Multiple instance learning with noisy labels based on symmetry loss

Multiple instance learning (MIL) serves as a crucial form of weakly supervised learning, characterized by each training bag containing multiple instances with unknown labels, and each bag being assigned either a positive or negative label. The majority of prior studies have operated under the assumption that the labels in training datasets are entirely accurate. However, in many real-world scenarios, various factors such as limited knowledge may lead to labeling errors, resulting in label noise. Motivated by these challenges, this paper investigates MIL problems with label noise. To address this issue, we propose two frameworks: area under the receiver operating characteristic curve (AUC) maximization and balanced error rate (BER) minimization that highlight the advantages of incorporating symmetric loss. Moreover, we provide theoretical analysis to establish error bounds and ensure consistency. Finally, experimental results on seven datasets using six different losses not only demonstrate the effectiveness but also emphasize how introducing symmetric loss leads to superior performance. Furthermore, experiments on a convex barrier hinge loss further validate the significance of maintaining asymmetric condition despite its lack of symmetry everywhere.