Models of learning in the visual system: dependence on retinal eccentricity
In the primary visual cortex of primates relatively more space is devoted to the representation of the central visual field in comparison to the representation of the peripheral visual field. Experimentally testable theories about the factors and mechanisms which may have determined this inhomogeneo...
|PDF Full Text
No Tags, Be the first to tag this record!
|In the primary visual cortex of primates relatively more space is devoted to the representation of the central visual field in comparison to the representation of the peripheral visual field. Experimentally testable theories about the factors and mechanisms which may have determined this inhomogeneous mapping may provide valuable insights into general processing principles in the visual system. Therefore, I investigated to which visual situations this inhomogeneous representation of the visual field is well adapted, and which mechanisms could support its refinement and stabilization during individual development. Furthermore, I studied possible functional consequences of the inhomogeneous representation for visual processing at central and peripheral locations of the visual field.
Vision plays an important role during navigation. Thus, visual processing should be well adapted to self-motion. Therefore, I assumed that spatially inhomogeneous retinal velocity distributions, caused by static objects during self-motion along the direction of gaze, are transformed on average into spatially homogeneous cortical velocity distributions. This would have the advantage that the cortical mechanisms, concerned with the processing of self-motion, can be identical in their spatial and temporal properties across the representation of the whole visual field. This is the case if the arrangement of objects relative to the observer corresponds to an ellipsoid with the observer in its center. I used the resulting flow field to train a network model of pulse coding neurons with a Hebbian learning rule. The distribution of the learned receptive fields is in agreement with the inhomogeneous cortical representation of the visual field. These results suggest that self motion may have played an important role in the evolution of the visual system and that the inhomogeneous cortical representation of the visual field can be refined and stabilized by Hebbian learning mechanisms during ontogenesis under natural viewing conditions.
In addition to the processing of self-motion, an important task of the visual system is the grouping and segregation of local features within a visual scene into coherent objects. Therefore, I asked how the corresponding mechanisms depend on the represented position of the visual field. It is assumed that neuronal connections within the primary visual cortex subserve this grouping process. These connections develop after eye-opening in dependence on the visual input. How does the lateral connectivity depend on the represented
position of the visual field? With increasing eccentricity, primary cortical receptive fields become larger and the cortical magnification of the visual field declines. Therefore, I investigated the spatial statistics of real-world scenes with respect to the spatial filter-properties of cortical neurons at different locations of the visual field. I show that correlations between collinearly arranged filters of the same size and orientation increase with increasing filter size. However, in distances relative to the size of the filters, collinear correlations decline more steeply with increasing
distance for larger filters. This provides evidence against a homogeneous cortical connectivity across the whole visual field with respect to the coding of spatial object properties.
Two major retino-cortical pathways are the magnocellular (M) and the parvocellular (P) pathways. While neurons along the M-pathway display temporal bandpass characteristics, neurons along the P-pathway show temporal lowpass characteristics. The ratio of P- to M-cells is not constant across the whole visual field, but declines with increasing retinal eccentricity. Therefore, I investigated how the different temporal response-properties of neurons of the M- and the P-pathways influence self-organization in the visual cortex, and discussed possible consequences for the coding of visual objects at different locations of the visual field. Specifically, I studied the influence of stimulus-motion on the self-organization of lateral connections in a network-model of spiking neurons with Hebbian learning. Low stimulus velocities lead to horizontal connections well adapted to the coding of the spatial structure within the visual input, while higher stimulus velocities lead to connections which subserve the coding of the stimulus movement direction. This suggests that the temporal lowpass properties of P-neurons subserve the coding of spatial stimulus attributes (form) in the visual cortex, while the temporal bandpass properties of M-neurons support the coding of spatio-temporal stimulus attributes (movement direction). Hence, the central representation of the visual field may be well adapted to the encoding of spatial object properties due to the strong contribution of P-neurons. The peripheral representation may be better adapted to the processing of motion.