Confrontation with Noise in Selective Classification
Haobo Xu · 2022 IEEE Conference on Telecommunications, Optics and Computer Science (TOCS) · 2022
Selective prediction is a powerful tool which enables machine learning models to make predictions with more confidence. In this paper, a selective classifier based on deep neural network (DNN) is applied for image classification tasks with training label noise or Out-of-distribution (OOD) data are implemented. Furthermore, the deep abstaining classifier (DAC) is introduced as a training data cleanser with the DNN model, and the performance of training a DNN after using DAC for cleansing, which is called post-DAC DNN, is compared with that of the original DNN classifier. It is shown that the post-DAC DNN significantly outperforms the baseline model either when the training data is uniformly corrupted or has randomized single-class label noise.