse from F.C. Donders Center for Cognitive Neuroimaging, Nijmegen NL

Brain-Computer Interface Research at the F.C. Donders Centre


This talk provides a brief overview of Brain-Computer Interface (BCI) research at the F.C. Donders Centre for Cognitive Neuroimaging (FCDC). We have developed a BCI system using multi-channel magnetoencephalogram (MEG) or electroencephalogram (EEG) signals for real-time control of a computer game and a robotic arm using a motor imagery paradigm. The FCDC BCI system is implemented in Matlab as part of the 'FieldTrip' open source toolbox and is based on a modular architecture which handles offline and online data streaming and event-based communication between any number of autonomous processing modules (agents) over which various signal analysis and classification tasks can be distributed. Our approach to BCI signal analysis so far has been to identify and extract a small number of underlying source signals whose time-frequency characteristics exhibit task-related modulation, and to subsequently combine these sources to form features or classifier ensembles. This approach lends itself to employing computationally efficient spatial filtering and time-frequency decomposition techniques for feature extraction and classification in the online case. On methodological aim of ongoing work is to automate as much as possible the identification of these ensembles of task-related sources, and how these can be robustly estimated in the first place.

Wednesday, 04.06.08 14.30 FR6046

Felix Biessmann

Combining pharmacological MRI and electrophysiology: problems and benefits

Pharmacological magnetic resonance imaging (phMRI) is a non-invasive tool for studying the modulatory effects of pharmacological agents on the large-scale brain networks that underlie cognition. However, the relation between these effects on functional imaging signals and the underlying neural activity is unclear. We have combined phMRI with electrophysiological recordings of neural activity to link effects at the level of imaging signals to those observed in electrical recordings from neuronal populations. This imposes additional demands on the data analysis: first of all, electrophysiological recordings have to be cleaned from artefacts arising from strong magnetic field gradients; secondly phMRI data analysis is a challenging task in itself due to the large dimensionality of the data; and finally electrophysiology and MRI describe brain activity on two distinct spatio-temporal scales. Our project focuses on employing machine learning methods in order to tackle those problems. The aim is to find analysis methods that make it possible to investigate the influence of pharmacological agents on brain activity in a unified framework.

Wednesday, 28.05.08 14.00 FR6046

Nikolay Jetchev

Topics: robot control, internal representations, grasping of a stick with a hand as a first example

Wednesday, 21.05.08 14.00 FR6046

Fouad Channir

Brain Computer Interface and Investigation of “resonance like” frequencies, using Somatosensory Steady State Evoked Potentials

Wednesday, 14.05.08 14.00 FR6046

Martijn Schreuder

The P300 response to auditory cues differing in spatial location

Most BCI research is conducted with the aim of applying the results at some stage to patient groups with severe motor disabilities and paralyzes. Especially for late-stage ALS patients with totally locked-in syndrome, such BCI would be of great interest. For several reasons, many of these patients have difficulty with using vision. Therefore, we investigate an auditory P300 BCI that may lead to a multiclass BCI.

Friday, 25.04.08 14.00 FR6046

Christine Steinhoff

Integrative Visualization of Array-CGH, Expression and Supplementary Variables

The intuitive way to analyze different omics data is 'separately and consecutively'; e.g. first determine regions with copy number aberrations and then look for differentially expressed genes. Only recently have integrative or joint approaches -where different data are fused before analysis- been published. Still, these approaches do not integrate interesting covariate data like tumor grading and staging. We established a new data analysis pipeline for: 1) joint unsupervised analysis of aCGH and microarray data 2) integration of supplementary categorical data. We have developed an analysis and visualization pipeline that comprises four parts: discretization, binary mapping, gene filtering and visualization with 'multiple correspondence analysis with supplementary variables' (MCASV). For some steps we considered different options (hyperparameters), e.g. we studied three microarray discretization procedures, corresponding to different biological objectives. The first two steps transform the three matrices (aCGH, microarray, covariates) to a common binary format. MCASV has been developed in the context of social sciences but to our knowledge has not been used in computational biology. The global binary matrix is projected onto a MCA bidimensional space, where vicinity between genes and covariates can be visualized and quantified. We demonstrate our visualization approach on a published breast cancer dataset (Pollack et al, 2002). Most importantly, the approach can easily be generalized to various kinds and numbers of input data.

Monday, 31.03.08 12.30 FR 5516 (!)

Gunnar Rätsch Friedrich Miescher Laboratory of the Max Planck Society, Tübingen

Machine Learning Methods for Discovering Common Sequence Variation in Arabidopsis thaliana

After giving an overview of the research performed in the group (Gene finding, tiling array analysis and large scale learning methods for sequence analysis), I will focus on two recently published papers (Zeller 2008, Clark et. al. 2007) on genome resequencing. The genomes of individuals from the same species vary in sequence as a result of different evolutionary processes. To examine the patterns of, and the forces shaping, sequence variation in Arabidopsis thaliana, we performed high-density array resequencing of 20 diverse strains (accessions). More than 1 million nonredundant single-nucleotide polymorphisms (SNPs) were identified at moderate false discovery rates (FDRs), and 4% of the genome was identified as being highly dissimilar or deleted relative to the reference genome sequence. With this technology, the detection rate for isolated SNPs is typically high. It is, however, greatly reduced when other polymorphisms are located near a SNP as multiple mismatches inhibit hybridization to arrayed oligonucleotides. Contiguous tracts of suppressed hybridization therefore typify polymorphic regions such as clusters of SNPs or deletions. We developed a machine learning method, designated margin-based Prediction of Polymorphic Regions (mPPR), to predict Polymorphic Regions (PRs) from resequencing array data. Conceptually similar to Hidden Markov Models, the method is trained with discriminative learning techniques related to Support Vector Machines, and accurately identifies even very short polymorphic tracts (<10 bp). Our predictions provide a valuable resource for evolutionary genetic and functional studies in A. thaliana, and our method is applicable to similar data sets in other species (currently done on rice and mouse).


-- Main/MikioBraun - 18 Nov 2008