Feature integration in generative models of visual cortex

Abstract:

The visual system of humans and primates combines localized features or image patches into representations which become more and more complex in higher visual areas. Generative models allow to study these integration processes in a well-defined probabilistic framework that allows to link cognitive functions to neural dynamics and to behaviour. In our work, we use such an approach to investigate two important aspects of visual feature integration: contour detection and contextual modulation. In particular, we identify a biophysically plausible neuronal model which predicts human behaviour in contour integration experiments to almost 100%. A similar approach allows to find a coherent explanation for a range of experimental results concerning the modulation of neuronal responses to stimuli inside their receptive fields by stimulus context. We will demonstrate that the probabilistic framework not only provides a functional interpretation of empirical observation, but also suggests specific neural mechanisms supporting these functions.