Gradient normalization systems and methods for adaptive loss balancing in deep multitask networks

摘要

Systems and methods for training a multitask network is disclosed. In one aspect, training the multitask network includes determining a gradient norm of a single-task loss adjusted by a task weight for each task, with respect to network weights of the multitask network, and a relative training rate for the task based on the single-task loss for the task. Subsequently, a gradient loss function, comprising a difference between (1) the determined gradient norm for each task and (2) a corresponding target gradient norm, can be determined. An updated task weight for the task can be determined and used in the next iteration of training the multitask network, using a gradient of the gradient loss function with respect to the task weight for the task.