a very similar representation is derived. Features
defined in terms of simple functionals of an edge input layer are used to represent
components of the face, making use of the statistics of the edges on the training set.
They model only three locations (two eyes and mouth) on the object, which are
chosen by the user, training multiple features for each. In the approach suggested
here, many more locations on the object are represented using one feature for
each. These locations are identified automatically in training. One problem with
over-dedicating resources to only three locations is the issue of noise and occlusion
which may eliminate one of the three. Our representation for faces for example,
is at a much lower resolution, e.g. 14 pixels between the eyes on average. At