Multi-view Deep Gaussian Process Model And Its Application In Pathological Diagnosis

Multi-view learning is a promising hot research topic in machine learning,which aims to integrate data information from different views to improve learning performance.In recent years,multi-view learning has made great progress.Although various multi-view learning algorithms have been proposed,it is still challenging to jointly explore the information contained in different views under the principle of consistency and complementarity.The recent deep Gaussian processes(DGPs)model has good uncertainty estimation,strong nonlinear mapping ability and good generalization performance,which is an attractive Bayesian probability model.DGPs model is composed of multi-layer potential variables and adopts the hierarchical structure of Gaussian process mapping.Recently,DGPs model has shown excellent performance in different machine learning tasks.However,DGPs only focuses on single view data and cannot be directly applied to multi-view scenarios,which limits its further application to broader learning tasks in reality.The combination of multi-view learning and deep Gaussian process has important significance and value in both theoretical research and practical application.In order to solve the above issues,this dissertation explores the multi-view deep Gaussian processes models from four machine learning tasks: representation learning,supervised learning,view-missing data and pathology image diagnosis,which expands the application range of multi-view deep Gaussian processes models from both theoretical and practical aspects.Specifically,the main work of this dissertation is as follows:(1)To learn comprehensive representation of multi-view data,we propose a multiview deep Gaussian processes model for representation learning(Mv DGPs).The proposed model inherits the advantages of deep Gaussian processes and multi-view learning,which can learn more efficient multi-view data representations.Mv DGPs consists of two stages.The first stage is multi-view data representation learning,which is mainly used to learn a comprehensive representation of multi-view data.The second stage is classifier design,which aims to select an appropriate classifier to better utilize the representation obtained in the first stage.Compared with DGPs,Mv DGPs can model asymmetrically according to the statistical properties of different views,thus better describing the differences between different views.Experimental results on multiple real-world datasets show that the proposed model is effective and outperforms other comparative multi-view representation learning approaches.(2)To handle labeled multi-view data,a multi-view deep Gaussian processes model for supervised learning is proposed(Sup Mv DGP).The model uses the labels of views to further improve performance,and utilizes quantitative uncertainty information as a supplement to help practitioners make better use of the prediction results.According to the diversity of views,the proposed model can build asymmetric depth structure to better model different views,so as to make full use of the characteristics of each view.We provide an effective variational inference method for solving the proposed model.Finally,we conducted comprehensive comparative experiments on multiple real-world datasets to evaluate the performance of the proposed model.The experimental results show that Sup Mv DGP has achieved better results in multiple tasks,which proves its effectiveness and superiority.Meanwhile,we perform a case study showing that SupMv DGP has the ability to provide uncertainty estimates over other deep models,which can help people treat the prediction results better in high-risk applications.(3)To address view-missing scenarios in multi-view learning,a partial multi-view deep Gaussian cross generation model is proposed(PMv CG).The model aims to jointly model views according to the principles of consistency and complementarity,and eventually learn the comprehensive representation of partial multi-view data.PMv CG can discover cross-view associations by learning view-shared and view-specific features of different views in the representation space.The missing views are reconstructed by the latent variables of the existing views,and then applied in turn to further train the model.At the same time,we integrate uncertainty into the representation to improve performance.We design a variational inference and iterative optimization algorithm to effectively solve PMv CG.Finally,we conducted experiments on multiple datasets to verify the performance of PMv CG.The experimental results indicate that PMv CG achieves satisfactory results and outperforms other comparative methods under different experimental settings.(4)To facilitate the research on computer-aided diagnosis methods,a multi-view deep learning model for pathology image diagnosis is proposed(Mv PID),which combines image features with multi-view deep networks.Specifically,first,the entire slice image is segmented into different non-overlapping sub-slices.Then,we extract different image features from sub-slices as different views for multi-view learning.Subsequently,we propose to use a view-specific deep Gaussian process to extract the unique information of different views,and use a view-common autoencoder network to integrate the information of different views into a common representation.The learned representation is input into the downstream classifier to realize automatic pathology diagnosis.Experimental results on real pathology data show that the proposed approach is reasonable.The best classification performance far exceeds the diagnostic accuracy of pathologists,which demonstrates the application potential of the proposed Mv PID.In this dissertation,we combine deep Gaussian processes and multi-view learning for the first time,and propose several multi-view deep Gaussian processes models for different machine learning tasks.The proposed models enrich the multi-view learning methods and expand the application scope of the deep Gaussian processes.The effectiveness and efficiency of the proposed models are verified in different experimental settings.