多源数据判别特征学习的结构化自编码方法

Video image data analysis is a frontier research topic in the field of artificial intelligence and machine learning. Complex backgrounds and diverse variations make the tasks including feature extraction, fusion and matching very difficult. The characteristics of multi-source, heterogeneity and label deficiency further complicate the data analysis tasks. .This project studies structured auto-encoder methods for intelligent analysis of multi-source video image data. The main research contents and innovations include: 1) It performs structured regularizations on the latent variables in the framework of variational auto-encoder networks, and then integrates the structured auto-encoders with adversarial training strategy to improve the discriminability of network features; 2) It introduces non-Euclidean measures into the triplet loss function. It also prunes the positive and negative sample pairs reasonably to simultaneously ensure classification accuracy and reduce the computational complexity. 3) It tackles the double transfer tasks, including both identity and style attributes, and improve classification accuracy of the multi-modal image data..This project will provide new learning framework and optimization algorithms for multi-source image feature extraction and discriminant classification. It will improve the discrimination, robustness and efficiency for data analysis methods further. Preliminary results also provide a good research basis for this project.

视频图像数据分析是人工智能与机器学习领域的前沿研究课题。数据背景复杂与变化多样使得特征提取、融合与匹配都十分困难，多源异构与标签缺失的情景进一步加剧了数据分析任务的难度。.本项目研究多源视频图像数据智能分析的结构化自编码方法，主要研究内容和创新之处包括：1）在变分自编码框架下对隐变量做出结构化约束，将结构化自编码与对抗式训练结合起来，提高输出特征的判别性；2）在三元组损失函数中嵌入非欧度量并合理剪枝正负样本对，保证分类精度的同时有效减小计算复杂度与存储成本；3）针对数据呈现的多模态特点，解决风格与身份的双重迁移学习与高效率计算问题，有效提高测试样本的分类精度。.本项目将为多源视频图像数据的特征提取与判别分类问题提供新的学习框架和优化算法，进一步提高数据分析方法的判别性、稳健性和高效性。前期成果也为本项目的开展提供了良好的研究基础。

项目研究了多源数据的判别特征学习问题，基于度量学习原理、最优传输原理和深度网络架构建立了结构化自编码框架，提出了类别属性的隐变量学习方法、局部正则化的度量学习方法和多源数据分析的自适应深度网络模型，实现了跨域视觉语义关联性刻画与局部结构对齐。建立了最优传输与条件分布嵌入之间的理论联系，得到了基于条件分布对齐的跨域泛化误差分析结果，并提出高效率的深度学习方法，阐明了最优传输代价动态更新与判别信息迁移之间的内在关联。提出了基于生成式训练策略的多源域自适应表征学习算法簇，揭示了多源视觉数据信息融合与不变量描述的低维表征学习规律，增强了模型在跨域或异分布场景下的泛化能力。 .项目已在国际高水平期刊发表论文2篇。