SCCS Colloquium - Aug 7, 2019
|Date:||August 7, 2019|
|Time:||15:00 - 15:30|
Samuel Weber: Exploiting the Data Hierarchy with Geometry Aware Sparse Grids for Image Classification
This is a Bachelor's thesis submission talk. Samuel is advised by Kilian Röhner.
In this thesis the concept of a data hierarchy and stencils that can exploit the data hierarchy are presented. Through establishing the data hierarchy, a set of different strong coarsened images of the original data, the image classification capabilities of geometry aware sparse grids are improved. Complex stencils can be applied on lower resolutions, that otherwise would be infeasible on the original resolution, while still being able to apply simpler stencils to the original data. Also, the new class of hierarchical parent stencils is defined. These stencils work vertically on the data hierarchy, using in some cases, less interaction terms than the direct-neighbor-stencil with a similar classification accuracy. A smart use of the data hierarchy and multiple stencils allows us to apply sparse grid image classification in cases that would have been infeasible before.
Keywords: Geometry-aware sparse grid, image classification, data hierarchy