CASIA OpenIR  > 模式识别国家重点实验室  > 先进数据分析与学习
非对称哈希学习方法研究
达铖
Subtype博士
Thesis Advisor向世明 ; 孟高峰
2019-05-21
Degree Grantor中国科学院大学
Place of Conferral中国科学院自动化研究所
Degree Name工学博士
Degree Discipline模式识别与智能系统
Keyword哈希学习 非对称哈希 二值编码 多值编码 深度学习 监督学习 无监督学习
Abstract

       信息技术的高速发展推动全球数据总量的急剧攀升。在大数据时代中,信息检索由于高额的计算和存储开销面临巨大挑战。因此,如何快速、高效且精确地从海量数据中检索出特定样本是当前机器学习领域的研究热点。哈希学习作为一种常用的近似最近邻搜索方法,凭借其高效的检索效率和低额的存储消耗受到了研究者的广泛关注。

       非对称哈希学习方法构建不同的哈希函数来分别编码数据库样本和查询样本。因其灵活的编码机制、优异的检索性能和高效的优化方式,非对称哈希学习方法成为当前哈希学习领域的研究热点。本文开展非对称哈希学习方法相关研究,取得若干成果。论文的主要贡献包含以下几个方面:

 

  1. 提出了一种基于度量嵌入的非对称哈希学习方法。该方法的核心思想是采用非对称编码策略来减少数据库样本的量化损失,同时保留查询样本的实数值信息。具体地,论文构建一种无监督非对称哈希学习模型,采用实数值编码和二值编码分别表示查询样本和数据库样本,并使用双线性函数来度量非对称编码间的相似性,然后嵌入至哈希学习过程中。实验结果表明所提方法的检索性能优于传统的无监督哈希学习方法。
  2. 提出了一种非对称多值哈希学习方法。该方法的核心思想是采用多整数值编码和实数值编码来分别表征数据库样本和查询样本。具体地,论文构建了一个多整数值约束的非对称哈希学习模型,并通过二值稀疏表示将其转化为一个易于优化的混合整数规划问题。在监督学习框架下,学习得到具有优异相似度保持性能的非二值编码,并在维持高效查询和低耗存储的同时提高检索精度。对比实验结果验证了所提方法的有效性。
  3. 提出了一种非线性非对称多值哈希学习方法。该方法的核心思想是在非对称多值哈希的框架下利用神经网络来实现查询样本的实数值映射。为此,论文提出了一种交替优化策略来实现对神经网络参数的学习和混合整数规划问题的高效求解。在此基础上,提出了一种数据库增量扩充方法来解决数据库样本更新问题。同时,从理论上指出非对称多值哈希学习方法中的距离度量本质上是一种查询敏感的加权汉明距离度量。对比实验结果验证了所提方法的有效性。
  4. 提出了一种深度无监督图哈希学习方法。该方法的核心思想是将图哈希学习方法与深度神经网络训练融合到统一的学习框架中。具体地,提出了一种非对称谱损失来保持深度神经网络的输出与二值编码间的局部近邻结构关系。在此框架下,利用全局图拉普拉斯构建小批量样本梯度来驱动深度神经网络参数的学习,并从全局保持样本间的亲和度关系。为此,提出了一种高效优化算法交替完成深度神经网络的训练和离散二值编码的优化。实验结果表明所提方法优于现有的无监督哈希学习方法。
Other Abstract

       The fast development of information technology has promoted the sharp rise of global data volume. In the era of big data, information retrieval methods face great challenges due to high computing and storage costs. Therefore, how to retrieve specific samples from massive data quickly, efficiently and accurately has attract extensive attentions in the field of machine learning. Technically, hash learning, as a common approximate nearest neighbor search method, has been widely concerned by researchers due to its high retrieval efficiency and low memory consumption.

       Asymmetric hashing methods utilize different hash functions to encode query and database samples. Due to its flexible coding mechanism, excellent retrieval performance and efficient optimization strategy, some efforts have been devoted to delving into the merits of the asymmetric form in similarity search problem over the past years. This thesis studies the asymmetric hashing learning technique and attains some new methods and results. The main contributions and novelties of this thesis are listed as follows.

  1. A metric-embedded asymmetric hashing method (MEAH) is proposed. In this method, the asymmetric encoding strategy is utilized to reduce the quantization error of database points and preserve well the real-valued information of query points. Specifically, an unsupervised asymmetric hashing method is proposed to encode query and database points as real-valued codes and binary ones. Moreover, a bilinear function that measures the similarity between the asymmetric codes is embedded into the hashing learning framework. Extensive experiments demonstrate the superiority of our approach over the several traditional unsupervised hashing methods.
  2. An asymmetric multi-valued hashing method (AMVH) is proposed. The core idea of this method is to leverage multi-integer-embeddings and real-valued ones to represent database and query points, respectively. Specifically, we introduce binary sparse representation to model the multi-integer-valued embedding for database points. By this way, the problem with multi-integer constraints can be formulated as a mixed integer programming problem, which can be readily optimized by the proposed alternative optimization algorithm. Casting on the framework of supervised learning, our approach is able to learn non-binary embeddings with powerful capability of similarity preservation. Therefore, the retrieval accuracy can be improved while maintaining efficiency of query and storage. Extensive experiments verify the effectiveness of the proposed method.
  3. A nonlinear asymmetric multi-valued hashing method (NAMVH) is proposed, which leverages multi-layer neural network to encode the query point as real-valued embedding in AMVH. To this end, a well-designed alternative optimization algorithm is proposed to efficiently solve the highly coupled problem consists of neural network learning and mixed integer programming problem. Moreover, we present a simple method for incremental extension of database points. Theoretically, we demonstrate that the similarity metric of NAMVH can be regarded as the query sensitive weighted Hamming distance metric. Extensive experiments on seven datasets verify the effectiveness of training, query and storage in NAMVH.
  4. An unsupervised deep graph hashing method (UDGH) is proposed, where deep hash function learning and graph structure preservation are seamlessly formulated into a unified framework. Specifically, the neighborhood structure between network outputs and binary codes is locally preserved in an asymmetric way, yielding the asymmetric spectral loss function. Moreover, the gradients of minibatch samples can be readily computed with the global graph Laplacian, so that the whole structure information can be dramatically considered in every minibatch. By this way, the graph Laplacian can regulate the network training sufficiently and directly. Extensive experiments demonstrate that our approach surpasses the existing unsupervised hashing methods.
Subject Area模式识别 ; 计算机神经网络
MOST Discipline Catalogue工学::控制科学与工程
Pages132
Language中文
Document Type学位论文
Identifierhttp://ir.ia.ac.cn/handle/173211/23842
Collection模式识别国家重点实验室_先进数据分析与学习
毕业生_博士学位论文
Recommended Citation
GB/T 7714
达铖. 非对称哈希学习方法研究[D]. 中国科学院自动化研究所. 中国科学院大学,2019.
Files in This Item:
File Name/Size DocType Version Access License
Thesis-达铖-final.pdf(11553KB)学位论文 开放获取CC BY-NC-SAApplication Full Text
Related Services
Recommend this item
Bookmark
Usage statistics
Export to Endnote
Google Scholar
Similar articles in Google Scholar
[达铖]'s Articles
Baidu academic
Similar articles in Baidu academic
[达铖]'s Articles
Bing Scholar
Similar articles in Bing Scholar
[达铖]'s Articles
Terms of Use
No data!
Social Bookmark/Share
All comments (0)
No comment.
 

Items in the repository are protected by copyright, with all rights reserved, unless otherwise indicated.