KEYNOTE SPEAKERS LIST
Tiberio Caetano, NICTA, Australia
Title: The Interplay of Statistical and Structural Pattern Recognition from a Machine Learning Perspective
Francis Bach, INRIA, France
Title: Structured Sparsity and Convex Optimization
José C. Príncipe, University of Florida, U.S.A.
Title: Measure of Statistical Dependence
Joachim M. Buhmann, ETH Zurich, Switzerland
Title: Context Sensitive Information: Model Validation by Information Theory
Kostas Triantafyllopoulos, University of Sheffield, U.K.
Title: Detecting Mean Reverted Patterns in Statistical Arbitrage
Tiberio Caetano studied Electrical Engineering, Physics and Computer Science at the Universidade Federal do Rio Grande do Sul (UFRGS), Brazil, where he obtained the PhD degree with highest distinction in 2004. The research part of the PhD program was undertaken at the Computing Science Department at the University of Alberta, Canada. He held a postdoctoral research position at the Alberta Ingenuity Centre for Machine Learning and is currently a senior researcher with the Statistical Machine Learning Group at NICTA. He is also an adjunct faculty member at the Research School of Information Sciences and Engineering, Australian National University. His research interests include machine learning, pattern recognition and computer vision.
The pattern recognition community has traditionally been segmented into two camps: the statistical and the structural/syntactical. While the former has focused on data-driven estimation issues, the latter has devoted particular attention to representational issues. In this talk I will discuss how these approaches have been converging in recent years thanks to developments in machine learning. In particular I will present how traditional questions asked in the structural camp, such as the matching of structural representations of objects, can naturally lend themselves to a statistical, data-driven approach. I will show several benefits of such integrated perspective, both in terms of improved accuracy and improved efficiency. I will show empirical results of the proposed techniques in different application domains, including graph matching in computer vision and image tagging under taxonomic prior knowledge. Finally, I will briefly present my personal perspective on future directions for the field.
Francis Bach is a researcher in the Willow INRIA project-team, in the Computer Science Department of the Ecole Normale Superieure, Paris, France. He graduated from the Ecole Polytechnique, Palaiseau, France, in 1997, and earned his PhD in 2005 from the Computer Science division at the University of California, Berkeley. His research interests include machine learning, statistics, optimization, graphical models, kernel methods, sparse methods and statistical signal processing. He has been awarded a starting investigator grant from the European Research Council in 2009.
The concept of parsimony is central in many scientificdomains. In the context of statistics, signal processing or machinelearning, it takes the form of variable or feature selection problems,and is commonly used in two situations: First, to make the model orthe prediction more interpretable or cheaper to use, i.e., even if theunderlying problem does not admit sparse solutions, one looks for thebest sparse approximation. Second, sparsity can also be used givenprior knowledge that the model should be sparse. In these two situations, reducing parsimony to finding models with low cardinalityturns out to be limiting, and structured parsimony has emerged as afruitful practical extension, with applications to image processing,text processing or bioinformatics. In this talk, I will review recentresults on structured sparsity, as it applies to machine learning andsignal processing (joint work with R. Jenatton, J. Mairal and G.Obozinski).
José C. Príncipe
University of Florida
José C. Príncipe (F'2000) is a Distinguished Professor of Electrical and Computer Engineering with the University of Florida, Gainesville. He is a BellSouth Professor and the Founding Director of the Computational Neuro-Engineering Laboratory, University of Florida. His current research interests are centered in advanced signal processing and machine learning, brain machine interfaces, and the modeling and applications of cognitive systems.
Second order statistics are still the most widely used measures of similarity. Independent Component Analysis showed how important it is to move beyond second order statistics, but independence is still "easy" compared with the concept of statistical dependence. This talk will provide a brief summary of the current methodologies to formulate dependence and will introduce new estimators, namely the generalized measure of association.
Joachim M. Buhmann
Joachim M. Buhmann leads the Machine Learning Laboratory in the Department of Computer Science at ETH Zurich where he has been a professor of Information Science and Engineering (Informatik) since October 2003.
He studied Physics at the Technical University Munich and obtained his PhD in Theoretical Biophysics under the supervision of Professor Klaus Schulten. His doctoral thesis investigated pattern recognition in neural networks. He then spent three years as a research associate and assistant professor at the University of Southern California, Los Angeles. In 1991 he worked at the Lawrence Livermore National Laboratory in California. He held a professorship for applied Computer Science at the University of Bonn, Germany from 1992 to 2003.
His research interests spans the areas of pattern recognition and data analysis, including machine learning, statistical learning theory and applied statistics. Application areas of his research include image analysis, medical imaging, acoustic processing and bioinformatics. More recently, he focused on information theory for machine learning.
He serves as president of the German Pattern Recognition Society since 2009, including serving on the board during 1994-2003. He was associate editor for IEEE Transactions on Neural Networks, IEEE Transactions on Image Processing and IEEE Transaction on Pattern Analysis and Machine Intelligence.
Learning patterns in data requires to extract interesting, statistically significant regularities in (large) data sets, e.g. detection of cancer cells in tissue microarrays or role mining in security permission management. Admissible solutions or hypotheses specify the context of pattern analysis problems which have to cope with model mismatch and noise in data. An information theoretic approach is developed which limits the precision of inferred solutions by noise adapted regularization. The trade-off between "informativeness" and "robustness" is mirrored in the balance between high information content and identifiability of solution sets, thereby giving rise to a new notion of context sensitive information. The effectiveness of the principle is demonstrated by model validation for spectral clustering based on different variants of graph cuts and by analyzing preference data to extract a total order of ranked items.
University of Sheffield
Dr. Triantafyllopoulos obtained a PhD from Warwick University (2002) after completing a BSc in Mathematics at Aristotle University of Thessaloniki (1996) and a MSc in Quality Management at Napier University of Edinburgh (1998). After a post-doctoral appointment at Bristol University (2001-2002) he took a lectureship at Newcastle University (2002-2004). He moved to Sheffield in February 2005. His research interests lie on Bayesian time series analysis and their applications. Areas he works include (a) financial applications with focus on volatility, forecasting, statistical arbitrage and asset allocation; (b) wavelet methods in statistics.
Statistical arbitrage is an important topic of financial econometrics.
Statistical arbitrage usually refers to statistical methods adopted to achieve profits, hence arbitrage. In particular, pairs trading, a simple yet effective subclass of statistical arbitrage, aims to select a pair of assets and at each trading point, the decision is to go long (buy one asset) while go short (short sell) the other asset. Assuming that the data stream process (defined as the difference of the price of the two assets at each time) is mean reverted or stationary, the above trading procedure, will result in profits. Unfortunately, very small number of financial data will form mean reverted spreads; instead it is more common that the data stream will be stationary for some periods and not stationary for other periods.
In this talk we discuss a pattern recognition approach to the detection of such trading periods. These are periods where the data stream is mean reverted, hence one can apply pairs trading. In fact, financial practitioners are exactly interested in such trading periods where it is difficult to recognize by eye whether the data stream is stationary or not.
Our methodology includes some new results on stationarity and it is illustrated with several data streams, consisting of prices of assets.