Novel Generative Semisupervised Learning Based on Fine-Grained Component-Conditional Class Labeling

Open Access
Lin, Chu-Fang
Graduate Program:
Electrical Engineering
Doctor of Philosophy
Document Type:
Date of Defense:
June 08, 2011
Committee Members:
  • David Jonathan Miller, Dissertation Advisor
  • David Jonathan Miller, Committee Chair
  • George Kesidis, Committee Member
  • William Kenneth Jenkins, Committee Member
  • Christopher Collins, Committee Member
  • Kenneth K Kuo, Committee Member
  • nearest neighbor classification
  • mixture models
  • inductive inference
  • generative models
  • semisupervised learning
  • Expectation- Maximization algorithm
This dissertation presents novel generative semisupervised mixtures with more fine-grained class label generation mechanisms than in previous semisupervised learning works. The proposed fine-grained models combine advantages of semisupervised mixtures, which achieve label extrapolation over a component, and nearest-neighbor(NN)/nearest-prototype classification, which achieve accurate classification in the vicinity of labeled samples. This dissertation includes several two-stage stochastic data generation mechanisms which involve first generating all the feature data and then, for the labeled data subset, generating each class label, not directly according to a component density, but rather based on unlabeled samples that were generated according to the component density. Variants of the proposed methods differ in whether or not class decisionmaking requires transduction. This generation mechanisms entail more sophisticated, exact E-step evaluations than for standard mixtures. These form the basis for generalized EM algorithms for learning. The proposed models are advantageous, compared to previous semisupervised mixtures, when within-component class proportions are not constant over the feature space region "owned by" a component. The practicality of this scenario is borne out by our experiments on UC Irvine data sets, which demonstrate consistent, substantial gains in classification accuracy over previous semisupervised mixtures and over both standard KNN and within-component KNN classification. Moreover, for small labeled fractions, the fine-grained methods also outperform supervised linear and nonlinear kernel support vector machines. For clustering and density estimation, the proposed methods outperform previous semisupervised approaches, but only give better results than standard unsupervised mixtures when there is both significant true component overlap and sufficient labeled examples to disambiguate components. We extend the original inductive fine-grained model by incorporating active learning methods. We propose two new uncertainty sampling methods in comparison with traditional entropy-based uncertainty sampling. Our experiments on UC Irvine data sets validate that the proposed active learning methods improve classification accuracy more than standard entropy-based active learning, especially when the labeled percentage is small. Moreover, we extend our inductive fine-grained model by introducing variable weighting to modulate the influence of samples and labels on EM parameter estimation. This new approach is shown to outperform previous weighting schemes. Another area where a significant amount of research is conducted to reduce supervision is image segmentation. We propose a three-dimensional (3D) tissue delineation framework to address the challenging of segmenting soft tissues in computed tomography (CT) images from diverse body types with little supervision. Conventional image processing methods are not suitable for automated delineation of soft tissue boundaries in CT image due to low contrast between soft tissues the variability in tissue location, shape, and size. To overcome these limitations, we have introduced a novel procedure, making use of a standard labeled template and an abdominal probabilistic atlas. Our algorithm seeks to find an optimal 3D tissue mapping to each new individual. The mappings are chosen to minimize a customized cost function that enforces tissue contiguity and relative tissue locations while accounting for tissue variability across individuals. Tissue boundaries in CT images that are missing due to low contrast are extrapolated using boundary information and anatomical knowledge from the mapped labeled template and the probabilistic atlas. Our approach leads to satisfactory tissue delineations for diverse body types. We view the tissue segmentation task as a classification problem and apply semisupervised learning on CT images. A mixture of expert classifier is applied on the CT soft tissue segment, which comprises multiple soft tissues, to further classify individual tissues. By extracting texture and distance features from the standard anatomical images in addition to gray-scale intensity from patient-specific CT images, we are able to classify seven soft tissues. A future work that combines our fine-grained models with the tissue delineation framework is suggested to obtain more accurate soft tissue segmentation.