In the recent years, deep learning has gained popularity for its ability to be utilized for several computer vision applications without any apriori knowledge. However, to introduce better inductive bias incorporating prior knowledge along with learnedinformation is critical. To that end, human intervention including choice of algorithm, data and model in deep learning pipelines can be considered a prior. Thus, it is extremely important to select effective priors for a given application. This dissertation explores different aspects of a deep learning pipeline and provides insights as to why a particular prior is effective for the corresponding application. For analyzing the effect of model priors, three applications which involvesequential modelling problems i.e. Audio Source Separation, Clinical Time-series (Electroencephalogram (EEG)/Electrocardiogram(ECG)) based Differential Diagnosis and Global Horizontal Irradiance Forecasting for Photovoltaic (PV) Applications are chosen. For data priors, the application of image classification is chosen and a new algorithm titled,“Invenio” that can effectively use data semantics for both task and distribution shift scenarios is proposed. Finally, the effectiveness of a data selection prior is shown using the application of object tracking wherein the aim is to maintain the tracking performance while prolonging the battery usage of image sensors by optimizing the data selected for reading from the environment. For every research contribution of this dissertation, several empirical studies are conducted on benchmark datasets. The proposed design choices demonstrate significant performance improvements in comparison to the existing application specific state-of-the-art deep learning strategies.

Severe forms of mental illness, such as schizophrenia and bipolar disorder, are debilitating conditions that negatively impact an individual's quality of life. Additionally, they are often difficult and expensive to diagnose and manage, placing a large burden on society. Mental illness is typically diagnosed by the use of clinical interviews and a set of neuropsychiatric batteries; a key component of nearly all of these evaluations is some spoken language task. Clinicians have long used speech and language production as a proxy for neurological health, but most of these assessments are subjective in nature. Meanwhile, technological advancements in speech and natural language processing have grown exponentially over the past decade, increasing the capacity of computer models to assess particular aspects of speech and language. For this reason, many have seen an opportunity to leverage signal processing and machine learning applications to objectively assess clinical speech samples in order to automatically compute objective measures of neurological health. This document summarizes several contributions to expand upon this body of research. Mainly, there is still a large gap between the theoretical power of computational language models and their actual use in clinical applications. One of the largest concerns is the limited and inconsistent reliability of speech and language features used in models for assessing specific aspects of mental health; numerous methods may exist to measure the same or similar constructs and lead researchers to different conclusions in different studies. To address this, a novel measurement model based on a theoretical framework of speech production is used to motivate feature selection, while also performing a smoothing operation on features across several domains of interest. Then, these composite features are used to perform a much wider range of analyses than is typical of previous studies, looking at everything from diagnosis to functional competency assessments. Lastly, potential improvements to address practical implementation challenges associated with the use of speech and language technology in a real-world environment are investigated. The goal of this work is to demonstrate the ability of speech and language technology to aid clinical practitioners toward improvements in quality of life outcomes for their patients.

Computer vision is becoming an essential component of embedded system applications such as smartphones, wearables, autonomous systems and internet-of-things (IoT). These applications are generally deployed into environments with limited energy, memory bandwidth and computational resources. This trend is driving the development of energy-effi cient image processing solutions from sensing to computation. In this thesis, diff erent alternatives are explored to implement energy-efficient computer vision systems. First, I present a fi eld programmable gate array (FPGA) implementation of an adaptive subsampling algorithm for region-of-interest (ROI) -based object tracking. By implementing the computationally intensive sections of this algorithm on an FPGA, I aim to offl oad computing resources from energy-ineffi cient graphics processing units (GPUs) and/or general-purpose central processing units (CPUs). I also present a working system executing this algorithm in near real-time latency implemented on a standalone embedded device. Secondly, I present a neural network-based pipeline to improve the performance of event-based cameras in non-ideal optical conditions. Event-based cameras or dynamic vision sensors (DVS) are bio-inspired sensors that measure logarithmic per-pixel brightness changes in a scene. Their advantages include high dynamic range, low latency and ultra-low power when compared to standard frame-based cameras. Several tasks have been proposed to take advantage of these novel sensors but they rely on perfectly calibrated optical lenses that are in-focus. In this work I propose a methodto reconstruct events captured with an out-of-focus event-camera so they can be fed into an intensity reconstruction task. The network is trained with a dataset generated by simulating defocus blur in sequences from object tracking datasets such as LaSOT and OTB100. I also test the generalization performance of this network in scenes captured with a DAVIS event-based sensor equipped with an out-of-focus lens.

Researchers have observed that the frequencies of leading digits in many man-made and naturally occurring datasets follow a logarithmic curve, with digits that start with the number 1 accounting for 30% of all numbers in the dataset and digits that start with the number 9 accounting for 5% of all numbers in the dataset. This phenomenon, known as Benford's Law, is highly repeatable and appears in lists of numbers from electricity bills, stock prices, tax returns, house prices, death rates, lengths of rivers, and naturally occurring images. This paper will demonstrate that human speech spectra also follow Benford's Law. This observation is used to motivate a new set of features that can be efficiently extracted from speech and demonstrate that these features can be used to classify between human speech and synthetic speech.

Due to their effectiveness in capturing similarities between different entities, graphical models are widely used to represent datasets that reside on irregular and complex manifolds. Graph signal processing offers support to handle such complex datasets. By extending the digital signal processing conceptual frame from time and frequency domain to graph domain, operators such as graph shift, graph filter and graph Fourier transform are defined. In this dissertation, two novel graph filter design methods are proposed. First, a graph filter with multiple shift matrices is applied to semi-supervised classification, which can handle features with uneven qualities through an embedded feature importance evaluation process. Three optimization solutions are provided: an alternating minimization method that is simple to implement, a convex relaxation method that provides a theoretical performance benchmark and a genetic algorithm, which is computationally efficient and better at configuring overfitting. Second, a graph filter with splitting-and-merging scheme is proposed, which splits the graph into multiple subgraphs. The corresponding subgraph filters are trained parallelly and in the last, by merging all the subgraph filters, the final graph filter is obtained. Due to the splitting process, the redundant edges in the original graph are dropped, which can save computational cost in semi-supervised classification. At the same time, this scheme also enables the filter to represent unevenly sampled data in manifold learning. To evaluate the performance of the proposed graph filter design approaches, simulation experiments with synthetic and real datasets are conduct. The Monte Carlo cross validation method is employed to demonstrate the need for the proposed graph filter design approaches in various application scenarios. Criterions, such as accuracy, Gini score, F1-score and learning curves, are provided to analyze the performance of the proposed methods and their competitors.

Operational efficiency of solar energy farms requires detailed analytics and information on each panel regarding voltage, current, temperature, and irradiance. Monitoring utility-scale solar arrays was shown to minimize the cost of maintenance and help optimize the performance of photovoltaic (PV) arrays under various conditions. This dissertation describes a project that focuses on the development of machine learning and neural network algorithms. It also describes an 18kW solar array testbed for the purpose of PV monitoring and control. The use of the 18kW Sensor Signal and Information Processing (SenSIP) PV testbed which consists of 104 modules fitted with smart monitoring devices (SMDs) is described in detail. Each of the SMDs has embedded, a wireless transceiver, and relays that enable continuous monitoring, fault detection, and real-time connection topology changes. Data is obtained in real time using the SenSIP PV testbed. Machine learning and neural network algorithms for PV fault classification is are studied in depth. More specifically, the development of a series of customized neural networks for detection and classification of solar array faults that include soiling, shading, degradation, short circuits and standard test conditions is considered. The evaluation of fault detection and classification methods using metrics such as accuracy, confusion matrices, and the Risk Priority Number (RPN) is performed. The examination and assessment the classification performance of customized neural networks with dropout regularizers is presented in detail. The development and evaluation of neural network pruning strategies and illustration of the trade-off between fault classification model accuracy and algorithm complexity is studied. This study includes data from the National Renewable Energy Laboratory (NREL) database and also real-time data collected from the SenSIP testbed at MTW under various loading and shading conditions. The overall approach for detection and classification promises to elevate the performance and robustness of PV arrays.

Dealing with relational data structures is central to a wide-range of applications including social networks, epidemic modeling, molecular chemistry, medicine, energy distribution, and transportation. Machine learning models that can exploit the inherent structural/relational bias in the graph structured data have gained prominence in recent times. A recurring idea that appears in all approaches is to encode the nodes in the graph (or the entire graph) as low-dimensional vectors also known as embeddings, prior to carrying out downstream task-specific learning. It is crucial to eliminate hand-crafted features and instead directly incorporate the structural inductive bias into the deep learning architectures. In this dissertation, deep learning models that directly operate on graph structured data are proposed for effective representation learning. A literature review on existing graph representation learning is provided in the beginning of the dissertation. The primary focus of dissertation is on building novel graph neural network architectures that are robust against adversarial attacks. The proposed graph neural network models are extended to multiplex graphs (heterogeneous graphs). Finally, a relational neural network model is proposed to operate on a human structural connectome. For every research contribution of this dissertation, several empirical studies are conducted on benchmark datasets. The proposed graph neural network models, approaches, and architectures demonstrate significant performance improvements in comparison to the existing state-of-the-art graph embedding strategies.