受限的玻尔兹曼机在背景建模和文本建模中的研究与应用

发布时间：2018-03-20 11:58

本文选题：受限玻尔兹曼机　切入点：背景建模　出处：《中国科学技术大学》2016年硕士论文　论文类型：学位论文

【摘要】：背景建模和文本建模分别是计算机视觉和自然语言处理两个人工智能领域关注的重点基础问题。这两个方向都是现今人工智能发展的前沿。背景建模旨在为视频序列生成一组稳定的背景,是前景检测、视频监控等应用的基础。传统的背景建模模型往往关注局部信息的利用,而实际上,在真实的复杂视频中,全局信息变化广泛存在,大到光照变化,小到物体阴影都随着时间发生变化。因此本文利用受限玻尔兹曼机针对全局信息进行建模,针对视频中相邻的两帧图片学习到视频的背景,并且设计了一种自适应的正则化项,能够强迫模型生成的背景保持稳定,也就使得模型能够准确找到背景信息。这是首次有科研工作将受限玻尔兹曼机应用于背景建模任务中。文本建模关注的是从文本中提取有用的信息,是自然语言处理领域的最常见的一类基础任务。本文关注的是近年来备受学术界关注的词向量模型工具包word2vec。目前,针对word2vec工具的解释尚缺严谨和充分的研究。由受限的玻尔兹曼机启发,本文设计了一种基于矩阵分解的模型——显式矩阵分解(EMF),并且从理论上证明了它与word2vec中的Skip-Gram Negative Sampling (SGNS)模型无条件等价,提出的模型相比于其他针对SGNS的解释工作有更好的解释性。此外,本文还从EMF模型的角度扩展出了一种有监督模型,能够结合单词类比数据,大大提高词向量在单词类比任务上的性能。
[Abstract]:Background modeling and text modeling are two important basic issues in the field of computer vision and natural language processing, respectively. These two directions are the frontier of the development of artificial intelligence. Background modeling is aimed at video. Sequence generates a stable set of backgrounds, Traditional background modeling models often focus on the use of local information, but in fact, in real complex video, global information changes widely exist, ranging from light changes. In this paper, the constrained Boltzmann machine is used to model the global information, and the background of the video is learned from the two adjacent frames of the video, and an adaptive regularization term is designed. The background that can force the generation of the model to remain stable, This is the first time that a limited Boltzmann machine has been applied to a background modeling task. Text modeling focuses on extracting useful information from text. Is the most common basic task in the field of natural language processing. This paper focuses on the word vector model toolkit word2vec. which has attracted much attention in recent years. The interpretation of the word2vec tool is not yet rigorous and adequate. Inspired by the restricted Boltzmann machine, In this paper, we design an explicit matrix factorization model based on matrix factorization, and prove that it is unconditionally equivalent to the Skip-Gram Negative Sampling model in word2vec. The proposed model is more explanatory than other interpretations of SGNS. In addition, this paper extends a supervised model from the perspective of EMF model, which can combine word analogy data. The performance of word vector in word analogy task is greatly improved.
【学位授予单位】：中国科学技术大学
【学位级别】：硕士
【学位授予年份】：2016
【分类号】：TP391.41

【相似文献】