This paper addresses the challenge of multi-task learning (MTL) by casting it as a multi-objective optimization problem. MTL involves solving multiple tasks jointly, sharing inductive bias between them, which can lead to conflicting objectives. The authors propose an efficient algorithm for MTL that scales to large-scale learning problems and achieves Pareto optimality. They introduce an upper bound on the multi-objective loss and show that optimizing this bound can be done with a single backward pass, significantly reducing computational overhead. The method is evaluated on various MTL problems, including digit classification, scene understanding (semantic segmentation, instance segmentation, and depth estimation), and multi-label classification, outperforming existing methods. The paper provides a comprehensive theoretical foundation and empirical validation, demonstrating the effectiveness of the proposed approach in handling complex MTL scenarios.This paper addresses the challenge of multi-task learning (MTL) by casting it as a multi-objective optimization problem. MTL involves solving multiple tasks jointly, sharing inductive bias between them, which can lead to conflicting objectives. The authors propose an efficient algorithm for MTL that scales to large-scale learning problems and achieves Pareto optimality. They introduce an upper bound on the multi-objective loss and show that optimizing this bound can be done with a single backward pass, significantly reducing computational overhead. The method is evaluated on various MTL problems, including digit classification, scene understanding (semantic segmentation, instance segmentation, and depth estimation), and multi-label classification, outperforming existing methods. The paper provides a comprehensive theoretical foundation and empirical validation, demonstrating the effectiveness of the proposed approach in handling complex MTL scenarios.