Seeing in color is a primordial aspect of our visual experience. Despite its
importance, it is still misunderstood what exact purpose our color vision serves.
A common belief is that object recognition, crucial to our survival, is a core
driving force in the development of our visual system, and our color perception
by extension. And indeed, color is known for improving our ability to recognize
objects.
In this thesis, I explored the limits to which Deep Neural Networks, optimized
for object recognition or color constancy, can explain and help us understand
our color vision. Using advanced feature visualization, stimuli generation and
representational analysis methods, I carefully examined the color vision of these
trained models, also comparing their artificial responses to biological visual
systems.
I find that both artificial and biological systems exhibit some striking differences,
but these are outweighed by the sheer number of similarities. These
similarities include (1) large computing power for the processing of color, (2)
single and double opponent units in their early processing stages, (3) more sensitivity
to variations in hue than saturation, and (4) color representations that
follow similar perceptual dimensions. Despite their limitations, Deep Neural
Networks can thus astonishingly explain many color properties of our visual
system. This thesis hence provides evidence that our color vision is largely
shaped for and motivated by a feedforward recognition of natural objects and
their surface colors.
Verknüpfung zu Publikationen oder weiteren Datensätzen
