Christoph Zetzsche
Universität Bremen
Nonlinear Neural Selectivity and the Statistics of Natural Scenes
Monday 11th of August 2014 at 02:00pm
560 Evans - ** note different time **
The classic linear filter model has been tremendously successful in
advancing our understanding of visual perception and of the neural
machinery of primary visual cortex. Its key feature is the selectivity
to spatial frequencies as obtained by wavelet-like bandpass filter
mechanisms with different sizes and orientations. All visual processing
beyond this linear stage remained somewhat enigmatic, however, and
despite a wealth of empirical information about various nonlinear
phenomena in visual perception and cortical processing we have never
again achieved a consensus like that for the standard model about a
suitable approach to the interpretation and modeling of these nonlinear
phenomena. In this talk I will suggest two basic nonlinear properties
that may be of help for structuring our understanding of nonlinear
vision. Both address the key issue of selectivity: In how far are the
nonlinear operations different from the selectivity obtained with linear
filters? The first aspect refers to the "tuning width" of selectivity as
determined by the shape of the nonlinear response surface in state space.
This response surface is curved, thereby increasing selectivity beyond
that obtainable with the planar response surfaces of linear filters. We
will show that this property can explain phenomena like the
overcompleteness of the representation in primary visual cortex, as well
as certain strange extraclassical receptive field effects, for example
the context-dependent switch from a suppressive to a facilitatory
influence of a given stimulus component. The second property is a new
concept for the "tuning direction": Beyond the selectivity for spatial
frequencies that is provided by linear filter mechanisms, the nonlinear
operations can provide a selectivity for AND-like combinations of
frequencies. If the frequencies differ in orientation this amounts to a
basic dimension of visual encoding: the nonlinear selectivity for
intrinsically two-dimensional signals. We show that this type of
selectivity can provide a unifying explanation for a large class of
recent observations on the selectivity of neurons in visual cortex.
Join Email List
You can subscribe to our weekly seminar email list by sending an email to
majordomo@lists.berkeley.edu that contains the words
subscribe redwood in the body of the message.
(Note: The subject line can be arbitrary and will be ignored)