本项目聚焦非独立同分布噪声条件下的稳健机器学习理论与方法研究。传统机器学习方法基于独立同分布噪声假设来选择损失函数，因而对实际应用中的非独立同分布噪声数据常表现出不稳健、不有效等问题。本项目拟基于损失函数的自适应学习策略，构造非独立同分布噪声下稳健高效的机器学习方法，解决其中的关键理论问题，并应用于典型图像和视频处理问题。具体包括：(1) 基于噪声分布建模和自步学习，建立针对非独立同分布噪声的损失函数学习理论与方法；(2) 将损失函数学习模型嵌入若干典型机器学习任务，设计构造稳健、高效的机器学习方法；(3) 针对高光谱图像恢复和视频去雨去雪等典型含有非独立同分布噪声的问题，应用损失函数学习策略构造新型有效算法。所获结果将建立起完整的针对非独立同分布噪声的损失函数建模新框架，为机器学习方法的稳健性问题提供全新的解决方案，同时可直接应用于高光谱图像恢复与视频去雨去雪等广泛图像/视频处理问题。

This project focuses on the theory and methods of robust machine learning with non-i.i.d. noises. The existing machine learning methods select loss function based on the i.i.d. noise assumption, and thus are less robust and effective to non-i.i.d noises in real applications. To this issue, this project plans to construct machine learning methods that are robust and effective to non-i.i.d noises based on adaptive loss function learning, develop corresponding theory and apply them to typical image and video processing problems. Specifically, (1) base on noise modeling and self-paced learning, we plan to develop models and methods for loss function learning with non-i.i.d noises; (2) for some typical tasks, we plan to develop robust and effective machine learning methods by the proposed loss function learning models; (3) by the proposed methodology, we plan to develop effective algorithms for some typical applications with non-i.i.d noises, such as hyperspectral image restoration and rain and snow removal in videos. The desired achievements will give build a full framework for loss function modeling with non-i.i.d noises, and thus provide a new solution for the robustness problems in machine learning. Besides, it also has wide applications in image and video processing, such as hyperspectral image restoration and rain and snow removal in videos.