ARCHIVE Snapshot of the Rätsch lab website (Tübingen, then New York) as of February 2016: preserved, not maintained. The group today: bmi.inf.ethz.ch
Personal tools
Home → Rätsch Lab → People → Research Topics → Machine Learning
 
Document Actions

Machine Learning

Machine Learning offers tools for computational analyses of complex phenomena. Given a set of labeled observations, one trains a learning machine to be able to make accurate predictions about future observations.

For instance, given a set of known splice sites in the genome, a learning machine can recognize the location of other such sites. Machine Learning techniques detect and exploit statistical regularities that are hidden in existing observations without requiring an explicit model of the problem. Such data-driven approaches excel in situations where a detailed understanding of the underlying biological mechanisms is lacking but data is abundant, which is frequently the case in Genome Biology.

The group has been developing kernel-based algorithms capable of dealing with heterogeneous and structured input data derived, in particular, from biological sequences. Specifically, we have developed multiple kernel learning algorithms that allow the optimal combination of heterogeneous data through an elegant and theoretically sound framework. Additionally, we have developed very efficient string data structures that allow us, in combination with innovative SVM optimization techniques, to solve sequence analysis tasks with up to 50 million training examples and other tasks with 7 billion test examples. With these new developments we have enabled large-scale sequence analyses and established state-of-the-art learning techniques in the field of genomic sequence analysis. The application of these techniques has led to the most accurate genomic signal detectors that currently exist.

An important limitation of kernel methods is that the resulting prediction functions are difficult to interpret and thus to translate into biologically meaningful insights. For specific sequence kernels, we have therefore developed methods capable of elucidating the learning result. Taking the correlation structure of the features into account, importance scores are computed for each relevant feature.

The group has been developing suitable learning techniques that allow us to predict structured outputs on genomic sequences and, in particular, segmentations thereof. These techniques are used in various projects related to gene structure prediction, the analysis of transcriptional activity, and the identification of genomic polymorphisms (see Computational Genome Annotation and Computational Transcriptomics).

Research Projects

Large Scale Learning with String Kernels
The goal of this project is to develop new sequence kernels that are particularly well suited for sequence analysis tasks as well as algorithms that efficiently compute the optimal SVM solution when using these kernels.
Boosting
We consider so-called boosting algorithms which combine hypotheses generated by another learning algorithm to accurately solve two-class classification tasks. We are interested in their convergence properties as well as their accuracy in the presence of noise in the training data.
Efficient Algorithms for Structured Output Learning
We work on the development of novel inference algorithms for predicting structured outputs, such as gene structures, RNA secondary structures, etc. We are particularly interested in designing efficient learning methods that can be used for problems arising in computational biology.
Extracting Discriminative Motifs Based on SVM Detectors
At the heart of many important bioinformatics problems, such as gene finding and function prediction, is the classification of biological sequences. Frequently the most accurate classifiers are obtained by training support vector machines (SVMs) with complex sequence kernels. However, a cumbersome shortcoming of SVMs is that their learned decision rules are very hard to understand for humans and cannot easily be related to biological facts. In this project we aim to develop new techniques to improve the interpretability of predictive methods including SVMs.
Domain Adaption and Multi-task Learning

People specializing in this area

Graduate Students

Nico Görnitz

Nico Görnitz

More about Nico Görnitz...

Publications