Prototypical Classifier for Robust Class-Imbalanced Learning

Deep neural networks have been shown to be very powerful methods for many supervised learning tasks. However, they can also easily overfit training set biases, i.e., label noise and class imbalance. While both with noisy labels class-imbalanced received tremendous attention, existing works mainly focus on one of these two biases. To fill the gap, we propose Prototypical Classifier, which does not require fitting additional parameters given embedding network. Unlike conventional classifiers that are biased towards head classes, Classifier produces balanced comparable predictions all classes even though is class-imbalanced. By leveraging this appealing property, detect by thresholding confidence scores predicted where threshold dynamically adjusted through iteration. A sample reweighting strategy then applied mitigate influence labels. We test our method benchmark real-world datasets, observing obtains substaintial improvements compared state arts.