A framework for multi-task learning optimization in deep neural networks : Balancing task priorities for improved performance
Ravindra K. Moje, Bhavana S. Tiple, Shankar M. Patil, Tushar Jadhav, Aniket Prakashrao Munshi, Akshay Revekar · Journal of Information and Optimization Sciences · 2025
Deep neural networks (DNNs) have made multi-task learning (MTL) a powerful way to learn multiple linked tasks at once, using shared models to improve generalization and speed. But successfully matching task goals is still one of the biggest problems in MTL, since giving all tasks the same level of value often results in poor performance. To solve this problem, we suggest a new way to improve MTL in DNNs: job priorities should be balanced automatically based on how important they are. Our framework includes a task priority balance system that changes how much each job contributes to the total loss function as training goes on. This system uses measures that are relevant to the job, like how hard it is and how much data is available, to automatically assign resources and set learning priorities. By managing task priorities well, our system lets DNNs focus more on important tasks while giving the right amount of resources to less important ones, which makes all tasks run better overall. We show that our method works by doing a lot of tests on a bunch of different standard datasets from different fields. The results show that our system regularly does better than other MTL methods by being more accurate and more general, especially when job difficulties are uneven. Overall, our framework offers an adaptable and effective way to enhance MTL in DNNs, boosting performance by constantly balancing job priorities.