Samy Bengio, Quan Le
For classification problems, it is important that the classifier is trained with data which is likely to appear in the future. Discriminative models, because of their nature to focus on the boundary between classes rather than data itself, usually do not h ...
IDIAP2003
