Another reason to explore the space of PCA/ICA comparisons is to provide data
for the current debate over global versus local features. The basis vectors that define
any subspace can be thought of as image features. Viewed this way, PCA and ICA
architecture II produce global features, in the sense that every image feature is influenced
by every pixel. (Equivalently, the basis vectors contain very few zeroes.) Depending
on your preference, this makes them either susceptible to occlusions and
local distortions, or sensitive to holistic properties. Alternatively, ICA architecture
I produces spatially localized features that are only influenced by small parts of
the image. It has been argued that this will produce better object recognition, since
it implements recognition by parts [27]. If localized features are indeed superior, ICA
architecture I should outperform PCA and ICA architecture II.