SC²S Colloquium - December 13, 2012
From Sccswiki
Date: | November 28, 2012 |
Room: | 02.07.023 |
Time: | 3 pm, s.t. |
Christian Kirst: Transcription of Polyphonic Piano Music with Non-Negative Matrix Factorization and Support Vector Machines
This work suggests a framework for polyphonic piano music transcription. The algorithm applies a supervised Non-Negative Matrix Factorization (NMF) to the spectrogram and maps spectral activations to pitch activations. In order to classify notes features are computed based on these pitch activations. Support-Vector-Machines (SVM) then detect note onsets. Finally offsets for each onset are computed by a rather simple classification algorithm. Overall the framework shows that a NMF-based transcription framework is able to achieve state of the art accuracy.
Florian Zacherl: Using pre-computed matrices in a density-based classification method with sparse grids
Classification of huge data sets is a process that requires long computation times, especially with high-dimensional data. A way to reduce this is to use function approximation on a sparse grid. Still this might not be enough for big data sets. This IDP examines a further optimization by using a density-based approach that allows to do a relevant part of the work in advance without the actual data. More precisely, the solving of a linear system of equations is simplified by pre-computing, decomposing and storing the left hand side matrix of the system.
Philipp Müller, Stjepan Bakrac: Discontinuous Galerkin Methods for Shallow Water Equations
For simulations based on the Shallow Water equations (SWE) as they are used for tsunami simulations, the use of Discontinuous Galerkin schemes with higher-order basis functions received growing attention during the last years. This project aims to examine differences between integrals computed analytically and by numerical quadrature.