基于最大化时序可分性的序列数据特征变换理论和方法研究

Sequence feature representation is one of the core problems of intelligent sequence data analysis and processing, and the related feature space transformation for sequence data has far-reaching research significance in both theory and practice. Feature space transformation for sequence data (FSTS) transforms the features in sequences into a discriminative low-dimensional subspace to overcome the curse of dimensionality and improve the representation capability of sequence features. Conventional feature space transformation methods are based on the statistical independence assumption, and hence can not be directly applied to sequence features which have temporal dependencies. Therefore, FSTS by exploiting the temporal structures of sequences remains a challenging problem. Research in this area has not yet commenced. This research systematically studies the following contents on the theories and methods of FSTS from the perspective of maximizing the global temporal separability of the sequence classes: studying the methods for measuring the sequence statistics and separability between sequence classes; establishing the objective formulations of maximizing the separaiblity of different sequence classes in the underlying subspace under different discriminant criteria; analyzing the methods for the representation, approximation and inference of the latent alignments in the underlying subspace, and developing the optimization algorithms for each objective. The production of the research will provide important theoretical basis and methodological support for some fundamental problems in machine learning such as high-level sequence feature learning and optimization with latent structures, as well as some advanced applications in computer vision such as video action recognition.

序列特征表达是序列数据智能分析处理的核心研究问题之一，与之相关的序列特征变换在理论和应用方面具有重要的研究价值。序列特征变换将序列中的特征变换到一个具有强鉴别性的低维子空间，从而克服维数灾难并提高序列特征的表达能力。传统的特征降维方法以统计独立性假设为基础，无法直接应用到具有时序依赖关系的序列特征中，因此如何利用序列的时序结构对序列特征进行变换是个难点问题，这方面的研究还尚未展开。本项目从最大化时序可分性的角度系统地研究监督序列特征变换的理论和方法：研究序列统计量和序列类别可分性的度量方法；建立不同准则下最大化潜在子空间中序列类别可分性的目标函数；分析潜在子空间中隐含对齐结构的表达、近似和推理方法，并建立各个学习目标的优化算法。研究成果将为高层序列特征学习、隐含结构约束下目标优化等机器学习基础问题以及视频动作分类等计算机视觉前沿应用提供重要的理论依据和方法支撑。

序列特征变换的目标是从训练序列数据中学习一个变换将序列中的特征映射到具有强鉴别性的低维子空间，从而克服维数灾难并提高序列特征的表达能力。本项目从最大化时序可分性的角度深入研究了监督序列数据特征变换的理论和方法。在序列类别可分性度量方面，分别用不同方式的时序对齐对序列的时序结构进行建模，提出了基于产生式模型、基于序列距离和基于质心的序列统计量，进而构造了多种序列类间距离和类内散度的度量方法。在基于整体鉴别的目标函数方面，通过分析不同的时序可分性度量和隐含对齐结构在潜在子空间中的形式化表达，建立了不同鉴别准则下最大化时序可分性的目标函数。在含有隐含结构的目标函数优化方面，研究了不同学习目标下隐含时序对齐结构的近似和推理方法，分别建立了基于特征分解、半定规划和交替迭代的优化算法。项目提出了一系列新的时序结构建模和序列特征变换方法，并在多个计算机视觉任务中进行了应用。项目在研期间发表了一系列CCF A类国际期刊和会议论文，研究成果对相关领域的研究进展和实际应用起到了一定的推动作用。