| Date: |
March 16, 2016
|
| Room: |
02.07.023
|
| Time: |
3:00 pm, s.t.
|
Gunnar König: Online Classification with Adaptive Sparse Grid Kernels
Sparse grids and Support Vector Machines have both been applied successfully to
classication problems. In this thesis, the two methods are combined: Sparse grids are
used to construct a new kernel that can be deployed as a nonlinear similarity measure for
Support Vector Machines.We will show, that training a Support Vector Machine equipped
with a sparse grid kernel indirectly tunes the parameters of the underlying grid. Thus, it
is possible to optimize a sparse grid using online algorithms that were orignally developed for
Support Vector Machines. More specically we will examine how online Support Vector
Machine learning algorithms can be used to train a sparse grid to seperate the classes of
a dataset. Research on sparse grids for classication has shown that it does not suce to
tune the grid's surplus vector in order to achieve good performance in terms of accuracy
and efficiency. Moreover it is crucuial to adapting the structure of the grid as well to t
the actual problem instance. Hence I propose a new method that allows to rene sparse
grid kernels eciently in the context of online Support Vector learning. We will see, that
the resulting algorithm achieves good results on benchmark datasets.