This paper addresses the challenges of multi-task learning, particularly the optimization difficulties that arise when multiple tasks are learned simultaneously. The authors identify three key conditions that cause detrimental gradient interference: conflicting gradients, high positive curvature, and large differences in gradient magnitudes. To mitigate these issues, they propose a method called "Projecting Conflicting Gradients" (PCGrad), which projects a task's gradient onto the normal plane of another task's gradient if they conflict. This approach is model-agnostic and can be combined with existing multi-task architectures. The paper provides theoretical analysis and empirical results showing that PCGrad significantly improves data efficiency and performance in both supervised and reinforcement learning tasks. The authors also discuss the broader impact of their work, highlighting potential applications in computer vision, autonomous driving, and robotics, while acknowledging the risks associated with machine learning systems.This paper addresses the challenges of multi-task learning, particularly the optimization difficulties that arise when multiple tasks are learned simultaneously. The authors identify three key conditions that cause detrimental gradient interference: conflicting gradients, high positive curvature, and large differences in gradient magnitudes. To mitigate these issues, they propose a method called "Projecting Conflicting Gradients" (PCGrad), which projects a task's gradient onto the normal plane of another task's gradient if they conflict. This approach is model-agnostic and can be combined with existing multi-task architectures. The paper provides theoretical analysis and empirical results showing that PCGrad significantly improves data efficiency and performance in both supervised and reinforcement learning tasks. The authors also discuss the broader impact of their work, highlighting potential applications in computer vision, autonomous driving, and robotics, while acknowledging the risks associated with machine learning systems.