Max Planck Institute for Dynamics and Self-Organization -- Department for Nonlinear Dynamics and Network Dynamics Group
Personal tools
Log in

BCCN AG-Seminar

Tuesday, 29.11.2005 16 c.t.

Systematic errors of contour detection are predicted by computational mechanisms for contour integration

by Nadja Schinkel
from Universität Bremen

Contact person: J. Michael Herrmann


Seminarraum Haus 2, 4. Stock (Bunsenstr.)


Contour integration is a process fundamental to object recognition. However, its neural mechanisms are still not well understood. Contours can be characterised by an association field which describes conditional link probabilities between oriented edges. These link probabilities can be used to generate well defined contours or vice versa to extract contours from a stimulus. In a neurophysiological context, the association field is often identified with horizontal interactions between distant neurons in early visual areas. While in typical biologically motivated neural networks (e.g., Li, Sagi, Hess) this lateral synaptic input is added to the afferent input from the visual stimulus, probabilistic models multiply these two input sources in order to optimally compute the contour's saliency. It is therefore an open question which model class can explain human contour integration best. Extensive numerical simulations using stimuli generated by an association field, show similar performances for the probabilistic-multiplicative and the additive model and thus cannot rule out one model class. However, psychophysical experiments with humans reveal that contour detection errors are not made randomly, but are highly correlated among different subjects. Hence, a model describing contour integration in the brain should not only explain human contour detection performance but should also reproduce these systematic errors made by subjects. Comparison between misdetections of humans and mispredictions of the two models reveal a high correlation between the multiplicative model and human behaviour. For a wide range of biologically plausible model parameters investigated so far, the correlations between the additive model and humans are lower than those of the multiplicative model and humans, and also lower as the correlations between different human observers. Pending further studies this gives a strong hint for the involvement of multiplication in contour integration mechanisms of the human brain.

back to overview