GradNorm

A gradient balancing strategy in multi-task learning that dynamically adjusts the loss weights of each task so that gradient norms tend toward consistency, thereby alleviating gradient conflicts among tasks.