The problem of attributing causes to effects is pervasive in science, medicine, economy and almost every aspects of our everyday life involving human reasoning and decision making. One important goal of causal modeling is to unravel enough of the data generating process to be able to make predictions under manipulations of the system of interest by an external agent (experiments).
Being able to predict the result of actual or potential experiments is very useful because experiments are often costly and sometimes impossible or unethical to perform. For instance, in policy-making, one may want to predict "the effect on a population health status" of "forbidding to smoke in public places", before passing a law. This example illustrates the case of an experiment, which is possible, but expensive. Forcing people to smoke would constitute an unethical experiment.

The need for assisting policy making and the availability of massive amounts of “observational” data prompted the proliferation of proposed causal discovery techniques. Each scientific discipline has its favorite approach (e.g. Bayesian networks in biology and structural equation modeling in social sciences), not necessarily reflecting better match of techniques to domains, but rather historical tradition. Standard benchmarks are needed to foster scientific progress.

Beyond our original motivation, there are many "classical" machine learning problems, which require devising experiments. Hence the "virtual lab" of the causality workbench in a great resource to organize machine learning challenges.

June 2-3, 2008: WCCI 2008, workshop to discuss the results of the first challenge.

Optimization techniques pierced through many chapters of OR, because their utilization “facilitates the choice and the implementation of more effective solutions which, typically, may involve complex interactions among people, materials and money problems” (www.euro-online.org). The latter challenge stimulates a study of existing techniques as well as development of new concepts for optimization. Conference EurOPT-2008 aims at a more exhaustive study of optimization and optimal decision making towards knowledge-based technologies.

The scope of the International conference “Continuous Optimization and Knowledge-Based Technologies” is to overview trends, to gain a common attitude towards latest challenges in continuous optimization and advanced applications for knowledge-based technologies, such as robust optimization, optimization in finance, supply chain management and data mining. We wish to contribute to education in Europe and the world, and to a deepening and initializing of scientific collaboration in the family of EURO and among our peoples.

The modelling of continuous-time stochastic processes from uncertain (discrete) observations is an important task that arises in a wide range of applications, such as in climate modelling, tracking, finance and systems biology. Although observations are in general only available at discrete times, the underlying system is often a continuous-time one. Hence, the physics or the dynamics are formulated by systems of differential equations, the observation noise and the process uncertainty being modelled by several stochastic sources.

When dealing with stochastic processes, it is natural to take a probabilistic approach. For example, we may incorporate prior knowledge about the dynamics by providing probability distributions on the unknown functions. In contrast to models that are only data driven, it is hoped that incorporating domain knowledge in the inference process will improve performance in practice. The main challenges in this context are how to deal with continuous-time objects, how to do inference and how to be agnostic about the deterministic driving forces and the sources of uncertainty.

The workshop will provide a forum for discussing the open problems arising in dynamical systems, and in particular continuous-time stochastic processes. It will focus both on the mathematical aspects/theoretical advances and the applications. Another important aim is to bridge the gap between the different communities (data assimilation, machine learning, optimal control, systems biology, finance, ...) and favour interactions. Hence, the workshop will be of interest to researchers from statistics, computer science, mathematics, physics and engineering. We also hope that the workshop will provide new insights in this exciting field and serve as a starting point for new research perspectives and future collaborations.

We welcome abstract contributions which will be peer-reviewed and selected for presentation. The abstract should not be longer than two A4 pages and in PDF format. It should be sent to ais@cs.ucl.ac.uk by 20 April 2008.

After the workshop, all presentations will be made available on the web as video lectures. Authors will be invited to submit a full paper version of their work for a collective volume that will summarise the major contributions to this meeting.

Organizers

• Cédric Archambeau, University College London.
• Neil Lawrence, University of Manchester.
• Ian Roulstone, University of Surrey.
• John Shawe-Taylor, University College London.
• Andrew Stuart, University of Warwick.

The machine learning team of the "Laboratoire d'Informatique Fondamentale de Marseille" (ML-LIF) is a Joint Research Unit of the Universite de Provence and the CNRS. All of the 5 professors/assistant professors of the team are involved in the organisation of a spring school in machine learning that will be held at the end of next May. The particularity of this school, compared to the machine learning summer schools usually supported by PASCAL, is that it is targeted to an attendance made of people coming from various scientific fields such as mere statistical inference, bioinformatics, machine vision. This short note summarizes some of the salient aspects regarding this school; it particularly stresses the diusion of machine learning over a wide range of research areas.

A second item that is covered by this note deals with the teaching experience of the machine learning team members. It brings some clues as to how undergraduate and graduate students positively perceive machine learning and may fail to sustain their interest when the theoretical aspects of the eld are tack-led.

In making advances within Computational Systems Biology there is an acknowledged need for the ongoing development of both probabilistic and mechanistic, possibly multi-scale, models of complex biological processes. In addition to such models the development of appropriate and efficient inferential methodology to identify and reason over such models is necessary. Examples of the progress which has been made in our understanding of modern biology by the exploitation of such methodology include model based inference of p53 activity; uncovering the evolution of protein complexes and understanding the circadian clock in plants; details of which were presented at the LICSB workshops.

Driven by application areas ranging from biology to the World Wide Web, research in Data Mining and Machine Learning is nowadays increasingly focusing on the analysis of structured data. Of particular interest is data that consists of interrelated parts or is characterized by collections of objects that are interrelated and linked together into complex graphs and structures. Following in the footsteps of the highly successful MLG workshops in the past, MLG 2008 again will be the premier forum for bringing together different sub-disciplines within Machine Learning and Data Mining that focus on the analysis of structured data. The workshop is actively seeking contributions dealing with all forms of structured data, including but not limited to graphs, trees, sequences, relations and networks. Contributions are invited from all relevant disciplines, such as for example

• Statistical Relational Learning
• Inductive Logic Programming
• Kernel Methods for Structured Data
• Probabilistic Models for Structured Data
• Graph Mining
• (Multi-)relational Data Mining
• Methods for Structured Outputs
• Network Analysis

Program committee

PASCAL NoE organizes the first Challenges Workshop that will take place in Southampton, April 11th-13th, in 2005. This Workshop presents the best relevant contributions related to the tasks of 4 PASCAL challenges that have been organized through 2004 and 2005.

Many modern machine learning algorithms reduce to solving large-scale linear, quadratic or semi-definite mathematical programming problems. Optimization has thus become a crucial tool for learning, and learning a major application of optimization. Furthermore, a systematic recasting of learning and estimation problems in the framework of mathematical programming has encouraged the use of advanced techniques from optimization such as convex analysis, Lagrangian duality and large scale linear algebra. This has allowed much sharper theoretical analyses, and greatly increased the size and range of problems that can be handled. Several key application domains have developed explosively, notably text and web analysis, machine vision, and speech all fuelled by ever expanding data resources easily accessible via the web.

This special topic is intended to bring closer optimization and machine learning communities for further algorithmic progress, particularly for developing large-scale learning methods capable of handling massive document and image datasets.

Topics of interest include:

• Mathematical programming approaches to machine learning problems, like semi-definite programming, interior point methods, sequential convex programming, gradient-based methods, etc.
• Optimisation on graphical models for machine learning, belief propagation.
• Efficient training of Support Vector Machines, incremental SVMs, optimization over kernels.
• Convex relaxations of machine learning problems.
• Applications involving large scale databases, such as data mining, bioinformatics, multimedia.

The workshop examines and invites discussion on a range of methods that have been developed for dimension reduction and feature selection. This is a core topic which has been addressed theoretically in many guises from the perspectives of boosting, eigenanalysis, optimisation, latent structure analysis, bayesian methods and traditional statistical approaches to name a few. As an applied technique many algorithms exist for feature selection and all real-world applications of machine learning include some aspect of this in their implementation.

In line with the Thematic Programme 'Linking Learning and Statistics with Optimisation' the workshop focuses on the integration between for example the statistical (frequentist and Bayesian) aspects as well as optimisation issues raised by subspace identification. We feel the workshop provides a real opportunity for interaction between different areas of research and its focus on a strongly applicable family of methods will promote active discussion between different areas of the research community.

Topics considered and contributions are sought in the following areas:

• Dimension reduction techniques, subspace methods
• Random projection methods
• Boosting
• Statistical analysis methods
• Bayesian approaches to feature selection
• Latent structure analysis/Probabilistic LSA
• Optimisation methods
• Novel applications of feature selection algorithms
• Open problems in the domain

Natural signal processing systems have the ability to combine impressions from different senses. This ability enables animals and humans to extract information from and understand noisy and complex environments. An example of this can be seen in most human to human interactions where speech, facial expression, smell, gestures, haptics, etc play a role. While each of these modalities have been modelled at high levels of sophistication, multi-media modeling is still in its infancy.

Multimodal signal processing includes the challenge of handling several sources of information at the same time. Topics of particular importance are multi-stream training and decoding, joint modality integration, fusion of multiple decision streams, confidence estimation and conversion between modalities.

The aim of the workshop is to:

• Introduce problems related to multimodal integration to machine learning researchers.
• Identify and compare different integration strategies.
• Determine the major challenges in using more than one modality.
• Propose novel methods to improve the state of the art.

Speakers will be asked to give a short presentation of their own work including a demonstration on real data. Further more speakers are asked to reserve a minimum of five minutes of their talk to a discussion of the main questions of the workshop. That is, a discussion of the integration strategy used in relation to other possible strategies and a discussion of what major challenges the speaker finds of importance to the field.

The workshop will concentrate on two aspects of multimodal signal processing namely sound and image integration and wearable computing. Further more the physiological/phychological side will be considered in a single talk. The workshop will build on experiences from the "JOINT AMI/PASCAL/IM2/M4 Workshop on Multimodal Interaction and Related Machine Learning Algorithms" (Martigny June 2004) and the "Machine Learning meets the User Interface Workshop" (NIPS 2003).

Organizers

• Tue Lehn-Schiøler
• Samy Bengio
• Lars Kai Hansen
• Stephane Canu
• Jan Larsen