RECEPTIVE FIELD
MODELS FAIL
TO PREDICT RESPONSES OF V1 NEURONS TO NATURAL MOVIES
Bruno A.
Olshausen1,
Jonathan L. Baker2, Shih-Cheng Yen2,
Charles M. Gray2
1Center
for
Neuroscience, UC Davis, Davis, CA, USA
2Department
of
Cell Biology and Neuroscience, Montana State University, Bozeman, MT,
USA
Much effort has
gone into
characterizing the receptive field properties of visual cortical
neurons.
However, the extent to which receptive field models are capable of
accounting
for the responses of visual neurons to natural scenes is largely
unexplored. In
this study, we compare the activity of V1 neurons in the anaesthetized
cat,
obtained in response to a natural movie, to the predictions of a simple
receptive field model based on taking a weighted sum of image pixels in
space
and time. For each neuron, the weights of the receptive field model are
derived
from the reverse correlation map obtained in response to an M-sequence.
For
those neurons having clear substructure in their receptive field
(simple
cells), the predicted response for a natural movie is obtained by
computing the
spatial inner-product between the receptive field weights and the movie
and
convolving over time. This is then compared with the actual response,
as
measured from the PSTH in response to multiple repetitions of a 30
second
natural movie sequence. We find that many V1 neurons exhibit brisk,
punctate
responses to events occuring in the movies that are highly reliable
across
trials. Most of these events are unpredicted by the receptive field map
obtained from the M-sequence. Even when the receptive field model is
over-fitted to the movie response, many events still remain unpredicted
by the
receptive field. The results suggest that the responses of V1 neurons
are
subject to strong non-linearities, possibly mediated by intracortical
circuitry, that are evoked by the structure of time-varying natural
images.
Support
Contributed By: NEI
R01-EY12478