INCORPORATING TEST INPUTS INTO LEARNING

Student: Zehra Cataltepe
Faculty: Prof. Yaser S. Abu-Mostafa
Sponsoring Organizations: NSF

In some applications (credit card approval, medical image diagnosis), in addition to the training data, we know the inputs where the system will be tested. We demonstrate a method that incorporates test inputs into learning. The method performs better than using the training data only, especially for noisy problems or when the training set size is small.

The improvement in test error for a real problem ( liver data ) is illustrated in the following figure:

Detailed Report