Search results

Title: FUNCTION APPROXIMATION WITH KERNEL METHODS
Creator: Zhou, Xuan
Date: 2015, 2015-12
Description: This dissertation studies the problem of approximating functions of d variables in a separable Banach space Fd. In particular we are...
Show moreThis dissertation studies the problem of approximating functions of d variables in a separable Banach space Fd. In particular we are interested in convergence and tractability results in the worst case setting and in the average case setting. The symmetric positive definite kernel in both settings is of a product form Kd(x, t) := d =1 1 − α2 + α2 Kγ (x , t ) for all x, t ∈ Rd. The kernel Kd generalizes the anisotropic Gaussian kernel, whose tractability properties have been established in the literature. For a fixed d, we study rates of convergence, which indicate how quickly approximation errors decay. Since rates of convergence can deteriorate quickly as d increases, it is desirable to have dimension-independent convergence rates, which corresponds to the concept of strong polynomial tractability. We present sufficient conditions on {α }∞ =1 and {γ }∞ =1 under which strong polynomial tractability holds for function approximation problems in Fd. Numerical examples are presented to support the theory and guaranteed automatic algorithms are provided to solve the function approximation problem in a straightforward and efficient way. viii
Ph.D. in Applied Mathematics, December 2015
Show less

Title: DEVELOPMENT OF COMPUTER-AIDED DIAGNOSIS METHODS IN MAMMOGRAPHY
Creator: Wang, Juan
Date: 2015, 2015-12
Description: Computer-aided diagnosis (CAD) is developed as a diagnostic aid to provide a “second opinion” in diagnosis of breast cancer in early stage....
Show moreComputer-aided diagnosis (CAD) is developed as a diagnostic aid to provide a “second opinion” in diagnosis of breast cancer in early stage. Clustered microcalcifications (MCs) can be an important early sign of breast cancer. The goal of this work is to develop automatic CAD methods in mammography for breast cancer. Its contribution consists of both development of machine learning algorithms and study of related issues in detection and diagnosis of breast cancer with clustered MCs. First, a bi-thresholding scheme is proposed for reduction of false-positives (FPs) associated with linear structures in MC detection. An unified classifier with dummy variable modeling is further developed to reduce the FPs caused by both linear structures and MC-like noise patterns. It is demonstrated that both of the proposed algorithms can reduce FPs in MC detection, and thus, improve the detection accuracy significantly. Second, a spatial density modeling approach is investigated to quantify the spatial distribution of the MCs in a cluster when the MC detection is inaccurate. A spatial density function (SDF) is defined such that the extracted features are more robust to the presence of FPs and false-negatives (FNs) in MC detection. The results show that the features extracted from the SDF can achieve better class separation while being robust to the variations in MC detection when compared with those extracted from a traditional region-based method. Third, a retrieval-boosted approach is studied to discriminate between the benign and malignant MC lesions. A retrieval strategy is proposed to boost the classification performance by taking into account the similarity both in image features and in pathology. An adaptive Adaboost classifier, which can be adapted to the retrieved cases at a low computational cost, is applied to demonstrate the benefit of the retrieval strategy. The results show that the retrieval-boosted approach can signifishow that the features extracted from the SDF can achieve better class separation while being robust to the variations in MC detection when compared with those extracted from a traditional region-based method. Third, a retrieval-boosted approach is studied to discriminate between the benign and malignant MC lesions. A retrieval strategy is proposed to boost the classification performance by taking into account the similarity both in image features and in pathology. An adaptive Adaboost classifier, which can be adapted to the retrieved cases at a low computational cost, is applied to demonstrate the benefit of the retrieval strategy. The results show that the retrieval-boosted approach can significantly outperform its baseline classifier and that inclusion of pathology information in the retrieval can further improve the classification accuracy. Fourth, the perceptual similarity of MC lesions by radiologists is studied. The issues investigated include the degree of variability in the similarity ratings, the impact of this variability on agreement between readers in retrieval of similar lesions, and the factors contributing to the readers’ similarity ratings. The results indicate that perceptually similar lesions could be of diagnostic value in diagnosis for clustered MCs. Fifth, the feasibility of modeling the perceptual similarity of MC lesions is investigated. A support vector regression (SVR) is applied to model the perceptual similarity of clustered MCs, and a feature saliency analysis derived from SVR is used to determine the most relevant image features among a large set of candidate features. The results demonstrate that the relevant features are consistent in radiologists’ similarity ratings among different MC lesions, indicating that the perceptual similarity of MC lesions by radiologists can be effectively modeled. Finally, whether retrieval of similar images can effectively assist radiologists in diagnosis of clustered MCs is investigated. A retrieval system for relevant images is designed by considering both perceptually similar image features and the likelihood of malignancy of the lesion under consideration. An observer study is conducted to evaluate the diagnostic value of the proposed retrieval system. The results indicate that the proposed retrieval system has the potential to improve the reader’s ability in diagnosis of breast cancer with clustered MCs.
Ph.D. in Electrical Engineering, December 2015
Show less

Title: TEMPORAL AND SPATIOTEMPORAL MODELS FOR SHORT-CRIME PREDICTION
Creator: Liu, Xiaomu
Date: 2017, 2017-07
Description: One of the most important aspects of predictive policing is identifying the likely time and place of crime occurrences so as to prevent future...
Show moreOne of the most important aspects of predictive policing is identifying the likely time and place of crime occurrences so as to prevent future crimes. The ability to make short-term predictions may be of particular importance for optimizing police resource allocation. The goal of this study is to investigate the temporal and spatiotemporal pattern of crime in the city of Chicago and to build corresponding predictive models. First, a temporal model for forecasting citywide violent crime time count is proposed. This model is composed of a long-term trend and short-term variations using data of time, weather and crime. The importance of model reproducibility is addressed in this study to produce low-complexity models. We introduce an approach that provides a way to extend the model selection criterion to both prediction accuracy and model reproducibility. The experimental results show that models produced by this approach outperform several simple time-series models. It is also found that these models typically include fewer variables; therefore, they are more interpretable, and may provide superior generalization error. Next we develop a framework that provides predictions for tomorrow’s violent crime counts at the level of a police district. The procedures include citywide daily violent crime count prediction, violent crime density estimation, and distributing citywide predictions to districts according to the estimated densities. In order to estimate the crime spatial densities, we use mesh modeling and demonstrate that a mesh model can be used as the structure for modeling the spatial variation of crime rate since it is well adapted to the inhomogeneous crime distribution. The experimental results show that our method provides more-accurate forecasts than those given by historical crime statistics. One aspect of studying spatial pattern of crimes is identifying geographical regions with similar crime characteristics. Specifically, we illustrate applying unsupervised clustering techniques to segment the city into sub-regions. We explore the use of Gaussian mixture models combined with a Markov random field for the purpose of regularization. We also propose a framework for the evaluation of clustering models without knowing the ground truth, which can present a more-complete picture for model selection in unsupervised clustering problems. Finally, we develop a spatiotemporal prediction method that predicts the locations where violent crimes or property crimes are most likely to occur tomorrow. Crime incidents are rasterized by a spatiotemporal grid. Other factors that affect the time and location preferences of criminal activities are also leveraged and represented by that grid. Each spatiotemporal grid cell is treated as an example for training and testing our models. We also explore whether pooling data from various sub-regions based on spatial clustering can improve model performance. The experimental results show that our models are more accurate than conventional hot-spot models. It is found that the effects of using different training samples are not consistent, which depends on target crime type.
Ph.D. in Electrical Engineering, July 2017
Show less

Title: INDIVIDUAL-BASED RISK MODELS FOR CRIME PREVENTION AND MEDICAL PROGNOSIS
Creator: Haro Alonso, David
Date: 2018, 2018-05
Description: Parallel trends are currently taking place in the fields of crime and medicine, in which the focus is shifting from a reactive stance to a...
Show moreParallel trends are currently taking place in the fields of crime and medicine, in which the focus is shifting from a reactive stance to a proactive one. Both fields have traditionally been reactive, with police responding to 911 calls after a crime has occurred, and patients seeking medical care after symptoms have already appeared. In the field of crime, social-services programs, law-enforcement agencies, sociologists, and criminologists are studying ways to prevent crime, instead of merely reacting to it. A similar trend, known as preventive medicine, is concerned with addressing the causes of disease and not just focusing on treatment of disease that has already emerged. If crime and disease are to be prevented, it is important to understand the early warning signs of risk, to anticipate and treat problems before they occur. This can be accomplished via mathematical risk models that can evaluate an individual’s risk based on leading indicators. In this thesis I develop such models for two real-world problems in crime prevention and one in preventive medicine. A major focus of this thesis is to emphasize the accuracy of the ranking of risk for situations in which the allocation of resources must be prioritized to the highest-risk individuals. This is especially true in a social-services program designed to reduce crime, where the number of available social workers may be limited. In the first part of the thesis, I describe a novel method of risk modeling based on the probabilistic framework of a conditional random field, in which a machine-learning regressor is embedded. This is applicable in situations where an individual’s risk of an adverse outcome is partly dependent on the risk levels of others. We have applied this technique to develop a model that assesses an individual’s near-term risk of becoming a victim or arrestee in a shooting or homicide in Chicago. The model was developed as an informational tool for a pilot crime-prevention program that aims to offer social services to at-risk persons with the aim of providing opportunities for life changes that may reduce their crime risk. In the second part of the thesis, I describe a new model with a similar goal—to identify individuals at risk of involvement in crime—but aims to provide information for use in smaller cities that have a more typical array of crime concerns than Chicago. We developed the model as part of a current partnership with the Elgin Police Department, where a social-services intervention program under development will incorporate our model in identifying persons who might benefit from assistance. In the last part of the thesis, I describe a risk assessment algorithm for the medical field, which we developed in partnership with Cedars-Sinai Medical Center, Los Angeles, CA. In this work, we sought to demonstrate to the cardiology field (and the broader medical field) that machine learning can provide a better framework for risk stratification in medicine than traditional statistical methods such as logistic regression, which are the norm in that field. We also showed that, contrary to concerns by medical practitioners, machine learning can provide a solution that is easy to interpret.
Ph.D. in Electrical Engineering, May 2018
Show less

Title: MACHINE LEARNING METHODS FOR PREDICTING GLOBAL AND LOCAL CRIME IN AN URBAN AREA
Creator: Navarro Comes, Eric
Date: 2018, 2018-05
Description: In recent years there has been growing interest in development of computer methods that can model and predict crime events for crime...
Show moreIn recent years there has been growing interest in development of computer methods that can model and predict crime events for crime prevention in law enforcement agencies. A popular example is the creation of crime density maps which are used to provide early-warning information about potential hotspots of crime in an urban area. One important aspect of predictive policing is to identify the time and place of likely crime occurrence so as to prevent future crime events. The ability to make accurate, timely predictions can be particularly important for optimizing police resource allocation in an urban area. This thesis focuses on developing machine learning techniques in crime analysis and prevention for predicting the overall crime trend in an urban area, as well as the likelihood of crime occurrence in a given local area during a time period. By using crime data extracted from the Citizen and Law Enforcement Analysis and Reporting (CLEAR) system in the city of Chicago, we demonstrate that state-of-the-art learning algorithms can achieve improved prediction accuracy over traditional methods based on time series models. We then study prediction techniques for determining the likelihood of crime occurrence at a specific local area during a given time window. We demonstrate these techniques in the operational framework of the Strategic Decision Support Centers (SDSCs) in the Chicago Police Department, where only a small number (up to six) surveillance cameras can be monitored simultaneously at any given time in a single district. We apply prediction techniques to select those cameras that most likely have crime events happening within their viewsheds during a determined time window, thereby maximizing the crime monitoring efficiency. Using these models, we can achieve higher accuracy than the methods based on local crime density alone.
M.S. in Electrical Engineering, May 2018
Show less

Title: MODELING THE INFORMATION CONTENT OF THE LIMIT ORDER BOOK BY BAGGING
Creator: Li, Wenyi
Date: 2018
Description: I propose a bagging tree framework to study the information content of the limit order book in U.S. equity market. By measuring the...
Show moreI propose a bagging tree framework to study the information content of the limit order book in U.S. equity market. By measuring the predictability and profitability of the order book data up to 5 levels, I find that the limit orders book is informative. In addition to market orders, limit orders behind the best bid and ask prices also contributes to short-term future price movements. Finally, I design simple strategies to show that this information content can be effectively and consistently translated to economic value. My results may provide important implications for both researchers and market practitioners.
Show less

Title: Removing Confounds in Text Classification for Computational Social Science
Creator: Landeiro Dos Reis, Virgile
Date: 2018
Description: Nowadays, one can use social media and other online platforms to communicate with friends and family, write a review for a product, ask...
Show moreNowadays, one can use social media and other online platforms to communicate with friends and family, write a review for a product, ask questions about a topic of interest, or even share details of private life with the rest of the world. The ever-increasing amount of user-generated content has provided researchers with data that can offer insights on human behavior. Because of that, the field of computational social science - at the intersection of machine learning and social sciences - has soared in the past years, especially within the field of public health research. However, working with large amounts of user-generated data creates new issues. In this thesis, we propose solutions for two problems encountered in computational social science and related to confounding bias.First, because of the anonymity provided by online forums, social networks, or other blogging platforms through the common usage of usernames, it is hard to get accurate information about users such as gender, age, or ethnicity. Therefore, although collecting data on a specific topic is made easier, conducting an observational study with this type of data is not simple. Indeed, when one wishes to run a study to measure the effect of a variable on another variable, one needs to control for potential confounding variables. In the case of user-generated data, these potential confounding variables are at best noisily observed or inferred and at worst not observed at all. In this work, we wish to provide a way to use these inferred latent attributes in order to conduct an observational study while reducing the effect of confounding bias as much as possible. We first present a simple matching method in a large-scale observational study. Then, we propose a method to retrieve relevant and representative documents through adaptive query building in order to build the treatment and control groups of an observational study.Second, we focus on the problem of controlling for confounding variables when the influence of these variables on the target variable of a classification problem changes over time. Although identifying and controlling for confounding variables has been assiduously studied in empirical social science, it is often neglected in text classification. This can be understood by the fact that, if we assume that the impact of confounding variables does not change between the training and the testing data, then prediction accuracy should only be slightly affected. Yet, this assumption often does not hold when working with user-generated text. Because of this, computational science studies are at risk of reaching false conclusions when based on text classifiers that are not controlling for confounding variables. In this document, we propose to build a classifier that is robust to confounding bias shift, and we show that we can build such a classifier in different situations: when there are one or more observed confounding variables, when there is one noisily predicted confounding variable, or when the confounding variable is unknown but can be detected through topic modeling.
Show less

Title: Language, Perception, and Causal Inference in Online Communication
Creator: Wang, Zhao
Date: 2021
Description: With the proliferation of social media platforms, online communication is becoming increasingly popular. The nature of a wide audience and...
Show moreWith the proliferation of social media platforms, online communication is becoming increasingly popular. The nature of a wide audience and rapid spread of information make these platforms attractive to public entities, organizations, and individuals. Marketers use these platforms to advertise their products and collect customer feedbacks (e.g. Amazon, Airbnb, Yelp, IMDB). Politicians use these platforms to directly speak with the public and canvass for votes (e.g., Twitter, Youtube, Snapchat). Individuals use these platforms to connect with friends and share daily life (e.g., Twitter, Facebook, Instagram, Weibo). The various platforms allow users to build public image and increase reputation through a fast and cheap way. However, due to the lack of regulations and low effort of online communication, some users try to manage their public impression using vague and tricky expressions during communication, making it hard for the audience to identify the authenticity of the public messages. Studies across many disciplines have shown that words and language play an important role in effective communication but the nature and extent of this role remain murky. Prior works have investigated wording effect on audience perception, but we still need automatic methods to estimate the causal effect of lexical choice on human perception in large scale. Getting insights into the treatment effect of subtle linguistic signals is crucial for intelligent language understanding and text analysis.The causal estimation of wording effect on perception also provides us an alternative way to understand the causal relationship between word features and perception labels. Comparing with correlational associations between features and labels, which is typically learned by statistical machine learning models, we find inconsistencies between the causal and correlational associations. These inconsistencies suggest possible spurious correlations in text classification and it's significant to address this issue by applying causal inference knowledge to guide statistical classifiers.In this thesis, our first goal is to investigate wording effect in online communication and study causal inference in text. We start from a deceptive marketing task to quantify entities' word commitment from online public messaging and identify potentially inauthentic entities. We then propose several frameworks to estimate the causal effects of word choice on audience perception by adapting Individual Treatment Effect estimation from causal inference literature to our problem of Lexical Substitution Effect estimation. The findings from these projects motivate us to explore our second goal of applying causal inference knowledge to improve statistical model robustness. Specifically, we study the causal and correlational associations in text and discover possible spurious correlations in text classifiers. Then, by extending the causal discovery, we propose two frameworks to improve text classifier robustness and fairness either by directly removing bias correlations or by training a robust model with automatically generated counterfactual samples.
Show less

Title: Advances in Machine Learning: Theory and Applications in Time Series Prediction
Creator: London, Justin J.
Date: 2021
Description: A new time series modeling framework for forecasting, prediction and regime switching for recurrent neural networks (RNNs) using machine...
Show moreA new time series modeling framework for forecasting, prediction and regime switching for recurrent neural networks (RNNs) using machine learning is introduced. In this framework, we replace the perceptron with an econometric modeling unit. This cell/unit is a functionally dedicated to processing the prediction component from the econometric model. These supervised learning methods overcome the parameter estimation and convergence problems of traditional econometric autoregression (AR) models that use MLE and expectation-maximization (EM) methods which are computationally expensive, assume linearity, Gaussian distributed errors, and suffer from the curse of dimensionality. Consequently, due to these estimation problems and lower number of lags that can be estimated, AR models are limited in their ability to capture long memory or dependencies. On the other hand, plain RNNs suffer from the vanishing and gradient problem that also limits their ability to have long-memory. We introduce a new class of RNN models, the $\alpha$-RNN and dynamic $\alpha_{t}$-RNNs that does not suffer from these problems by utilizing an exponential smoothing parameter. We also introduce MS-RNNs, MS-LSTMs, and MS-GRUs., novel models that overcome the limitations of MS-ARs but enable regime (Markov) switching and detection of structural breaks in the data. These models have long memory, can handle non-linear dynamics, do not require data stationarity or assume error distributions. Thus, they make no assumptions about the data generating process and have the ability to better capture temporal dependencies leading to better forecasting and prediction accuracy over traditional econometric models and plain RNNs. Yet, the partial autocorrelation function and econometric tools, such as the the ADF, Ljung-Box, and AIC test statistics, can be used to determine optimal sequence lag lengths to input into these RNN models and to diagnose serial correlation. The new framework has capacity to characterize the non-linear partial autocorrelation of time series and directly capture dynamic effects such as trends and seasonality. The optimal sequence lag order can greatly influence prediction performance on test data. This structure provides more interpretability to ML models since traditional econometric models are embedded into RNNs. The ability to embed econometric models into RNNs will allow firms to improve prediction accuracy compared to traditional econometric or traditional ML models by creating a hybrid utilizing a well understood traditional econometric model and a ML. In theory the traditional econometric model should focus on the portion of the estimation error that is best managed by a traditional model and the ML should focus the non-linear portion of the model. This combined structure is a step towards explainable AI and lays the framework for econometric AI.
Show less

Title: Unsupervised Learning of Visual Odometry Using Direct Motion Modeling
Creator: Andrei, Silviu Stefan
Date: 2020
Description: Data for supervised learning of ego-motion and depth from video is scarce and expensive to produce. Subsequently, recent work has focused on...
Show moreData for supervised learning of ego-motion and depth from video is scarce and expensive to produce. Subsequently, recent work has focused on unsupervised learning methods and achieved remarkable results which surpass in some instances the accuracy of supervised methods. Many unsupervised approaches rely on predicted monocular depth and so ignore motion information. Moreover, unsupervised methods which do incorporate motion information do so only indirectly by designing the depth prediction network as an RNN. Hence, none of the existing methods model motion directly. In this work, we show that it is possible to achieve superior pose estimation results by modeling motion explicitly. Our method uses a novel learning-based formulation for depth propagation and refinement which transforms predicted depth maps from the current frame onto the next frame where it serves as a prior for predicting the next frame's depth map. Experimental results demonstrate that the proposed approach surpasses state of the art techniques for the pose prediction task while being better or on par with other methods for the depth prediction task.
Show less

Title: Machine Learning On Graphs
Creator: He, Jia
Date: 2022
Description: Deep learning has revolutionized many machine learning tasks in recent years.Successful applications range from computer vision, natural...
Show moreDeep learning has revolutionized many machine learning tasks in recent years.Successful applications range from computer vision, natural language processing to speech recognition, etc. The success is partially due to the availability of large amounts of data and fast growing computing resources (i.e., GPU and TPU), and partially due to the recent advances in deep learning technology. Neural networks, in particular, have been successfully used to process regular data such as images and videos. However, for many applications with graph-structured data, due to the irregular structure of graphs, many powerful operations in deep learning can not be readily applied. In recent years, there is a growing interest in extending deep learning to graphs. We first propose graph convolutional networks (GCNs) for the task of classification or regression on time-varying graph signals, where the signal at each vertex is given as a time series. An important element of the GCN design is filter design. We consider filtering signals in either the vertex (spatial) domain, or the frequency (spectral) domain. Two basic architectures are proposed. In the spatial GCN architecture, the GCN uses a graph shift operator as the basic building block to incorporate the underlying graph structure into the convolution layer. The spatial filter directly utilizes the graph connectivity information. It defines the filter to be a polynomial in the graph shift operator to obtain the convolved features that aggregate neighborhood information of each node. In the spectral GCN architecture, a frequency filter is used instead. A graph Fourier transform operator or a graph wavelet transform operator first transforms the raw graph signal to the spectral domain, then the spectral GCN uses the coe"cients from the graph Fourier transform or graph wavelet transform to compute the convolved features. The spectral filter is defined using the graph’s spectral parameters. There are additional challenges to process time-varying graph signals as the signal value at each vertex changes over time. The GCNs are designed to recognize di↵erent spatiotemporal patterns from high-dimensional data defined on a graph. The proposed models have been tested on simulation data and real data for graph signal classification and regression. For the classification problem, we consider the power line outage identification problem using simulation data. The experiment results show that the proposed models can successfully classify abnormal signal patterns and identify the outage location. For the regression problem, we use the New York city bike-sharing demand dataset to predict the station-level hourly demand. The prediction accuracy is superior to other models. We next study graph neural network (GNN) models, which have been widely used for learning graph-structured data. Due to the permutation-invariant requirement of graph learning tasks, a basic element in graph neural networks is the invariant and equivariant linear layers. Previous work by Maron et al. (2019) provided a maximal collection of invariant and equivariant linear layers and a simple deep neural network model, called k-IGN, for graph data defined on k-tuples of nodes. It is shown that the expressive power of k-IGN is equivalent to k-Weisfeiler-Lehman (WL) algorithm in graph isomorphism tests. However, the dimension of the invariant layer and equivariant layer is the k-th and 2k-th bell numbers, respectively. Such high complexity makes it computationally infeasible for k-IGNs with k > 3. We show that a much smaller dimension for the linear layers is su"cient to achieve the same expressive power. We provide two sets of orthogonal bases for the linear layers, each with only 3(2k & 1) & k basis elements. Based on these linear layers, we develop neural network models GNN-a and GNN-b, and show that for the graph data defined on k-tuples of data, GNN-a and GNN-b achieve the expressive power of the k-WL algorithm and the (k + 1)-WL algorithm in graph isomorphism tests, respectively. In molecular prediction tasks on benchmark datasets, we demonstrate that low-order neural network models consisting of the proposed linear layers achieve better performance than other neural network models. In particular, order-2 GNN-b and order-3 GNN-a both have 3-WL expressive power, but use a much smaller basis and hence much less computation time than known neural network models. Finally, we study generative neural network models for graphs. Generative models are often used in semi-supervised learning or unsupervised learning. We address two types of generative tasks. In the first task, we try to generate a component of a large graph, such as predicting if a link exists between a pair of selected nodes, or predicting the label of a selected node/edge. The encoder embeds the input graph to a latent vector space via vertex embedding, and the decoder uses the vertex embedding to compute the probability of a link or node label. In the second task, we try to generate an entire graph. The encoder embeds each input graph to a point in the latent space. This is called graph embedding. The generative model then generates a graph from a sampled point in the latent space. Di↵erent from the previous work, we use the proposed equivariant and invariant layers in the inference model for all tasks. The inference model is used to learn vertex/graph embeddings and the generative model is used to learn the generative distributions. Experiments on benchmark datasets have been performed for a range of tasks, including link prediction, node classification, and molecule generation. Experiment results show that the high expressive power of the inference model directly improves latent space embedding, and hence the generated samples.
Show less

Title: Machine learning applications to video surveillance camera placement and medical imaging quality assessment
Creator: Lorente Gomez, Iris
Date: 2022
Description: In this work, we used machine learning techniques and data analysis to approach two applications. The first one, in collaboration with the...
Show moreIn this work, we used machine learning techniques and data analysis to approach two applications. The first one, in collaboration with the Chicago Police Department (CPD), involves analyzing and quantifying the effect that the installation of cameras had on crime, and developing a predictive model with the goal of optimizing video surveillance camera location in the streets. While video surveillance has become increasingly prevalent in policing, its intended effect on crime prevention has not been comprehensively studied in major cities in the US. In this study, we retrospectively analyzed the crime activities in the vicinity of 2,021 surveillance cameras installed between 2005 and 2016 in the city of Chicago. Using Difference-in-Differences (DiD) analysis, we examined the daily crime counts that occurred within the fields-of-view of these cameras over a 12-month period, both before and after the cameras were installed. We also investigated their potential effect on crime displacement and diffusion by examining the crime activities in a buffer zone (up to 900 ft) extended from the cameras. The results show that, collectively, there was an 18.6% reduction in crime counts within the direct viewsheds of all of the study cameras (excluding District 01 where the Loop -Chicago's business center- is located). In addition, we adapted the methodology to quantify the effect of individual cameras. The quantified effect on crime is the prediction target of our 2-stage machine learning algorithm that aims to estimate the effect that installing a videocamera in a given location will have on crime. In the first stage, we trained a classifier to predict if installing a videocamera in a given location will result in a statistically significant decrease in crime. If so, the data goes through a regression model trained to estimate the quantified effect on crime that the camera installation will have. Finally, we propose two strategies, using our 2-stage predictive model, to find the optimal locations for camera installations given a budget. Our proposed strategies result in a larger decrease in crime than a baseline strategy based on choosing the locations with higher crime density.The second application that forms this thesis belongs to the field of model observers for medical imaging quality assessment. With the advance of medical imaging devices and technology, there is a need to evaluate and validate new image reconstruction algorithms. Image quality is traditionally evaluated by using numerical figures of merit that indicate similarity between the reconstruction and the original. In medical imaging, a good reconstruction strategy should be one that helps the radiologist perform a correct diagnosis. For this reason, medical imaging reconstruction strategies should be evaluated on a task-based approach by measuring human diagnosis accuracy. Model observers (MO) are algorithms capable of acting as human surrogates to evaluate reconstruction strategies, reducing significantly the time and cost of organizing sessions with expert radiologists. In this work, we develop a methodology to estimate a deep learning based model observer for a defect localization task using a synthetic dataset that simulates images with statistical properties similar to trans-axial sections of X-ray computed tomography (CT). In addition, we explore how the models access diagnostic information from the images using psychophysical methods that have been previously employed to analyze how the humans extract the information. Our models are independently trained for five different humans and are able to generalize to images with noise statistic backgrounds that were not seen during the model training stage. In addition, our results indicate that the diagnostic information extracted by the models matches the one extracted by the humans.
Show less

Title: SOLID-STATE SMART PLUG DEVICE
Creator: Deng, Zhixi
Date: 2022
Description: Electrical faults are a leading cause of residential fire, and flexible power cords are particularly susceptible to metal or insulation...
Show moreElectrical faults are a leading cause of residential fire, and flexible power cords are particularly susceptible to metal or insulation degradation that may lead to a variety of electrical faults. Smart Plugs are a type of plug-in device controlling electrical loads via wireless communication for consumer market. However, there is lack of circuit protection features in existing Smart Plug products. Moreover, there is no previous product or research on Smart Plug with circuit protection features. This thesis introduces a new Smart Plug 2.0 concept which offers all-in-one protection against over-current, arc, and ground faults in addition to the smart features in Smart Plug products. It aims at preventing fire and shock hazards caused by degraded or damaged power cords and electrical connections in homes and offices. It offers microsecond-scale time resolution to detect and respond to a fault condition, and significantly reduces the electrothermal stress on household electrical wires and loads. A new arc fault detection method is developed using machine learning models based on load current di/dt events. The Smart Plug 2.0 concept has been validated experimentally. A 120V/10A solid-state Smart Plug 2.0 prototype using power MOSEFTs is designed and tested. It has experimentally demonstrated the comprehensive protection features against all types of electrical faults.
Show less

repository.iit

Search the repository

Enabled Filters

Refine Results

Type

Date

Department

Subject

Creator

Rights