EIF - Mislabelled sample detection accuracy for EIF versus IF

Mislabelled sample detection EIF vs IF

In mislabelled experiment. We inject 10% training data with wrong class labels. We then rank their Influences according to the confusion pairs correspond to the top1 wrong testing class. After ranking in ascending order, harmful samples get lower influence values, thus have higher rank. Our EIF is comparable in performance than IF in mislabelled data detection.

ProxyNCA++ Noisy Level = 0.01 ProxyNCA++ Noisy Level = 0.05 ProxyNCA++ Noisy Level = 0.1

SoftTriple Noisy Level = 0.01 SoftTriple Noisy Level = 0.05 SoftTriple Noisy Level = 0.1

Page updated

Google Sites

Report abuse