Neurobiological Background - Hierarchical Neural Networks for Image Interpretation - page 34

While the exact role of the feedback connections is not yet fully understood,
it can be safely stated that there is strong evidence that recurrent connections are
crucial for the performance of the visual system. This is not obvious when probing it
with isolated, high contrast stimuli which can be processed in a single feed-forward
sweep. However, in the regular mode of operation, natural visual stimuli contain
much ambiguity, e.g. due to occlusions, low contrast, and noise. In these situations,
feedback is used to bias low-level decisions based on attention and on the context of
partial scene interpretations. It seems that only after the match between higher-level
models and the low-level visual stimuli, a visual perception is relayed to prefrontal
areas and becomes conscious.
2.7 Conclusions
In Chapter 4, an architecture for computer vision will be introduced that is motivated
by the ventral pathway of the human visual system. It resembles key features of that
system, such as:
- computation by simple processing elements arranged in layers,
- retinotopic organization of interwoven feature maps,
- local recurrent connection structure with specific excitation and unspecific inhi-
bition,
- hierarchy of representations with increasing feature complexity, receptive field
size, invariance, and number of features,
- iterative refinement of image interpretation,
- integration of top-down, bottom-up, and lateral influences, and
- adaptation to the statistics of visual stimuli through learning.
Not all aspects covered in the previous chapter will be used in the remainder of
the thesis. For instance, the proposed architecture focuses on the ventral processing
stream and does not reflect the dorsal processing. Furthermore, eye movements, the
log-polar mapping between the retina and V1, and color processing are not investi-
gated, although they are important for the performance of the human visual system.
The reason for this restriction is that coverage of all these aspects would compete
for the available resources with the in-depth discussion of the selected aspects.
It is also important to note that the degree of biological realism in the remainder
of the thesis will only be very limited. The aim of the proposed architecture is not to
model neurobiological data, as is done in computational neuroscience, but to solve
computer vision problems based on inspiration from the human visual system.