Workshops

In the past five years machine learning classifiers have met great interest in the field of cognitive neuroscience. They have been used to make predictions about the mental state of subjects directly from fMRI data, as well as study the neural encoding of specific mental contents in the human brain, in ways that transcend the limitations of conventional methods. The goals of this workshop are to present the current challenges in the field from the points of view of cognitive neuroscience and machine learning practitioners, as well as bring to the fore new work that goes beyond mere decoding to the study of the structure present in fMRI activity. More broadly, we will be concerned with the implications of multivariate decoding results for theories of cognitive neuroscience, as well as the manner in which those theories may influence the practice of decoding.

This workshop will focus on algorithms, statistical tests, metrics, and analysis tools for assessing the quality of learning systems in the cotext of an actual real-world task, the design of learning systems for deploying them in safety-critical problem, the stability of online learning, tradeoffs between robustness and risks in making decisions, and on all aspects of validation and testing of learning systems.

Trading exploration and exploitation plays a key role in a number of learning tasks. For example the bandit problem provides perhaps the simplest case in which we must decide a trade-off between pulling the arm that appears most advantageous and experimenting with arms for which we do not have accurate information. Similar issues arise in learning problems where the information received depends on the choices made by the learner. Learning studies have frequently concentrated on the final performance of the learned system rather than consider the errors made during the learning process. For example reinforcement learning has traditionally been concerned with showing convergence to an optimal policy, while in contrast analysis of the bandit problem has attempted to bound the extra loss experienced during the learning process when compared with an a priori optimal agent. This workshop provides a focus for work concerned with on-line trading of exploration and exploitation, in particular providing a forum for extensions to the bandit problem, invited presentations by researchers working in related areas in other disciplines, as well as discussion and contributed papers.

The identification of an effective function to compare examples is essential to several machine learning problems. For instance, retrieval systems entirely depend on such a function to rank the documents with respect to their estimated similarity to the submitted query. Another example is kernel-based algorithms which heavily rely on the choice of an appropriate kernel function. In most cases, the choice of the comparison function (also called, depending on the context and its mathematical properties, distance metric, similarity measure, kernel function or matching measure) is done a-priori, relying on some knowledge/assumptions specific to the task. An alternative to this a-priori selection is to learn a suitable function relying on a set of examples and some of its desired properties. This workshop is aimed at bringing together researchers interested in such a task.

Systems that adapt to their users have the potential to tailor the system behavior to the specific needs and preferences of their users. The purpose of this workshop is to bring together researchers from academia and industry to summarize previous work; evaluate the need for user adaptive systems; and discuss the main difficulties that arise in designing and implementing such systems. This one day workshop will allow people working on different types of user adaptive systems to exchange ideas and to learn from each other's experience.

Understanding how world knowledge can be grounded in sensori-motor experience has been a long-standing goal of philosophy, psychology, and artificial intelligence. So far this goal has remained distant, but recent progress in machine learning, cognitive science, neuroscience, engineering, and other fields seems to bring nearer the possibility of addressing it productively. The objective of this workshop is to provide cross-fertilization of ideas between diverse research communities interested in this subject. The workshop will be comprised of invited talks by 5-6 of the top people from a variety of disciplines related to experience based knowledge representations. The speakers will share their area-specific knowledge and understanding of these issues with the workshop attendees. The workshop will conclude with a poster session populated with work submitted by the community at large.

Continuous attractor neural networks (CANNs) are a special type of recurrent networks that have been studied in many neuro- and cognitive science areas such as modelling hypercolumns, movement generation, spatial navigation, working memory, population coding, attention, saccade initiation and decision making. They have been also applied to engineering problems such as robot control. Such neural field models of the Wilson-Cowan-Amari type, or bump models, are a fundamental type of neural circuitry that underlies the general mechanisms for neural systems encoding continuous stimuli and categorizing objects. The goal of the workshop is to bring together researchers from diverse areas to solidify the existing research on CANNs, identify important issues that need to be solved, and explore their potential applications to artificial learning systems.

There has been a large body of work developed in the last 10 years on the human pose estimation and tracking from video. Many of these methods are based on well founded statistical models and machine learning techniques. Progress however has been limited because of the lack of common datasets and error metrics for quantitative comparison. The goal of this workshop is to (1) establish the current state of the art in the human pose estimation and tracking from single and multiple camera views, (2) discuss future directions in the field, and (3) introduce a benchmark database and error metrics for comparing current and future methods. To this end a new (HumanEva) dataset for evaluation of articulated human motion will be introduced.

Revealing and modeling the hidden state-space of dynamical systems is a fundamental problem in signal processing, control theory, and learning. Classical approaches to this problem include hidden Markov models, reinforcement learning, and various system identification algorithms. More recently, the problem has been approached by such modern machine learning techniques as kernel methods, Bayesian and Gaussian processes, latent variables, and the information bottleneck. Moreover, dynamic state-space learning is the key mechanism in the way organisms cope with complex stochastic environments, i.e., biological adaptation. One familiar example of a complex dynamic system is the authorship system in the NIPS community. Such a system can be described by both internal variables, such as links between NIPS authors, and external environment variables, such as other research communities. This complex system, which generates a vast number of papers each year, can be modeled and investigated using various parametric and non-parametric methods. In this workshop we intend to review and confront different approaches to dynamical system learning, with various applications in machine learning and neuroscience. In addition, we hope this workshop will familiarize the machine learning community with many real-world examples and applications of dynamical system learning. Such examples will also serve as the basis for the discussion of such systems in the workshop.

This workshop explores the use of causality with predictive models in order to assess the results of given actions. Such assessment is essential in many domains, including epidemiology, medicine, ecology, economy, sociology and business. Predictive models simply based on event correlations do not model mechanisms. They allow us to make predictions in a stationary environment (no change in the distribution of all the variables), but do not allow us to predict the consequence of given actions. For instance, smoking and coughing are both predictive of respiratory disease. One is a cause and the other a symptom. Acting on the cause can change the disease state, but not acting on the symptom. Understanding the effect of interventions has been the goal of most causal models but their complexity has limited their use to a few hundreds variables. Feature selection on the other hand can handle thousands of variables at the same time but does not make a difference between causes and symptoms. By confronting the hypothesis underlying causality and feature selection approaches, this workshop aims at investigating new approaches to extract causal relationships from data.

During the last decade, many areas of statistical machine learning have reached a high level of maturity with novel, efficient, and theoretically well founded algorithms that increasingly removed the need for heuristics and manual parameter tuning, which dominated the early days of neural networks. Reinforcement learning (RL) has also made major progress in theory and algorithms, but is somehow lagging behind the success stories of classification, supervised, and unsupervised learning. Besides the long-standing question for scalability of RL to larger and real world problems, even in simpler scenarios, a significant amount of manual tuning and human insight is needed to achieve good performance, e.g., as in exemplified in issues like eligibility factors, learning rates, the choice of function approximators and their basis functions for policy and/or value functions, etc. Some of the reasons for the progress of other statistical learning disciplines comes from connections to well- established fundamental learning approaches, like maximum-likelihood with EM, Bayesian statistics, linear regression, linear and quadratic programming, graph theory, function space analysis, etc. Therefore, the main question of this workshop is to discuss, how other statistical learning techniques may be used to developed new RL approaches in order to achieve properties including higher numerical robustness, easier use in terms of open parameters, probabilistic and Bayesian interpretations, better scalability, the inclusions of prior knowledge, etc.

The field of computational biology has seen dramatic growth over the past few years, both in terms of new available data, new scientific questions, and new challenges for learning and inference. In particular, biological data is often relationally structured and highly diverse, well-suited to approaches that combine multiple weak evidence from heterogeneous sources. These data may include sequenced genomes of a variety of organisms, gene expression data from multiple technologies, protein expression data, protein sequence and 3D structural data, protein interactions, gene ontology and pathway databases,genetic variation data (such as SNPs), and an enormous amount of textual data in the biological and medical literature. The goal of this workshop is to present emerging problems and machine learning techniques in computational biology. The workshop will include invited and submitted talks from experts in the fields of biology, bioinformatics and machine learning. We encourage contributions describing either progress on new biological problems or work on established problems using methods that are substantially different from standard approaches. Deadline for contributed abstracts: October 31, 2006.

The analysis of audio signals is central to the scientific understanding of human hearing abilities as well as in engineering applications such as sound localisation, hearing aids or music information retrieval. In recent years, there is an increasing interest for a Bayesian treatment and the application of graphical models which together permit increasingly refined analyses and representations of the acoustic signals. Such techniques are quite natural since acoustical time series can be conveniently modelled using hierarchical signal models by incorporating prior knowledge from physics or studies of human cognition and perception. The workshop will address advances in modelling and inference techniques for acoustics and how to integrate relevant prior information from sources such as neurobiology and form useful acoustic representations.

Open source tools have recently reached a level of maturity which makes them suitable for building large-scale real-world systems. At the same time, the field of machine learning has developed a large body of powerful learning algorithms for a wide range of applications. Inspired by similar conferences in bioinformatics (BOSC) or statistics (useR), our aim is to build a forum for open source software in machine learning. The workshop's ultimate goal is to bring existing and freshly developed toolboxes and algorithmic implementations to people's attention.

A Brain-Computer Interface (BCI) is a novel augmentative communication system that translates human intentions - reflected by suitable brain signals - into a control signal for an output device such as a computer application or a neuroprosthesis. The crucial point is that the system works without using the normal output pathways of peripheral nerves and muscles, just the manifestation of thought. In developing a BCI system many fields of research are involved, such as classification, signal processing, neurophysiology, measurement technology, psychology, control theory. Since the 1970s several labs succeeded in building BCI systems that can provide feedback control from thoughts. The classical approach is to establish EEG-based control by setting up a system that is controlled by a specific EEG feature which is known to be susceptible to conditioning and to let the subjects learn the voluntary control of that feature, a process that may well require several weeks or month. In recent years a machine learning approach to BCI was pioneered by Fraunhofer FIRST Berlin (NIPS*01) and the IDIAP Institute Martigny in which the system adapts to the individual brain signatures of each user, thereby drastically reducing or even eliminating the need for subject training. Still we are faced with a lot of problems, e.g., the dramatic inter-subject variability in performance and the question how to adapt classifiers online to cope with the changing characteristics of background activity. Furthermore several labs started to investigate Brain-Computer Interfaces that use brain signals recorded invasively from the surface of the cortex or from inside the brain. Due to the challenging problems in the design of the BCI control it is of uttermost necessity to bring together researchers from all involved fields to discuss possible solutions and new directions of research.

There is great interest in sensory coding. Studies of sensory coding typically involve recording from sensory neurons during stimulus presentation, and the investigators determine which aspects of the neuronal response are most informative about the stimulus. These studies are left with a decoding problem: are the discovered codes, sometimes quite exotic, ultimately decoded by downstream networks and used by the nervous system to guide behavior? In our one-day workshop, researchers with many different backgrounds will evaluate what we know about neuronal decoders and suggest new strategies, both experimental and computational, for addressing the decoding problem.

The modelling of continuous-time dynamical systems from uncertain observations is an important task that comes up in a wide range of applications ranging from numerical weather prediction over finance to genetic networks and motion capture in video. Often, we may assume that the dynamical models are formulated by systems of differential equations. In a Bayesian approach, we may then incorporate a priori knowledge about the dynamics by providing probability distributions over the unknown functions, which correspond for example to driving forces and appear as coefficients or parameters in the differential equations. Hence, such functions become stochastic processes in a probabilistic Bayesian framework. Gaussian processes (GPs) provide a natural and flexible framework in such circumstances. The use of GPs in the learning of functions from data is now a well-established technique in Machine Learning. Nevertheless, their application to dynamical systems becomes highly nontrivial when the dynamics is nonlinear in the (Gaussian) parameter functions as closed form analytical posterior predictions (even in the case of Gaussian observation noise) are no longer possible. Moreover, their computation requires the entire underlying Gaussian latent process at all times (not just at the discrete observation times). Hence, inference of the dynamics would require nontrivial sampling methods or approximation techniques. The aim of this workshop is to provide a forum for discussing open problems related to stochastic dynamical systems, their links to Bayesian inference and their relevance to Machine Learning.

Many machine learning algorithms assume that the training and the test data are drawn from the same distribution. Indeed many of the proofs of statistical consistency, etc., rely on this assumption. However, in practice we are very often faced with the situation where the training and the test data both follow the same conditional distribution, p(y|x), but the input distributions, p(x), differ. For example, principles of experimental design dictate that training data is acquired in a specific manner that bears little resemblance to the way the test inputs may later be generated. The aim of this workshop will be to try and shed light on the kind of situations where explicitly addressing the difference in the input distributions is beneficial, and on what the most sensible ways of doing this are.

Autonomous robot navigation in unstructured outdoor environments remains a critical challenge for tasks such as reconnaissance, search and rescue and automated driving. This workshop addresses two main components necessary for formulating open problems in outdoor navigation within the theoretical framework of Machine Learning. The first is concerned with using color cameras as the primary sensing modality for learning models of traversable terrain over the long term. The second is concerned with learning the necessary locomotion required to allow legged robots to efficiently move through rough outdoor terrain.

A new approach to analyzing and training recurrent neural networks (RNNs) has emerged over the last few years. The central idea is to regard a sparsely connected recurrent circuit as a nonlinear, excitable medium, which is driven by input signals (possibly in conjunction with feedbacks from readouts). This recurrent circuit is --like a kernel in Support Vector Machine applications-- not adapted during learning. Rather, very simple (typically linear) readouts are trained to extract desired output signals. Despite its simplicity, it was recently shown that such simple networks have (in combination with feedback from readouts) universal computational power, both for digital and for analog computation. There are currently two main flavours of such networks. Echo state networks were developed from a mathematical and engineering background and are composed of simple sigmoid units, updated in discrete time. Liquid state machines were conceived from a mathematical and computational neuroscience perspective and usually are made of biologically more plausible, spiking neurons with a continuous-time dynamics. Generic cortical microcircuits are seen from this perspective as explicit implementations of kernels (in the sense of SVMs), that therefore are not required to carry out specific nonlinear computations (as long as their individual computations and representations are sufficiently diverse). Obviously this hypothesis provides a new perspective of neural coding, experimental design in neurobiology, and data analysis. This workshop will cover theoretical aspects of this approach, applications to concrete engineering tasks, as well as results of first neurobiological experiments that have tested predictions of this new model for cortical computation.

Regular competitive events have helped drive research progress and further the state-of-the-art in many areas, such as: data mining (KDD Cup), planning and scheduling (International Planning Competition) and multi-agent robotics (Robocup). Over the past two years the reinforcement learning community has held two competitive meetings. This workshop will be the culmination of the First Annual Reinforcement Learning Competition. The competition will use the same evaluation software and similar problem sets as the previous competitions, but also include several new events. This competition will feature a pentathlon: the agent that achieves the best performance across all five domains will be declared the winner. The winners will be invited to describe their approach at the workshop, with the aim of improving the participants' expertise on applying reinforcement learning techniques in practice.

When training a learning machine, both practical and theoretical considerations may yield to split the problem into multiple levels of inference. Typically, at the lower level, the parameters of individual models are optimized and at the second level the best model is selected, e.g. via cross-validation. But, there may be more than two levels of inference and cross-validation is not the only way of addressing the resulting optimization problem. This workshop will revisit the problem of model selection, with the goal of bridging the gap between theory and practice. A * game of model selection is organized *, check the web-site!

With increasing pressure for accessing, understanding and translating information available in different languages, Machine Learning has the potential to provide much needed technology for multilingual applications. However, there are also specific issues and pitfalls, such as the need to work with limited resources, to integrate prior linguistic knowledge, to scale up to very large text corpora, and to provide evaluation measures that correspond to human assessments. This workshop will provide the opportunity for researchers interested in Machine Learning and/or Computational Linguistics to analyse and discuss challenges in applying ML to multilingual information access, possible solutions, and existing applications.

Dimensionality reduction has been one of the most active research areas of machine learning in the past few years. Novel algorithms for nonlinear dimensionality reduction (Isomap, locally linear embedding, local tangent space alignment, etc.) and supervised dimensionality reduction (neighborhood components analysis, max-margin matrix factorization, support vector decomposition, etc.) have taken significant steps toward overcoming deficiencies in traditional methods like PCA and Fisher's LDA. It is intuitive that dimensionality reduction helps to visualize data, reduce computational complexity, and avoid overfitting. But in practice there are still relatively few applications which make use of new dimensionality reduction techniques. The goal of this workshop is to understand how to match the capabilities of new nonlinear and supervised dimensionality reduction techniques with practical applications in science, engineering, and technology. We hope to achieve this by bringing together researchers who develop these techniques and those who apply them. A successful workshop will lead to new directions for application-oriented dimensionality reduction research and ignite cross-fertilization between different application domains. We are especially interested in applications from biology (particularly neuroscience and genetics), psychology, human and computer vision, auditory signal processing, and text analysis.