Enhanced task attention with adversarial learning for dynamic multi-task CNN

Highlights：

• We propose a novel learning framework of multi-task CNN to enhance task attention through tuning the TTC of the shared subnet DMT-CNN with adversarial learning. Supervision information introduced with adversarial learning makes each task subnet focus more on its target task without limiting the model flexibility.

• We enhance the task attention with a task discriminator shared by all layers of all subnets to improve the sharing mechanism.

• We design the even-label strategy to construct the adversarial learning structure in a dynamic multi-task model.