Urs Köster
Researcher, Neuroinfomatics Group
HIIT Basic Research Unit
Research interests
I am studying statistical models of natural images. Following the hypothesis that the visual cortex is matched to the statistics of the data it receives, I am trying to model properties of neurons in primary visual cortex. Using unsupervised learning techinques, based on ICA and other statistical methods, I am working towards biologically plausible multi-layer network models. These models are capable of forming powerful internal representations of the training data, allowing them to remove redundancy from the data before further processing, and to perform Bayesian inference to reconstruct incomplete data.
Publications
Urs Köster, Jussi Lindgren, Michael Gutmann and Aapo Hyvärinen
Learning Natural Image Structure with a Horizonal Product Model
ICA2009
pdf
|
Urs Köster, Jussi T. Lindgren and Aapo Hyvärinen
Estimating Markov Random Field Potentials for
Natural Images
ICA2009
pdf
|
Urs Köster and Aapo Hyvärinen
A Two-Layer Model of Natural Stimuli Estimated with Score Matching
Pending Review
pdf
|
U. Köster and A. Hyvärinen.
A two-layer ICA-like model estimated by Score Matching.
Proc. Int. Conf. on Artificial Neural Networks (ICANN2007)
pdf
[Shows how to learn weights both layers of a two-layer
model which generalizes ICA, and the models below.]
|
A. Hyvärinen and U. Köster.
Complex Cell Pooling and the Statistics of Natural Images .
Network: Computation in Neural Systems, 18:81-100, 2007.
pdf
[Shows that independent subspace analysis (see above) is a
really better model for natural images, in the sense of a statistical
criterion, than independent component analysis.]
|
A. Hyvärinen and U. Köster.
FastISA: A fast fixed-point algorithm for Independent Subspace Analysis
ESANN2006: 14th European Symposioum on Artificial Neural Networks
PDF
|
Talks (selection)
|
Redwood Seminar 2006 PDF
|
|
Redwood Seminar 2009 PDF
|
Posters
| 
