INCORPORATING TEST INPUTS INTO LEARNING
MOTIVATION: In learning-from-examples we are given a training set and our aim is to find a model having the smallest error on an unknown test set. In some applications (credit card approval, medical image diagnosis), we know the test inputs.
Our GOAL is to incorporate the test inputs into learning in order to achieve a smaller test error.
METHOD: We incorporate test inputs into learning by adding a term to the usual training error. The augmented term is scaled by a parameter which is determined by the data at hand.
The augmented error prevents the model from overfitting noisy training data.
ACHIEVEMENTS: Our method:
PUBLICATIONS: