Abstract
Stimulation of the extra-classical surrounds of receptive fields leads to highly nonlinear effects not explainable within the classical linear filter paradigm. We investigate these effects from two perspectives. On the one hand, we study a nonlinear multi-layer architecture which consists of linear filter mechanisms and simple ON/OFF nonlinearities. The filter mechanisms in each layer are learned (PCA, ICA) to yield an optimal adapation to the statistical properties of natural scenes. The resulting nonlinear processing properties are then compared with recent neurophysiological data on extra-classical properties. The same is done in an alternative approach by a class of nonlinear models that is based on Volterra-Wiener theory and generalized measures from differential geometry. Here the basic nonlinear interactions are AND-like (multiplicative) and the selected nonlinear combinations are intended to yield a higher-order whitening of the polyspectra of natural images. We show that both types of nonlinear architectures can capture basic extra-classical properties, like surround suppression effects, and we reveal the existence of equivalence classes in which one and the same input-output behavior can be obtained by models which differ apparently substantial in their structure. From a theoretical perspective, the concept of AND-like combinations of spatial frequency components is an attractive nonlinear extension of the classical linear spatial filter paradigm, since it is quite general and comprises, for example, also all linear-nonlinear-linear schemes with squaring rectifiers. Regarding the actual implementation of models, explicit AND operations seem to yield a somewhat more general class, since they enable a higher sensitivity and can describe subtle effects, like the components of intracellular recordings, or the complete cancellation of inhibitory influences, which are difficult to understand in terms of nonlinear inhibition schemes.