Abstract
Learning with noisy labels is a crucial task for training accurate deepneural networks. To mitigate label noise, prior studies have proposed variousrobust loss functions, particularly symmetric losses. Nevertheless, symmetriclosses usually suffer from the underfitting issue due to the overly strictconstraint. To address this problem, the Active Passive Loss (APL) jointlyoptimizes an active and a passive loss to mutually enhance the overall fittingability. Within APL, symmetric losses have been successfully extended, yieldingadvanced robust loss functions. Despite these advancements, emergingtheoretical analyses indicate that asymmetric losses, a new class of robustloss functions, possess superior properties compared to symmetric losses.However, existing asymmetric losses are not compatible with advancedoptimization frameworks such as APL, limiting their potential andapplicability. Motivated by this theoretical gap and the prospect of asymmetriclosses, we extend the asymmetric loss to the more complex passive loss scenarioand propose the Asymetric Mean Square Error (AMSE), a novel asymmetric loss. Werigorously establish the necessary and sufficient condition under which AMSEsatisfies the asymmetric condition. By substituting the traditional symmetricpassive loss in APL with our proposed AMSE, we introduce a novel robust lossframework termed Joint Asymmetric Loss (JAL). Extensive experiments demonstratethe effectiveness of our method in mitigating label noise. Code available at:https://github.com/cswjl/joint-asymmetric-loss