Invited Session Fri.3.MA 376

Friday, 15:15 - 16:45 h, Room: MA 376

Cluster 12: Life sciences & healthcare [...]

Methods from discrete mathematics in systems biology


Chair: Utz-Uwe Haus



Friday, 15:15 - 15:40 h, Room: MA 376, Talk 1

Stefanie Jegelka
On fast approximate submodular minimization and related problems

Coauthors: Jeff A. Bilmes, Hui Lin


Machine learning problems often involve very large data sets. To test algorithms quickly, we aim to extract a suitable subset of a large training corpus. This is a submodular minimization problem, but the size of the data renders current exact methods very impractical. Graph cuts can be an alternative, but may not be able to efficiently represent any submodular function. We therefore approximate the objective function by a sequence of graph-representable functions. This leads to an efficient approximate minimization algorithm. It turns out that the underlying model not only helps represent submodular functions, it also enhances applications of graph cuts in computer vision, representing non-submodular energy functions that improve image segmentation results.



Friday, 15:45 - 16:10 h, Room: MA 376, Talk 2

Edward Roualdes
Non-parametric species delimitation based on branching rates

Coauthors: David Weisrock, Ruriko Yoshida


Many probabilistic tests have been developed to delimit species based on the coalescent model. These computational efforts rely primarily on parametric models that try to account for known variance found in genetic processes. Unfortunately, this variance is difficult to model precisely. Using non-parametric tests, we develop a method to delineate species by estimating the time species change from growth (e.g., Yule models) to a coalescence process without constraining the processes to a particular model. Using simulated gene trees from a known species tree, we compare our non-parametric method to established parametric methods.



Friday, 16:15 - 16:40 h, Room: MA 376, Talk 3

Giovanni Felici
Logic data mining in the presence of noisy data

Coauthor: Emanuel Weitschek


In this work we consider a method for the extraction of knowledge from data. The knowledge is represented as disjunctive normal form (DNF) logic formulas that identify with high precision subsets of the training data. The method is mainly designed for classification purposes, but can be profitably deployed for information compression and data analysis in general. It is based on three main steps: discretization, feature selection and formula extraction. For each step, a mathematical optimization problem is formulated and solved with ad hoc algorithmic strategies.
The method is designed to perform exact separation of training data, and can thus be exposed to overfitting when a significant amount of noise is present in the available information. We analyze the main problems that arise when this method deals with noisy data and propose extensions to the discretization, feature selection and formula extraction steps; we motivate these extensions from a theoretical standpoint, and show with experimental evidence how they operate to remove the effect of noise on the mining process.


  Illinois Payday Loans should not be obtained by people who do not have the capacity to repay the lenders. This is a permit to the world of pleasure and the lasting sex. Cialis Super Active online is a drug, the quality level of which is determined by its action speed.