The neurosciences present some of the steepest challenges to machine learning. Among diverse problem settings and approaches, certain commonalities can be identified. Nearly always there is a very high-dimensional input structure - particularly relative to the number of exemplars, since each data point is usually gathered at a high cost in time and money. To avoid overfitted solutions, inference must therefore make considerable use of domain knowledge from physics, neurophysiology and anatomy. Solutions typically occupy a relatively small subspace of the input representation, the rest being made up of noise that may be of much larger magnitude (often composed largely of the manifestations of other neurophysiological processes, besides the ones of interest). In finding generalizable solutions, one usually has to contend with a high degree of variability, both between individuals and across time, leading to problems of covariate shift and non-stationarity. In all cases, even the high-dimensional raw input is a vastly simplified reflection of the underlying processes and structures. New ways of measuring relevant information, and new ways of transforming the data, are therefore still waiting to be found, leading to feature representations that are more relevant, less noisy, or more transferrable between experimental sessions and subjects.

While BCI is primarily regarded as an alternative means of communication for people with no voluntary motor control, subject groups with diverse cognitive deficits may also benefit from this line of research. BCI-technology can provide subjects with feedback on their neural signals, which may be utilized to learn how to volitionally induce beneficial brain states. One of our contributions to this field is the development of BCI-technology for the control of robotic devices, which we pursue in collaboration with our robotics laboratory. The combination of robotic devices with BCI-technology may be of particular value in the domain of stroke rehabilitation, as BCI-control of robotic-based physical therapy may result in enhanced rehabilitation outcomes. We evaluate the viability of such approaches in collaboration with the Department of Neurosurgery at the University Clinics.

In contrast to the engineering approach of BCI, another role that machine learning can play in neuroscientific research is that of analysis-by-synthesis. We pursue a particular interest in causal inference problems that arise in the context 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 are developing new tools in collaboration with the research group on causal inference at the MPI. In particular, we develop information-theoretical techniques for geometrically representing causal interactions, and kernel methods for detecting statistical dependencies in experimental time series. We apply these methodologies to neurophysiological recordings in order to provide new insight into the mechanisms of perception, and allow computational theories of these mechanisms to be advanced.