by C. Zetzsche, F. Roehrbein
Abstract:
The neural mechanisms of early vision can be explained in terms of an information-theoretic optimization of the neural processing with respect to the statistical properties of the natural environment. Recent applications of this approach have been successful in the prediction of the linear filtering properties of ganglion cells and simple cells, but the relations between the environmental statistics and cortical nonlinearities, like those of end-stopped or complex cells, are not yet fully understood. Here we present extensions of our previous investigations of the exploitation of higher-order statistics by nonlinear neurons. We use multivariate wavelet statistics to demonstrate that a strictly linear processing would inevitably leave substantial statistical dependencies between the outputs of the units. We then consider how the basic nonlinearities of cortical neurons - gain control and ON/OFF half-wave rectification - can exploit these higher-order statistical dependencies. We first show that gain control provides an adaptation to the polar separability of the multivariate probability density function (PDF), and, together with an output nonlinearity, enables an overcomplete sparse coding. We then consider how the remaining higher-order dependencies between different units can be exploited by a combination of basic ON/OFF point nonlinearities and subsequent weighted linear combinations. We consider two statistical optimization schemes for the computation of the optimal weights: principal component analysis (PCA) and independent component analysis (ICA). Since the intermediate nonlinearities transform some of the higher-order dependencies into second-order dependencies even the basic PCA approach is able to exploit part of the redundancies. ICA ignores this second-order structure, but can exploit higher-order dependencies. Both schemes yield a variety of nonlinear units which comprise the typical nonlinear processing properties, such as end-stopping, side-stopping, complex-cell properties and extra-classical receptive field properties, but the `ideal' complex cells seem only to occur with PCA. Thus, a combination of ON/OFF nonlinearities with an integrated PCA-ICA strategy seems necessary to exploit the statistical properties of natural images.
Reference:
Nonlinear and extra-classical receptive field properties and the statistics of natural scenes (C. Zetzsche, F. Roehrbein), In INET, Informa UK Limited, volume 12, 2001.
Bibtex Entry:
@Article{Zetzsche2001,
author = {C. Zetzsche and F. Roehrbein},
title = {Nonlinear and extra-classical receptive field properties and the statistics of natural scenes},
journal = {{INET}},
year = {2001},
volume = {12},
number = {3},
pages = {331--350},
month = {jan},
abstract = {The neural mechanisms of early vision can be explained in terms of an information-theoretic optimization of the neural processing with respect to the statistical properties of the natural environment. Recent applications of this approach have been successful in the prediction of the linear filtering properties of ganglion cells and simple cells, but the relations between the environmental statistics and cortical nonlinearities, like those of end-stopped or complex cells, are not yet fully understood. Here we present extensions of our previous investigations of the exploitation of higher-order statistics by nonlinear neurons. We use multivariate wavelet statistics to demonstrate that a strictly linear processing would inevitably leave substantial statistical dependencies between the outputs of the units. We then consider how the basic nonlinearities of cortical neurons - gain control and ON/OFF half-wave rectification - can exploit these higher-order statistical dependencies. We first show that gain control provides an adaptation to the polar separability of the multivariate probability density function (PDF), and, together with an output nonlinearity, enables an overcomplete sparse coding. We then consider how the remaining higher-order dependencies between different units can be exploited by a combination of basic ON/OFF point nonlinearities and subsequent weighted linear combinations. We consider two statistical optimization schemes for the computation of the optimal weights: principal component analysis (PCA) and independent component analysis (ICA). Since the intermediate nonlinearities transform some of the higher-order dependencies into second-order dependencies even the basic PCA approach is able to exploit part of the redundancies. ICA ignores this second-order structure, but can exploit higher-order dependencies. Both schemes yield a variety of nonlinear units which comprise the typical nonlinear processing properties, such as end-stopping, side-stopping, complex-cell properties and extra-classical receptive field properties, but the `ideal' complex cells seem only to occur with PCA. Thus, a combination of ON/OFF nonlinearities with an integrated PCA-ICA strategy seems necessary to exploit the statistical properties of natural images.},
doi = {10.1080/net.12.3.331.350},
publisher = {Informa {UK} Limited},
url = {10.1080/net.12.3.331.350">http://dx.doi.org/10.1080/net.12.3.331.350},
}