Multi Task Pareto Optimization Algorithms for Balanced Performance Across Diverse Objectives

Abstract

A challenge of multitask learning is balancing competing tasks, especially for high-dimensional deep neural networks. Previous approaches often overlook the fact that there are multiple objectives and bias toward the dominant task, resulting in poor performance in all tasks. This study presents a Multi-Task Pareto Optimization Algorithm (MTPOA) combining gradient projection with dynamic task weighting to properly balance the performance of the different and heterogeneous tasks. The framework is tested on the depth estimation and semantic segmentation tasks, where the NYU Depth V2 dataset is used. Task-specific gradients are calculated, Pareto-efficient projections are made to resolve conflicts, and model parameters are adapted to keep the model converging while all objectives are optimized. Experimental results show that MTPOA improves over the basic approaches such as weighted sum, PCGrad, and MGDA. The depth RMSE gets reduced to 1.81 as compared to 2.15 (Weighted Sum), and the segmentation accuracy gets increased to 86.7% from 81.2%. Pareto front analysis validates the results of MTPOA solutions, which are found to be more Pareto efficient (0.74) and more hypervolume (0.52) compared to baseline solutions. The framework also shows better convergence, using 80 epochs for depth and 85 epochs for segmentation, which suggests that the framework is more stable to train. These results confirm MTPOA as a scalable, conflict-resistant multi-task deep learning approach that is efficient in terms of computational resources and maintains the high performance of each task. The study offers a solid theoretical basis for implementing the Pareto-based multi-task optimization in a variety of AI applications.