Neuronal Dynamics

The mammalian brain is a highly complex, hierarchically organized dynamical system which is very efficient in processing information from the sensory level and from internal sources. In different projects, we investigate how the particular dynamics of neurons and neural circuits supports important cognitive functions, and how certain principles of information processing imply specific neural mechanisms and structures.

Our research combines bottom-up (dynamical systems) and top-down approaches (probabilistic, generative models) to focus on three main aspects of neural dynamics, namely rapid information processing with stochastic spikes, information integration and feature combination in the visual system, and adaptation of information processing to task and stimulus context. The majority of projects is pursued in close collaboration with experimentalists, in particular with human psychophysics lab of Manfred Fahle and monkey electrophysiology lab of Andreas Kreiter.

Contextual modulation and non-classical receptive fields

Receptive fields (RFs) capture the basic response properties of neurons or neuron populations to external stimuli, and are essential to understand information processing in the brain. Usually, RFs are experimentally characterized by performing reverse correlation or spike-triggered average on neural responses. However, these linear methods reveal only part of the stimuli (the 'classical' RF, cRF) that trigger a neuron's response:  in addition, neural activity is strongly modulated by contextual stimuli outside the cRF which would have no effect on the cell if they were presented alone (non-classical RFs).

We investigate these non-linear response properties both in realistic cortical network models, and in generative models. In particular, we find that ncRFs can be explained by an optimal inference process performed on a noisy stimulus. The shape of ncRFs is hereby determined by the statistics of natural scenes which are used as stimuli. The neural dynamics reflects an on-going information integration process, where different neurons compete for explaining part of the stimulus (collaboration with Sophie Deneve).

Contour Integration

Contour integration is a cognitive process where multiple, localized edge elements are linked into global percepts of contours. It serves to detect curves and borders of objects, and is thus important for segmenting and interpreting natural scenes. In different projects, we investigate putative neural mechanisms and useful computational strategies for contour integration.

Our approach is to use probabilistic, generative models to capture both contour generation and ideal contour integration in a single theoretical framework. Model predictions are tested against human behaviour in psychophysical experiments that use the same contour stimuli (collaboration with Sunita Mandon). Our models provide a quantitative account and a concise functional interpretation of human contour integration, and they predict novel mechanisms underlying this important cognitive process.