Leveraging convergence behavior to balance conflicting tasks in multi-task learning

Multi-Task Learning is a learning paradigm that uses correlated tasks to improve performance generalization. A common way learn multiple through the hard parameter sharing approach, in which single architecture used share same subset of parameters, creating an inductive bias between them during training process. Due its simplicity, potential generalization, and reduce computational cost, it has gained attention scientific industrial communities. However, often conflict with each other, makes challenging define how gradients should be combined allow simultaneous learning. To address this problem, we use idea multi-objective optimization propose method takes into account temporal behaviour create dynamic adjusts importance task backpropagation. The result give more are diverging or not being benefited last iterations, ensuring heading maximization all tasks. As result, empirically show proposed outperforms state-of-the-art approaches on conflicting Unlike adopted baselines, our ensures reach good generalization performances.