Dirichlet Processes, Chinese Restaurant Processes, and all that
Invited Speaker: Michael I. Jordan
Dept. of Electrical Engineering and Computer Science, Dept. of Statistics University of CaliforniaBerkeley,USA
Bayesian approaches to learning problems have many virtues, including their ability to make use of prior knowledge and their ability to link related sources of information, but they also have many vices, notably the strong parametric assumptions that are often invoked willy-nilly in practical Bayesian modeling. Nonparametric Bayesian methods offer a way to make use of the Bayesian calculus without the parametric handcuffs. In this talk I describe several recent explorations in nonparametric Bayesian modeling and inference, including various versions of "Chinese restaurant process priors" that allow flexible structures to be learned and allow sharing of statistical strength among sets of related structures. I discuss applications to problems in bioinformatics and information retrieval.
Privacy and Background Knowledge
Invited Speaker: Johannes Gehrke
Dept. of Computer Science at Cornell University and Faculty Associate Director of the Cornell Theory CenterNew York, USA
The digitization of our daily lives has led to an explosion in the collection of data by governments, corporations, and individuals. Protection of confidentiality of this data is of utmost importance. However, knowledge of statistical properties of private data can have significant societal benefit, for example, in decisions about the allocation of public funds based on Census data, or in the analysis of medical data from different hospitals to understand the interaction of drugs. I will start by introducing two application scenarios, privacy-preserving data analysis and privacy-preserving data publishing. I will show how in simple models background knowledge can lead to severe breaches of privacy in both applications, and I will describe how proper modeling of background knowledge can avoid privacy breaches. I will outline first algorithmic steps towards privacy-preserving data analysis and data publishing with background knowledge, and I will conclude with open problems.
Why Computers Need to Learn About Music
Invited Speaker: Gerhard Widmer
University Linz, Austria
The goal of this presentation is to convince the research community that music is much more than an interesting and "nice", but ultimately esoteric toy domain for machine learning experiments. I will try to show that right now is the time for machine learning to really make an impact in both the arts, the (music) sciences, and, not least, the music market. In order to demonstrate that, some impressions will be given of what computers can currently do with music. In the domain of classical music, I will show how machine learning can give us new insights into complex artistic behaviours such as expressive music performance, with examples ranging from the automatic discovery of characteristic stylistic patterns to automatic artist identification and even computers that learn to play music with "expression". In the (commercially more relevant) domain of popular music, the currently ongoing rapid shift of the music market towards digital music distribution opens myriads of application possibilities for machine learning, from intelligent music recommendation services to content-based music search engines to adaptive radio stations. Again, some ongoing work in this area will be briefly demonstrated. A number of challenges for machine learning research will be identified throughout the presentation, and my hope is that after the conferences, a large part of the ICML and ILP attendants will go back to their labs and get involved in machine learning and music right away.
Joint ICML/ILP Tutorial: Statistical Relational Learning
Invited Speaker: Lise Getoor
University of Maryland, USA
Statistical machine learning is in the midst of a "relational revolution". After many decades of focusing on independent and identically-distributed (iid) examples, many researchers are now studying problems in which the examples are linked together into complex networks. These networks can be a simple as sequences and 2-D meshes (such as those arising in part-of-speech tagging and remote sensing) or as complex as citation graphs, the world wide web, and relational data bases. Statistical relational learning raises many new challenges and opportunities. Because the statistical model depends on the domain's relational structure, parameters in the model are often tied. This has advantages for making parameter estimation feasible, but complicates the model search. Because the "features" involve relationships among multiple objects, there is often a need to intelligently construct aggregates and other relational features. Problems that arise from linkage and autocorrelation among objects must be taken into account. Because instances are linked together, classification typically involves complex inference to arrive at "collective classification" in which the labels predicted for the test instances are determined jointly rather than individually. Unlike iid problems, where the result of learning is a single classifier, relational learning often involves instances that are heterogeneous, where the result of learning is a set of multiple components (classifiers, probability distributions, etc.) that predict labels of objects and logical relationships between objects. In this tutorial, we will survey several of the major branches of this newly emerging field: rule-based approaches, frame-based approaches and stochastic/functional programming approaches. We will describe representational issues, learning and inference. Many of the approaches are based in some way on graphical models, and we will describe approaches which are based on both directed and undirected graphical models. We will describe several useful inference tasks such as link prediction, group detection and entity resolution and applications areas including citation graphs, the world wide web and social networks.