SCCS Colloquium - Oct 31, 2019
|Date:||October 31, 2019|
|Time:||15:00 - 16:00|
Moritz Spielvogel: Active Learning using Uncertainty Quantification
Deep Neural Networks need in general big amounts of data to get state-of-the-art results. Depending on the task, it may be very expensive to acquire labeled data whereas it may be easy to get big amounts of unlabeled data.
Active Learning aims to construct the most label-efficient data set by sampling the most informative or representative queries to be labeled by an oracle.
In this thesis there are different query methods presented in order to acquire the information gain of the model by an instance or to acquire the knowledge about an instance's representativenes about the whole data set in combination with the already labeled data set.
Keywords: Active Learning, Bayesian Neural Networks, Uncertainty Quantification
Dmitrij Boschko: Generalization and Parallelization of Sherman- Morrison System Matrix Updates for Sparse Grid Density Estimation
IDP submission talk. Dmitrij is advised by Kilian Röhner.
Sherman-Morrison rank-one updates have been used successfully for adaptive sparse grid density estimation. This allowed for regularization and adaptivity, but until now, this has only been possible in the offline/online splitting context using an orthogonal decomposition, such as tridiagonal. This IDP generalizes the old version of the Sherman-Morrison formula based sparse grid density estimation by using the Sherman-Morrison-Woodbury (SMW) formula. This allows for arbitrary decompositions in the offline phase and better supports parallelization by allowing for simultaneous adaptivity operations (rank-k updates), while keeping all features, such as regularization in between off/on phases. A parallelized/distributed version of the new SMW approach has been implemented and evaluated against the old version, and shows enhancements in terms of speed, while keeping the accuracy.
Johannes Stubenrauch: Boosting the Runtime Performance of a FEM Solver for Turbine Simulation
This master thesis provides a detailed performance analysis and enhancement of the open source finite element method solver Calculix used in production for turbine simulation and other problem domains. While already providing sufficient precision and correctness of results, the current implementation is not competitive with proprietary products regarding run-time performance, which effectively limits the simulation capabilities to smaller systems.
To this end, two complementary strategies for boosting performance are being evaluated to come up with a fine-grained design for an optimized implementation. Firstly, different methods of code transformations such as CPU/GPU parallelization and vectorization, language migration and others are investigated. This includes evaluating the usage of state-of-the-art libraries for numerical mathematics. Secondly, alternative algorithms or problem formulations with better theoretical execution time or complexity are evaluated. Thereby these are compared to the original algorithms to ensure stable or better quality of simulation results. The most promising code transformations are applied to the code for estimating the theoretical performance boost in real-world testing scenarios. In the final optimization design, a solution that is not bound to specialized hardware (such as FPGA, ASIC) is favoured to make the solution accessible to a wide audience.
Keywords: FEM, CUDA, Direct Solver, Indirect Solver, Finite Element