Flaw in computation of specificity

[marcelluethi]

Specificity is currently computed by comparing random samples to the training examples
as represented in the model (i.e. given by the scores). This is done in order to avoid alignment problems. In this approach the test error can be made smaller by having a less expressive model (i.e. a model with 0 parameter would always yield 0 error, as the only training sample in the model representation is the mean).

A better solution might be to align the model first with the training images and use those for the validation.