All talks will be held in lecture hall H6 at the western end of the main hall.
The time for a talk is 20 minutes plus 5 min discussion.
Wednesday, 2007-09-05 - from 16:00 to 17:45
Class imaging of hyperspectral satellite remote sensing data using FLSOM
|Presenting Author: Thomas Villmann|
|Authors: Thomas Villmann, F.-M. Schleif, E. Merenyi, M. Strickert, B. Hammer|
We propose an extension of the self-organizing map for supervised fuzzy classification learning, whereby uncertain (fuzzy) class information is also allowed for training data. The method is able to detect class similarities, which can be used for data vizualization. Applying a special functional metric, derived from of the Lp norms, we show the application of the method for classification and visualization of hyper-spectral data in satellite image remote sensing image analysis.
Label Propagation for Semi-Supervised Learning in Self-Organizing Maps
|Presenting Author: Lutz Herrmann|
|Authors: Lutz Herrmann, Alfred Ultsch|
Semi-supervised learning aims at discovering spatial structures in high-dimensional input spaces when insufficient background information about clusters is available. A particulary interesting approach is based on propagation of class labels through proximity graphs. The Self-Organizing Map itself can be seen as such a proximity graph that is suitable for label propagation. It turns out that Zhu's popular label propagation method can be regarded as a modification of the SOM's well known batch learning rule. In this paper, an approach for semi-supervised learning is presented. It is based on label propagation in trained Self-Organizing Maps. Furthermore, a simple yet powerful method for crucial parameter estimation is presented. The resulting clustering algorithm is tested on the fundamental clustering problem suite (FCPS).
Single pass clustering for large data sets
|Presenting Author: Barbara Hammer|
|Authors: Nikolai Alex, Barbara Hammer, Frank Klawonn|
The presence of very large data sets poses new problems to standard
neural clustering and visualization algorithms such as Neural Gas (NG)
and the Self-Organizing-Map (SOM) due to memory and
time constraints. In such situations, it is no longer possible to store
all data points in the main memory at once and only a few, ideally only one
run over the whole data set is still affordable to achieve a feasible
training time. In this contribution we propose single pass extensions
of the classical clustering algorithms NG and fuzzy-k-means which
are based on a simple patch decomposition of the data set and fast
batch optimization schemes of the respective cost function. The algorithms
maintain the benefits of the original ones including easy implementation and
interpretation as well as large flexibility and adaptability because of the
underlying cost function. We demonstrate the efficiency of the approach in a variety
Decision Manifolds: Classification Inspired by Self-Organization
|Presenting Author: Thomas Lidy|
|Authors: Georg Pölzlbauer, Thomas Lidy, Andreas Rauber|
We present a classifier algorithm that approximates the decision surface of labeled data by a patchwork of separating hyperplanes. The hyperplanes are arranged in a way inspired by how Self-Organizing Maps are trained. We take advantage of the fact that the boundaries can often be approximated by linear ones connected by a low-dimensional nonlinear manifold. The resulting classifier allows for a voting scheme that averages over the classifiction results of neighboring hyperplanes. Our algorithm is computationally efficient both in terms of training and classification. Further, we present a model selection framework for estimation of the paratmeters of the classification boundary, and show results for artificial and real-world data sets.