Robust Learning with Noisy Label Detection and Counterfactual Correction

Wenting Qi, Charalampos Chelmis

Abstract: Data quality is of paramount importance in the training process of any machine learning model. Recently proposed methods for noisy learning focus on detecting noisy labeled data instances by using a fixed loss value threshold, and exclude detected noisy data instances in subsequent training steps. However, a predefined, fixed loss value threshold may not always be optimal, and excluding the detected noisy data instances can hurt the size of the training set. In this paper, we propose a new method, NDCC, that automatically selects a loss threshold to identify noisy labeled data instance, and uses counterfactual learning to repair them. To the best of our knowledge, NDCC is the first work to explore the feasibility of using counterfactual learning in the noisy learning domain. We demonstrate the performance of NDCC on Fashion–MNIST and CIFAR–10 datasets under a variety of label noise environments. Experimental results show the superiority of the proposed method compared to the state–of–the–art, especially in the presence of severe label noise.