元数据驱动的损失项与正则项自动化学习方法

本项目聚焦元数据驱动的机器学习损失项与正则项自动化学习理论与方法研究。传统机器学习方法通常需要人为确定损失项与正则项的形式，无法充分反映数据的本质结构和采样偏差，限制了学习算法的性能。本项目拟基于元学习原理，构造元数据驱动的损失项与正则项学习模型与方法，解决其中的关键理论问题，并应用于若干典型实际问题。包括：（1）构建元数据驱动的损失项与正则项学习的一般模型与结合具体学习任务的求解算法；（2）从收敛性、稳健性、泛化性等方面出发，建立所提损失项与正则项学习方法的一般理论；（3）将所提策略推广至贝叶斯机器学习模型，构造元数据驱动的似然项与先验项学习方法；（4）针对智能通信、空对地目标检测、图像复原等典型实际问题，应用所提策略构造新型有效算法。所获结果将从理论、模型、算法与应用等多个层面建立一套完整的元数据驱动的损失项与正则项学习方法论，为机器学习模型在实际应用中的构建提供全新的自动化解决方案。

This project focuses on the theory and methods of automatically learning the loss and regularization terms, driven by meta-data, in machine learning models. The existing machine learning methods generally require manually specifying the loss and regularization terms, which cannot sufficiently reflect the intrinsic structure and sampling bias underlying data, and thus limits the performance of learning algorithms. To this issue, based on the principle of meta-learning, this project plans to construct automatic learning methods for the loss and regularization terms driven by meta-data, develop corresponding theories and apply them to typical real problems. Specifically, (1) we plan to develop the general models for automatically learning the loss and regularization terms driven by meta-data, and corresponding algorithms for specific machine learning tasks; (2) we plan to develop the general theories for the proposed methods, including convergence, robustness and generalization; (3) we plan to generalize the proposed strategy to Bayesian machine learning models, and correspondingly develop learning methods for the likelihood and prior terms; (4) by the proposed methodology, we plan to develop effective algorithms for some typical applications, such as intelligent wireless communication, ground target detection and image restoration. The desired achievements will build a full framework, from the aspects of theory, model, algorithm and application, for meta-data driven learning of the loss and regularization terms, and thus provide a new solution for automatically building machine learning models in real applications.