Nonlinear properties of V2 and V4 neurons emerge in multi-layer networks trained with natural images

by C. Zetzsche, U. Nuding, K. Schill

Abstract:

We investigated nonlinear multi-stage networks optimised to reduce statistical dependences in natural images. These networks serve as models for the neural information processing in the higher visual areas of primates (visual cortices V2, V4). We analysed the resulting units with regard to nonlinear selectivity and invariance properties. We show that the proposed network principle leads to units that are highly selective with respect to the input signal space and to units that are invariant with respect to certain stimulus classes. The stimulus selectivity is tested by a set of Cartesian, hyperbolic, and polar gratings as used in several physiological experiments. We compared the population histogram of the stimulus selectivity indices of the network units with the population histogram of macaque V2 units and show that the network can achieve the high selectivity observed in many V2 neurons. We also tested for nonlinear frequency interaction effects, and found substantial interactions for frequencies with the same orientation and for frequencies with different orientations. The former enable a selective encoding of edges and the latter are required for the selective encoding of intrinsically two-dimensional features like corners and terminators

Reference:

Nonlinear properties of V2 and V4 neurons emerge in multi-layer networks trained with natural images (C. Zetzsche, U. Nuding, K. Schill), In Perception, volume 35, 2006.

Bibtex Entry:

@Article{Zetzsche2006, author = {C. Zetzsche and U. Nuding and K. Schill}, title = {Nonlinear properties of V2 and V4 neurons emerge in multi-layer networks trained with natural images}, journal = {Perception}, year = {2006}, volume = {35}, abstract = {We investigated nonlinear multi-stage networks optimised to reduce statistical dependences in natural images. These networks serve as models for the neural information processing in the higher visual areas of primates (visual cortices V2, V4). We analysed the resulting units with regard to nonlinear selectivity and invariance properties. We show that the proposed network principle leads to units that are highly selective with respect to the input signal space and to units that are invariant with respect to certain stimulus classes. The stimulus selectivity is tested by a set of Cartesian, hyperbolic, and polar gratings as used in several physiological experiments. We compared the population histogram of the stimulus selectivity indices of the network units with the population histogram of macaque V2 units and show that the network can achieve the high selectivity observed in many V2 neurons. We also tested for nonlinear frequency interaction effects, and found substantial interactions for frequencies with the same orientation and for frequencies with different orientations. The former enable a selective encoding of edges and the latter are required for the selective encoding of intrinsically two-dimensional features like corners and terminators}, }