在社会生产生活智能化应用场景中，复杂智能系统通常由多个子任务组成，和单任务相比，多任务学习会给我们带来对于目标系统更有洞见性的理解。然而，如何构建有效的多任务学习模型？目前缺少充分的理论支持。现有模型的性能、可解释性和不确定性表示能力还有待提高。高斯过程作为自适应非参机器学习的主流方法之一，在构造多任务学习模型上有很多优点。本研究将基于高斯过程回归网络和具有核成分依赖结构的高斯过程，设计一种新结构的耦合高斯过程回归网络解决多任务学习问题，并创新性地考虑高斯过程回归网络中神经元的耦合机制，包括：1）高斯过程回归网络中神经元的依赖结构分析；2）神经元及其权重的耦合机制表示和建模；3）神经元的耦合形式、耦合程度的量化评估；4）神经元的耦合机制的可解释性和可视化。通过本项目的理论研究与实验验证，有望提高高斯过程对复杂多任务的解释和学习能力，对增强复杂多任务系统的理解和提升其智能化水平具有重大价值。

In the intelligent application scenarios of social production and life, complex intelligent systems usually consist of multiple sub-tasks. Compared with single-task, multi-task learning will give us a more insightful understanding of the target system. However, how to build an effective multi-task learning model? At present, there is a lack of sufficient theoretical support. The performance, interpretability and uncertainty representation capabilities of existing models need to be improved. As one of the main methods in adaptive nonparametric machine learning, Gaussian process (GP) has many advantages in constructing multi-task learning model. Based on the Gaussian process regression network (GPRN) and the GP with dependent kernel structure, this research will design a new structured coupled GPRN to solve the problem of multi-task learning, and innovatively consider the coupling mechanism of neurons in the GPRN. The research includes: 1) dependency structure analysis of neurons in GPRN; 2) representation and modeling of coupling mechanism between neurons and their weights; 3) quantitative evaluation of coupling form and coupling degree of neurons; 4) interpretability and visualization of coupling mechanism of neurons. Through the theoretical research and experimental verification of this project, it is expected to improve the interpretation and learning ability of GPs for complex multi-tasking. It is of great value to enhance the understanding of complex multi-task systems and improve their intelligence performance.