Prof. Dr. Bernhard Schölkopf
Organization: Max Planck Institute for Intelligent Systems
Phone number: +49 (0)7071 601 551
Department: Dept. of Empirical Inference
Position: Director, Dept. of Empirical Inference
Area: CIN Members
Scientific topic: Empirical Inference - Decoding Complexity
Field of Research
Uncertainties are present at many levels in biological and artificial adaptive systems. Even with data of high quality, gauging and combining a multitude of data sources and constraints in usually imperfect models of the world requires us to represent and process uncertain knowledge in order to take viable decisions. In the increasingly complicated settings of modern science, model structure or causal relationships may not be known a-priori.
The focus of our research directly addresses the question of unobserved variables, both at the level of algorithms for machine learning, and of individual applications of inference. New learning that can detect structures in experimental data processes, for example algorithms for pattern recognition, regression, density estimates, novelty detection and feature selection, are developed in the “Empirical Inference” department. We develop new algorithms and approximation techniques, assess their quality both in theory and in empirical evaluations, and apply them to challenging problems, focusing on the development of so-called kernel methods.
The basic problem with learning is “generalization”: the extracted laws are supposed to explain not just existing observations (the “training data”) correctly, but also to be applied new observations. This problem of induction not only touches fundamental issues in statistics but in the empirical sciences in general. These issues include the representation of data and existing knowledge and the complexity or capacity of explanations or models. If a given quantity of empirical observations can be adequately accounted for by a model of low complexity, then statistical learning theory guarantees that it is highly probable that future observations will be consistent with the model.
In experimental design, the goal is to optimize data sampling or measurement architectures so that inference or estimation can be realized faster or at lower costs. The expertise spanned by our group allows us to combine Bayesian techniques with ideas coming from adjacent fields. Many of these techniques and concepts come from or have been developed in the context of classical statistics. Processing data from brain-machine interfaces is however only one problem amongst many which can be solved with learning theory.
Brain-Computer Interfacing (BCI) for Communication
A brain-machine interface allows humans to interact with their environment without the aid of muscle power. For example, paralyzed patients can spell out words and form sentences using just their thoughts. However, this requires costly data analysis of the complex multidimensional signals that the brain generates to control this process. It is carried out with “support vector machines” which automatically extract regularities from a large amount of data, which are then used for control purposes. This research primarily aims to construct systems that could allow a paralyzed person to communicate, by decoding the user's intentions from measured brain signals. Currently, BCI systems based on electroencephalogram (EEG) signals do allow communication, and can be used even by people with very little remaining motor control, but they are still slow and difficult enough to use that they are not an attractive alternative to other communication methods. Among people who have absolutely no voluntary motor control, and who therefore stand to benefit most from BCI, a convincing demonstration of successful communication has yet to be published.
We have been working hands-on in collaboration with the Institute of Medical Psychology and Behavioral Neurobiology, and the Department of Neurosurgery at the University Clinics, in order to develop and test BCI systems with paralyzed patients. In addition, we pursue several laboratory-based lines of research. The contribution of machine learning to BCI is in developing, refining and applying algorithms to improve the accuracy with which neural signals are decoded, to interpret users' communication intentions more reliably and to reduce training times for the user. We regard algorithmic development and improvements in experimental methodology as significant contributions towards making clinical BCI systems a reality.
Brain-Computer Interfacing for Rehabilitation
While BCI is primarily regarded as an alternative means of communication for people with no voluntary motor control, also subjects with milder neurological deficits could benefit from this line of research. BCI-technology can provide them patients with such deficits with feedback on their altered neural signals, which may be utilized to learn how to induce beneficial changes of brain states disturbed by stroke or other types of brain disease. One of our contributions to this field is the development of BCI-technology for the control of robotic devices. The combination of robotic devices with BCI-technology might be of particular value in the domain of stroke rehabilitation through robotic-based physical therapy. We evaluate the viability of such approaches in collaboration with the Department of Neurosurgery at the University Clinics.
Another role that machine learning can play in neuroscientific research is that of analysis-by-synthesis. We have a particular interest in causal inference problems that arise in the context of neuroscientific research. Functional brain networks are highly recurrent due to their local organization in microcircuits and the bidirectional connections linking remote cortical areas. To understand how information processing is distributed in these structures, we develop new tools in collaboration with the research group on causal inference at the MPI. We apply these methodologies to neurophysiological recordings in order to provide new insight into the mechanisms of perception, and advance computational theories of these mechanisms.