Scope:

Many complex real-world tasks, for example, in image processing, brain-like computing, robotics, and financial applications involve high-dimensional, non-linear data. Geometrical analysis can help to understand the complexity of data and task which is a first step towards a solution. Here a variety of kernel methods offer efficient ways for classification, regression, and dimensionality reduction. This workshop will address algorithms and computational examples which involve kernel methods to solve complex tasks. The keynote will focus on fundamental geometrical concepts and pitfalls associated with manifold learning.

Topic:	Approaching Complex Tasks with Kernel Methods
Speaker:	Dr. Stephan Chalup is senior lecturer at the University of Newcastle, Australia. He is leader of the Interdisciplinary Machine Learning Research Group and coordinator of the Newcastle Robotics Laboratory. Currently he is visiting AIFB at the University of Karlsruhe.
	Download the slides

Speaker:	Prof. Dr. D. Seese
Topic:	Complexity in a Foreign Exchange Market
Abstract:	The talk discusses results of the paper Xiaotie Deng, Mao-cheng Cai: Approximation and Computation of Arbitrage in Frictional Foreign Exchange Market, Electronic Notes in Theoretical Computer Science, to appear. Main result of this paper is the NP-completeness and NP-completeness of the approximation problem for a foreign exchange market.
	Download the slides
Speaker:	Tobias Dietrich
Topic:	LIBSVM and R: A Hands-on Demonstration
Abstract:	In his talk Mr. Dietrich demonstrates how to use LIBSVM, a code library for Support Vector Machines, and R, a free software environment for statistical computing, for Support Vector Classification of real-world data.
	Download the slides
Speaker:	Max Christian Ullrich
Topic:	Predicting Foreign Exchange Rate Return Directions with Support Vector Machines
Abstract:	Forecasting financial time series is an important and complex problem in machine learning and statistics. This paper examines and analyzes the general ability of Support Vector Machine (SVM) models to correctly predict and trade daily EUR/GBP, EUR/JPY and EUR/USD exchange rate return directions. For this purpose, six SVM models with varying standard kernels along with one exotic p-Gaussian SVM are compared to investigate the separability of Granger-caused input data in high dimensional feature space. To ascertain their potential value as out-of-sample forecasting and quantitative trading tool, all SVM models are benchmarked against traditional forecasting techniques. We find that hyperbolic SVMs consistently perform well in terms of forecasting accuracy and in terms of trading performance via a simulated strategy. Moreover, it is found that p-Gaussian SVMs perform reasonably well in predicting EUR/GBP and EUR/USD return directions.
	Download the slides