Title: Confidence Scores Make Instance-dependent Label-noise Learning Possible

Abstract: Learning with noisy labels has drawn a lot of attention. In this area, most of recent works only consider class-conditional noise, where the label noise is independent of its input features. This noise model may not be faithful to many real-world applications. Instead, few pioneer works have studied instance-dependent noise, but these methods are limited to strong assumptions on noise models. To alleviate this issue, we introduce confidence-scored instance-dependent noise (CSIDN), where each instance-label pair is associated with a confidence score. The confidence scores are sufficient to estimate the noise functions of each instance with minimal assumptions. Moreover, such scores can be easily and cheaply derived during the construction of the dataset through crowdsourcing or automatic annotation. To handle CSIDN, we design a benchmark algorithm termed instance-level forward correction. Empirical results on synthetic and real-world datasets demonstrate the utility of our proposed method.