The research agenda of the Bethge Lab
can be summarized by the two basic questions originally posed by Hermann von Helmholtz (1867,1878): “What are the principles that govern how the visual pathways make inferences from the visual image? How do we use image information to compute these perceptual inferences?” A principal difficulty in the understanding of biological vision is the complexity of the inference problems we encounter both at the level of behavior as well as at the level of neuronal responses. This complexity mostly results from the large number of degrees of freedoms in the sensory input and in the neuronal responses. Using methods of statistical inference and learning theory, as well as signal processing, nonlinear dynamics and optimization theory, our research addresses the problem of perceptual inference from natural images and its neural basis at different levels:
- We develop mathematical generative models of natural images and image transformations using unsupervised learning methods. Particular emphasis is placed on quantitative comparisons of the performance of these models.
- We perform psychophysical studies in order to evaluate the relationship between natural image models and perception.
- We develop new efficient methods to predict the spike trains of neurons in response to natural stimuli with the goal of inferring the contribution of these neurons to the image processing performed in the early visual system. In particular, we build population response models for multi-cell recordings and we address the aspects of contrast adaptation, non-Gaussian stimuli and interspike correlations.