Image segmentation is widely used in a variety of computer vision tasks, such as object localization and recognition, boundary detection, and medical imaging. This thesis proposes deep learning architectures to improve automatic object localization and boundary delineation for salient object segmentation in natural images and for 2D medical image segmentation. First, we propose and evaluate a novel dilated dense encoder-decoder architecture with a custom dilated spatial pyramid pooling block to accurately localize and delineate boundaries for salient object segmentation. The dilation offers better spatial understanding and the dense connectivity preserves features learned at shallower levels of the network for better localization. Tested on three publicly available datasets, our architecture outperforms the state-of-the-art for one and is very competitive on the other two. Second, we propose and evaluate a custom 2D dilated dense UNet architecture for accurate lesion localization and segmentation in medical images. This architecture can be utilized as a stand-alone segmentation framework or used as a rich feature extracting backbone to aid other models in medical image segmentation. Our architecture outperforms all baseline models for accurate lesion localization and segmentation on a new dataset. We furthermore explore the main considerations that should be taken into account for 3D medical image segmentation, among them preprocessing techniques and specialized loss functions.

Reducing traffic accidents is an important public safety challenge, therefore, accident analysis and prediction has been a topic of much research over the past few decades. Using small-scale datasets with limited coverage, being dependent on extensive set of data, and being not applicable for real-time purposes are the important shortcomings of the existing studies. To address these challenges, we propose a new solution for real-time traffic accident prediction using easy-to-obtain, but sparse data. Our solution relies on a deep-neural-network model (which we have named DAP, for Deep Accident Prediction); which utilizes a variety of data attributes such as traffic events, weather data, points-of-interest, and time. DAP incorporates multiple components including a recurrent (for time-sensitive data), a fully connected (for time-insensitive data), and a trainable embedding component (to capture spatial heterogeneity). To fill the data gap, we have - through a comprehensive process of data collection, integration, and augmentation - created a large-scale publicly available database of accident information named US-Accidents. By employing the US-Accidents dataset and through an extensive set of experiments across several large cities, we have evaluated our proposal against several baselines. Our analysis and results show significant improvements to predict rare accident events. Further, we have shown the impact of traffic information, time, and points-of-interest data for real-time accident prediction.

Multimodal datasets contain an enormous amount of relational information, which grows exponentially with the introduction of new modalities. Learning representations in such a scenario is inherently complex due to the presence of multiple heterogeneous information channels. These channels can encode both (a) inter-relations between the items of different modalities and (b) intra-relations between the items of the same modality. Encoding multimedia items into a continuous low-dimensional semantic space such that both types of relations are captured and preserved is extremely challenging, especially if the goal is a unified end-to-end learning framework. The two key challenges that need to be addressed are: 1) the framework must be able to merge complex intra and inter relations without losing any valuable information and 2) the learning model should be invariant to the addition of new and potentially very different modalities. In this paper, we propose a flexible framework which can scale to data streams from many modalities. To that end we introduce a hypergraph-based model for data representation and deploy Graph Convolutional Networks to fuse relational information within and across modalities. Our approach provides an efficient solution for distributing otherwise extremely computationally expensive or even unfeasible training processes across multiple-GPUs, without any sacrifices in accuracy. Moreover, adding new modalities to our model requires only an additional GPU unit keeping the computational time unchanged, which brings representation learning to truly multimodal datasets. We demonstrate the feasibility of our approach in the experiments on multimedia datasets featuring second, third and fourth order relations.

Are two sets of observations drawn from the same distribution? This problem is a two-sample test. Kernel methods lead to many appealing properties. Indeed state-of-the-art approaches use the $L^2$ distance between kernel-based distribution representatives to derive their test statistics. Here, we show that $L^p$ distances (with $p\geq 1$) between these distribution representatives give metrics on the space of distributions that are well-behaved to detect differences between distributions as they metrize the weak convergence. Moreover, for analytic kernels, we show that the $L^1$ geometry gives improved testing power for scalable computational procedures. Specifically, we derive a finite dimensional approximation of the metric given as the $\ell_1$ norm of a vector which captures differences of expectations of analytic functions evaluated at spatial locations or frequencies (i.e, features). The features can be chosen to maximize the differences of the distributions and give interpretable indications of how they differs. Using an $\ell_1$ norm gives better detection because differences between representatives are dense as we use analytic kernels (non-zero almost everywhere). The tests are consistent, while much faster than state-of-the-art quadratic-time kernel-based tests. Experiments on artificial and real-world problems demonstrate improved power/time tradeoff than the state of the art, based on $\ell_2$ norms, and in some cases, better outright power than even the most expensive quadratic-time tests.

Information in electronic health records (EHR), such as clinical narratives, examination reports, lab measurements, demographics, and other patient encounter entries, can be transformed into appropriate data representations that can be used for downstream clinical machine learning tasks using representation learning. Learning better representations is critical to improve the performance of downstream tasks. Due to the advances in machine learning, we now can learn better and meaningful representations from EHR through disentangling the underlying factors inside data and distilling large amounts of information and knowledge from heterogeneous EHR sources. In this chapter, we first introduce the background of learning representations and reasons why we need good EHR representations in machine learning for medicine and healthcare in Section 1. Next, we explain the commonly-used machine learning and evaluation methods for representation learning using a deep learning approach in Section 2. Following that, we review recent related studies of learning patient state representation from EHR for clinical machine learning tasks in Section 3. Finally, in Section 4 we discuss more techniques, studies, and challenges for learning natural language representations when free texts, such as clinical notes, examination reports, or biomedical literature are used. W e also discuss challenges and opportunities in these rapidly growing research fields.

In this chapter, we provide a brief overview of applying machine learning techniques for clinical prediction tasks. We begin with a quick introduction to the concepts of machine learning and outline some of the most common machine learning algorithms. Next, we demonstrate how to apply the algorithms with appropriate toolkits to conduct machine learning experiments for clinical prediction tasks. The objectives of this chapter are to (1) understand the basics of machine learning techniques and the reasons behind why they are useful for solving clinical prediction problems, (2) understand the intuition behind some machine learning models, including regression, decision trees, and support vector machines, and (3) understand how to apply these models to clinical prediction problems using publicly available datasets via case studies.

Dimensionality reduction (DR) on the manifold includes effective methods which project the data from an implicit relational space onto a vectorial space. Regardless of the achievements in this area, these algorithms suffer from the lack of interpretation of the projection dimensions. Therefore, it is often difficult to explain the physical meaning behind the embedding dimensions. In this research, we propose the interpretable kernel DR algorithm (I-KDR) as a new algorithm which maps the data from the feature space to a lower dimensional space where the classes are more condensed with less overlapping. Besides, the algorithm creates the dimensions upon local contributions of the data samples, which makes it easier to interpret them by class labels. Additionally, we efficiently fuse the DR with feature selection task to select the most relevant features of the original space to the discriminative objective. Based on the empirical evidence, I-KDR provides better interpretations for embedding dimensions as well as higher discriminative performance in the embedded space compared to the state-of-the-art and popular DR algorithms.

--Multiview analysis aims at extracting shared latent components from data samples that are acquired in different domains, e.g., image, text, and audio. Classic multiview analysis, e.g., canonical correlation analysis (CCA), tackles this problem via matching the linearly transformed views in a certain latent domain. More recently, powerful nonlinear learning tools such as kernel methods and neural networks are utilized for enhancing the classic CCA. However, unlike linear CCA whose theoretical aspects are clearly understood, nonlinear CCA approaches are largely intuition-driven. In particular, it is unclear under what conditions the shared latent components across the veiws can be identified--while identifiability plays an essential role in many applications. In this work, we revisit nonlinear multiview analysis and address both the theoretical and computational aspects. We take a nonlinear multiview mixture learning viewpoint, which is a natural extension of the classic generative models for linear CCA. From there, we derive a nonlinear multiview analysis criteron. We show that minimizing this criterion leads to identification of the latent shared components up to certain ambiguities, under reasonable conditions. On the computation side, we propose an effective algorithm with simple and scalable update rules. A series of simulations and real-data experiments corroborate our theoretical analysis. Multiview analysis has been an indispensable tool in statistical signal processing, machine learning, and data analytics. In the context of multiview learning, a view can be understood as measurements of data entities (e.g., a cat) in a certain domain (e.g., text, image, and audio). Most data entities naturally appear in different domains. Multiview analysis aims at extracting essential and common information from different views. Compared with single-view analysis tools like principal component analysis (PCA), independent component analysis (ICA) [1], and nonnegative matrix factorization (NMF) [2], multiview analysis tools such as canonical correlation analysis (CCA) [3] have an array of unique features. For example, CCA has been shown to be more robust to noise and view-specific strong interference [4], [5]. The classic CCA has been extensively studied in the literature, ever since its proposal in statistics in the 1930s [3], [6]. Q. Lyu and X. Fu are with the School of Electrical Engineering and Computer Science, Oregon State University, Corvallis, OR 97331, United States. The transformations are supposed to'project' the views to a domain where the views share similar representations. Interestingly, the formulated optimization problem, although being nonconvex, can be recast into a generalized eigende-composition problem and solved efficiently [3], [7]. In recent years, many attempts have been made towards scaling up classic CCA to handle big datasets; see, e.g., [7].

Data augmentation has been widely applied as an effective methodology to prevent over-fitting in particular when training very deep neural networks. The essential benefit comes from introducing additional priors in visual invariance, and thus generate images in different appearances but containing the same semantics. Recently, researchers proposed a few powerful data augmentation techniques which indeed improved accuracy, yet we notice that these methods have also caused a considerable gap between clean and augmented data. This paper revisits this problem from an analytical perspective, for which we estimate the upper-bound of testing loss using two terms, named empirical risk and generalization error, respectively. Data augmentation significantly reduces the generalization error, but meanwhile leads to a larger empirical risk, which can be alleviated by a simple algorithm, i.e., using less-augmented data to refine the model trained on fully-augmented data . We validate our approach on a few popular image classification datasets including CIFAR and ImageNet, and demonstrate consistent accuracy gain. We also conjecture that this simple strategy implies a generalized approach to circumvent local minima, which is of value to future research on model optimization.

A continuing challenge for machine learning is providing methods to perform computation on data while ensuring the data remains private. In this paper we build on the provable privacy guarantees of differential privacy which has been combined with Gaussian processes through the previously published \emph{cloaking method}. In this paper we solve several shortcomings of this method, starting with the problem of predictions in regions with low data density. We experiment with the use of inducing points to provide a sparse approximation and show that these can provide robust differential privacy in outlier areas and at higher dimensions. We then look at classification, and modify the Laplace approximation approach to provide differentially private predictions. We then combine this with the sparse approximation and demonstrate the capability to perform classification in high dimensions. We finally explore the issue of hyperparameter selection and develop a method for their private selection. This paper and associated libraries provide a robust toolkit for combining differential privacy and GPs in a practical manner.