Neural Network Models and the Visual Cortex: The Missing Link between Orientation Selectivity and the Natural Environment

Abstract

Orientation selectivity is a basic property of neurones in the visual cortex of higher vertebrates. Such neurones can be seen to act as 'feature detectors', which provide an efficient cortical representation of the outside world. More recently, the removal of correlations between the signals of cortical neurones has been suggested as suitable theoretical concept for explaining the development of receptive fields. Corresponding neural network simulations yielded oriented 'receptive field' structures resembling those observed by neurophysiologists. The findings suggest that the 'decorrelation approach' can provide a causal relationship between characteristics of the physical world and brain function. However, we were able to reveal a basic deficit of the decorrelation approach which we illustrate by the construction of two artificial 'worlds', a 'Gaussian' one and an 'orientation-only' one. We show that, according to the decorrelation approach, oriented environmental features would be neither necessary nor sufficient for the development of oriented receptive fields. Thus the link between environmental structure and cortical orientation selectivity still awaits a theoretical explanation.