Prof. Dr. Ingo Steinwart
(Localized) learning with kernels
Was |
|
---|---|
Wann |
16.06.2017 von 12:00 bis 13:00 |
Wo | Eckerstraße 1, Raum 404, 4. OG |
Termin übernehmen |
vCal iCal |
Using reproducing kernel Hilbert spaces in non-parametric approaches
for regression and classification has a long-standing history. In
the first part of this talk, I will introduce these kernel-based
learning
(KBL) methods and discuss some existing statistical guarantees for
them.
In the second part I will present a localization approach that
addresses
the super-linear computational requirements of KBLs in terms of the
number
of training samples. I will further provide a statistical analysis
that
shows that the "local KBL" achieves the same learning rates as the
original,
global KBL. Furthermore, I will report from some large scale
experiments
showing that the local KBL achieves essentially the same test
performance
as the global KBL, but for a fraction of the computational
requirements.
In addition, it turns out that the computational requirements for
the local
KBLs are similar to those of a vanilla random chunk approach, while
the
achieved test errors are in most cases significantly better.
Finally, if time
permits, I will briefly explain, how these methods are being made
available
in a recent software package.