SC²S Colloquium - June 22, 2016
|Date:||June 22, 2016|
|Time:||3:00 pm, s.t.|
Sebastian Kreisel: Spatial Refinement for Sparse Grid Classifiers
In this thesis refinement strategies specific for the machine learning task classification are discussed. Although general strategies, such as surplus refinement, give good results, specialized strategies might offer improvements. For classification, areas where refinement is necessary are usually given by boundaries between multiple classes. Using a probability density estimation approach, these boundaries are indicated by intersections between probability density functions of different classes. The first strategy proposed in this thesis is based directly on these intersections. A second approach uses a sample of data points in order to locate areas to refine in. Lastly, a strategy based on grid points alone is examined using only the information provided by the approximated probability density functions. These three strategies are then evaluated using both artificial and real world data sets and compared to established refinement strategies.
Johann Maier: Implementation and Evaluation of Online-Classification Methods based on Sparse Grids
In this thesis, i want to implement and evaluate online classification methods based on sparse grids combined with stochastic gradient descent, density estimation and support vector machines. The implementation will be done in C++ and the algorithms should be integrated into the SG++ framework. Furthermore, i want to study convergence of the different methods, investigate grid refinement and develop a new refinement strategy using information gain criteria, which are commonly applied for decision tree construction.