Computational Neuroscience Seminar - LCN
11.09.09 Friday, 12h15, AAB032
Boris Vladimirskiy, Department of Physiology, Bern University, Switzerland (Homepage)
Hierarchical novelty-familiarity coding in the early visual cortex
Abstract:
Information processing in the visual system can be viewed as driven by
the statistical structure in natural stimuli due to evolutionary
adaptation processes. Predictive coding, in which population feedback
from higher areas carries expectations of lower-level activity
(familiarity signal), whereas the population feedforward (novelty)
signals carry discrepancies between the expectations and the stimuli,
has been suggested to serve as an organizing principle for the entire
cortical hierarchy (Lee and Mumford, 2003; Friston, 2005). Modeling
results (Olshausen and Field, 1996, 1997; Rao and Ballard, 1999) have
shown that the statistics of natural images can explain some receptive
field (RF) properties, but no attempt had been made at coding entire
images. Furthermore, the proposed connectivity appeared too slow to be
biologically compatible and unrealistic RFs with only 3 small patches of
5 natural images were used. We investigate how good predictive coding
actually is at coding entire natural stimuli using a natural topographic
connectivity and a hierarchy of processing levels, each effectively
performing fast visual recognition. Our neural network model is based on
the principle of prediction error minimization, natural to expect of an
organism in order for it to survive, and is neurobiologically feasible.
The network is trained on 1000 natural images, following which the
coding performance is evaluated on a set of 200 different images.
Despite a compression factor of 4 for each level, the image
reconstruction quality is quite good and strongly exceeds that of local
averaging, implying that the learning results in the extraction of
features characteristic of the set of natural images as a whole. With
our model, we are able to reproduce several classical and
extra-classical receptive-field effects in V1. Most importantly, the
proposed architecture allows for the simultaneous representation of
familiarity and novelty (e.g., a predator suddenly appearing in a
familiar scene) at several spatial scales and could be an effective way
for the visual system to combine fast hierarchical visual recognition
with higher information processing, such as providing a read-out signal
for attention or escape behavior, in the brain. back |