Semi-Supervised Learning for Scalable and Robust Visual Search

de Jun Wang

Unlike textual document retrieval, searching of visual data is still far from satisfactory. There exist major gaps between the available solutions and practical needs in both accuracy and computational cost. This thesis aims at the development of robust and scalable solutions for visual search and retrieval. Specifically, we investigate two classes of approaches: graph-based semi-supervised learning and hashing techniques. The graph-based approaches are used to improve accuracy, while hashing approaches are used to improve efficiency and cope with large-scale applications. A common theme shared between these two subareas of our work is the focus on semi-supervised learning paradigm, in which a small set of labeled data is complemented with large unlabeled datasets. Graph-based approaches have emerged as methods of choice for general semi-supervised tasks when no parametric information is available about the data distribution. It treats both labeled and unlabeled samples as vertices in a graph and then instantiates pairwise edges between these vertices to capture affinity between the corresponding samples. A quadratic regularization framework has been widely used for label prediction over such graphs. However, most of the existing graph-based semi-supervised learning methods are sensitive to the graph construction process and the initial labels. We propose a new bivariate graph transduction formulation and an efficient solution via an alternating minimization procedure. Based on this bivariate framework, we also develop new methods to filter unreliable and noisy labels. Extensive experiments over diverse benchmark datasets demonstrate the superior performance of our proposed methods. However, graph-based approaches suffer from the critical bottleneck in scalability since graph construction requires a quadratic complexity and the inference procedure costs even more. The widely used graph construction method relies on nearest neighbor search, which is prohibitive for large-scale applications. In addition, most large-scale visual search problems involve handling high-dimensional visual descriptors, thereby causing another challenge in excessive storage requirement. To handle the scalability issue of both computation and storage, the second part of the thesis focuses on efficient techniques for conducting approximate nearest neighbor (ANN) search, which is key to many machine learning algorithms, including graph-based semi-supervised learning and clustering. Specifically, we propose Semi-Supervised Hashing (SSH) methods that leverage semantic similarity over a small set of labeled data while preventing overfitting. We derive a rigorous formulation in which a supervised term minimizes the empirical errors on the labeled data and an unsupervised term provides effective regularization by maximizing variance and independence of individual bits. Experiments on several large datasets demonstrate the clear performance gain over several state-of-the-art methods without significant increase of the computational cost. The main contributions of the thesis include the following. Bivariate graph transduction: a) a bivariate formulation for graph-based semi-supervised learning with an efficient solution by alternating optimization; b) theoretic analysis from the view of graph cut for the bivariate optimization procedure; c) novel applications of the proposed techniques, such as interactive image retrieval, automatic re-ranking for text based image search, and a brain computer interface (BCI) for image retrieval. Semi-supervised hashing: a) a rigorous semi-supervised paradigm for hash functions learning with a tradeoff between empirical fitness on pair-wise label consistence and an information-theoretic regularizer; b) several efficient solutions for deriving semi-supervised hash functions, including an orthogonal solution using eigen-decomposition, a revised strategy for learning non-orthogonal hash functions, a sequential learning algorithm to derive boosted hash functions, and an extension to unsupervised cases by using pseudo labels. Two parts of the thesis - bivariate graph transduction and semi-supervised hashing - are complimentary and can be combined to achieve significant performance improvement in both speed and accuracy. Hash methods can help build sparse graphs in a linear time fashion and greatly reduce the data size, but they lack sufficient accuracy. Graph-based methods provide unique capabilities to handle non-linear data structures with noisy labels but suffer from high computational complexity. The synergistic combination of the two offers great potential for advancing the state-of-the-art in large-scale visual search and many other applications.