Document Actions

You are here: Home / Research / Meet the Scientist / Matthias Bethge

Matthias Bethge

Bernstein Awardee 2006,
Max Planck Institute for Biological Cybernetics


Matthias Bethge received the first Bernstein Award in the year 2006. The award is conferred annually by the Federal Ministry of Education and Research and enables excellent young scientists to establish their own research group.

Bethge studied physics in Göttingen. After his dissertation in Bremen, he did a post doc in the laboratory of Bruno Olshausen at the Redwood Neuroscience Institute in California. Since November 2005 Bethge has worked at the Max Planck Institute for Biological Cybernetics in the group of Bernhard Schölkopf.

The language of the brain

Neurons transmit signals in the form of short electrical impulses–they ‘fire’. Every sensation, from the scent of a rose to the play of colors in an evening sky, is converted into such a neuronal ‘morse code’ in the brain. Since his Diploma Thesis, Bethge has been interested in the question of neuronal infor-mation coding. Is a neuron‘s firing rate critical to information coding; does the combined activity pattern of several neurons play a role? Information coding in the brain must be extremely fast. Bethge’s calculations show that a maximal information rate can be reached, when neurons act according to an all-or-nothing law and either fire salvos of impulses or show only minimal activity. ‘Minor differences in the frequency of impulses can not be resolved in such a short time’, Bethge explains. In addition, the system can only function, if populations of neurons collectively code information. ‘It is a great challenge to analyze and understand such a system,’ says Bethge.

Not every piece of information that reaches the brain through the sense organs is processed. Rather, the brain can distinguish relevant from irrelevant information and filter the latter one out. But not only that, it can also interpret visual information. For example, the brain has the ability to deduce the three dimensional structure of objects from the two dimensional image that is projected onto our retina. Using different methodological approaches, Bethge addresses how the visual system can infer viable information from the enormous amount of sensory input.

A ‘top-down approach’, as Bethge calls it, is the abstract reflection, which principles the brain could use to optimally pro-cess visual information. ‘What would we do, if we wanted to build a machine which can recognize three-dimensional structures?’ asks Bethge. This may not only lead to a better understanding of biological vision, it can also advance applications in the area of artificial intelligence. ‘Research in the field of computer vision usually attempts to find solutions for very specific applications’, says Bethge. Only very narrow classes of objects can be recognized by computers. In order to do so, they resort to previous knowledge, which cannot be extrapolated to other object classes. The brain, in contrast, is capable of understanding the features of an abstract sculpture of which it has no previous knowledge, based only on the distribution of shadows and edges. ‘We try to understand the principles of such an analysis of shapes, so that a computer will at some point be able to interpret images with diverse contents.’

In a ‘bottom-up approach’ Bethge sets out from measure-ments of neuronal activity. From the way in which neurons reactto visual stimuli, he draws conclusions about image recognition. Current models of how the brain analyzes images generally assume a linear mode of information processing in several successive layers. Such models, however, reflect reality only in a limited way; they are not very good at predicting neuronal activity in the retina or the subsequent processing areas of the brain. Bethge is developing methods that go beyond such conventional models and also take the interaction of neurons within each layer into account.

Comparable to a computer program for image compression, the brain strives to dispose of redundant information. Not every pixel of a uniformly blue surface needs to be independently pro-cessed–missing regions can be easily filled in. So called ‘simple cells’ in the primary visual cortex preferentially react to edges and contours. Statistical models for image processing predict that edges are particularly independent image components. Through quantitative analysis, Bethge could show that other aspects of an image can similarly serve as independent image components. Together with Felix Wichman, also at the Max Planck Institute for Biological Cybernetics, Bethge tests his theoretical results using psychophysical methods. ‘Test persons can predict edges and contours very well, meaning that they are no less redundant than other aspects of an image’, explains Bethge. Given the enormous complexity of the brain, mathematical models and computer analyses are essential tools to unravel its function. But in the end, these models have to be measured against reality.