Natural Image Statistics and Divisive Normalization:
Modeling Nonlinearity and Adaptation in Cortical Neurons
Published in:
Statistical Theories of the Brain
Eds. R Rao, B Olshausen, and M Lewicki
To appear, 2001.
© MIT Press, 2001.
We have empirically examined the responses of multi-scale oriented basis
functions to natural images, and found that these responses exhibit
striking statistical dependencies, even when the basis functions are
chosen to optimize
independence (e.g.,
simoncelli97b,
buccigrossi97). Such
dependencies cannot be removed through further linear processing.
Rather, a nonlinear form of cortical processing is required, in which
the linear response of each basis function is rectified (and typically
squared) and then divided by a weighted sum of the rectified responses
of neighboring neurons. In earlier work, we have shown that this
model, with all parameters determined from the statistics of a set of
natural images, can account qualitatively for recent physiological
observations of suppression of V1 responses by stimuli presented
outside the classical receptive field
(simoncelli98d). Here, we
show that the model can account for responses to non-optimal stimuli.
In addition, we show that adjusting the model parameters according to
the statistics of recent visual input can account for
physiologically observed adaptation
effects.
Full Text (498k, ps.gz)
/
Full Text (138k, pdf)
/ Online Publications