基于深度学习的高维时间序列预测方法及其视频控制的应用

As one of the most sophisticated learning approaches in machine learning area, deep learning is receiving more and more attentions from both academic and industrial sections. Deep model mimics the mechanism of the information processing of intelligent human brain, i.e., the high level data representation and pattern abstraction. Those patterns are highly informative and crucial for further decision making. In this project, we develop a new deep model and corresponding learning algorithm for high-dimensional time series prediction by proposing a deep network with temporal structure, and integrating human-like communication between different neurons. The merits of this project springs from three aspects as following: 1) the proposed temporal structure is very suitable for modelling the temporal coherence in high-dimensional time series; 2) the proposed model has deep architecture to imitate the way that human brain process information; 3) the proposed deep model is more robust as we also introduce fuzzy theory to the model in the natural way that neuron communicate with each other. Finally, object recognition and tracking problems in video control will be adopted to verify the performance of the proposed deep learning techniques.

深度学习被誉为最接近人脑的机器学习方法，近年来在学术界与产业界引起广泛的关注。深度模型主要是模拟人脑处理信息的机制，也就是多层逐层地提取原始数据的高阶数据表示和数据特征。这些提取出来的特征就是原始数据里面的关键信息，被用于进一步的决策判断。 最近的研究表明，这种信息处理机制是有效的。本项目通过提出与时间相关的模糊深度模型及其学习方法，对高维时间序列进行建模、分析与预测。通过提出时间性深度网络结构，并引入类似人脑神经元的通信方式，得到更加智能的深度模型。新提出的深度模型的优点体现在三方面：1）拥有适合于高维时间序列的模型结构，即时间性的结构；2）拥有类似人脑的信息处理方式，即深层网络模型；3）拥有类似人脑神经元的通信方式，即引入的模糊理论。高维时间序列在实际中有着广泛的应用，本项目计划用视频控制、视频中目标识别与追踪等例子对所提出的模型的有效性进行验证，同时发挥其实际应用价值。

项目主要研究视频数据的表示学习算法，并解决视频分类、描述、目标追踪等任务。现已发表论文四篇，处于修改状态论两篇，处于投稿状态两篇。该项目的研究成果具有一定的理论和实际应用价值。例如，提出的视频数据的无监督学习算法，很多程度上缩小了有监督学习与无监督学习之间的差距，同时获得的视频特征具有更好的泛化性能；提出了深度模型的有效压缩方法，能够极大减少深度模型的参数，使之能够在小型移动平台使用；提出的深度时空模型的目标追踪模型，能够实时准确的追踪目标等。