Search

This paper aims to present a study on developing a program that assists nonprofit organizations in determining the ideal location for building their facilities based on community needs, thus maximizing their potential for success. The study highlights the importance of location in the success of nonprofit organizations, and the challenges they face in identifying suitable areas for their operations. The paper reviews existing literature on nonprofit organizations, location analysis, and data analysis techniques, and proposes a methodology for developing the program. The methodology involves data collection and analysis, and machine learning algorithms to predict community needs. The program provides a user-friendly interface for nonprofit organizations to access and analyze the data and offers recommendations for suitable locations based on their criteria. The study concludes that the proposed program can be a valuable tool for nonprofit organizations to make informed decisions about their location and maximize their potential for success in serving their communities.

Out-of-distribution (OOD) detection aims to identify test examples that do not belong to the training distribution and are thus unlikely to be predicted reliably. Despite a plethora of existing works, most of them focused only on the scenario where OOD examples come from semantic shift (e.g., unseen categories), ignoring other possible causes (e.g., covariate shift). In this paper, we present a novel, unifying framework to study OOD detection in a broader scope. Instead of detecting OOD examples from a particular cause, we propose to detect examples that a deployed machine learning model (e.g., an image classifier) is unable to predict correctly. That is, whether a test example should be detected and rejected or not is “model-specific”. We show that this framework unifies the detection of OOD examples caused by semantic shift and covariate shift, and closely addresses the concern of applying a machine learning model to uncontrolled environments. We provide an extensive analysis that involves a variety of models (e.g., different architectures and training strategies), sources of OOD examples, and OOD detection approaches, and reveal several insights into improving and understanding OOD detection in uncontrolled environments.

Software models are used to analyze and understand the properties of the system, providing stakeholders with an overview of how the system should work before actually implementing it. Such models are usually created informally, such as drawing sketches on a whiteboard or paper, especially during the early design phase, because these methods foster communication and collaboration among stakeholders. However, these informal sketches must be formalized to be useful in later applications, such as analysis, code generation, and documentation. This formalization process is often tedious, error-prone, and time-consuming. In an effort to avoid recreating formal models from scratch, this thesis presents SkeMo, a sketch-based software modeling tool. SkeMo is built on a CNN-based image classifier using 3000 input sketches of class diagram components and integrated into the functionality of an existing web-based model editor, the Instructional Modeling Language (IML), with a newly implemented touch interface. SkeMo was evaluated using a ten-fold cross-validation to assess the image classifier and through a user study involving 20 participants to collect metrics and feedback. The results demonstrate the promising potential of sketch-based modeling as an intuitive and efficient modeling practice, allowing users to quickly and easily create models to design complex software systems.

Diabetes is one of the highest causes of death in the world. Diabetes is caused when theblood glucose level is too high in the body. Gradually, high blood glucose leads to heartdisease, stroke, eye, and foot problems. To prevent the dreadful effects among people,early detection is required that would lead to proper medical treatment and change inlifestyle. Therefore, with the rise of machine learning we can predict if a patient hasdiabetes or not. Furthermore, we will integrate the trained model to a web applicationthat will connect the model to generate predictions in real-time considering factorsresponsible for diabetes like body mass index (BMI), age, insulin, etc. In this paper, we areusing the Pima Indian dataset that is originally from the National Institute of Diabetes,Digestive and Kidney Diseases for diabetes prediction model design using machinelearning. The proposed system in this paper is the Soft Voting ensemble classifier. Thealgorithm with the best accurate result was used in making predictions. This model wasdeployed to the web using flask (a python framework), it takes inputs from the user tomake predictions. This model is implemented using python programming language andflask (a web base framework) hosted in GCP. Soft Voting ensemble classifiers evenperform better than other classifiers with an accuracy of 91.55% which is quite promisingconsidering the other classification models in the literature for this problem.

The limitation of DNNs to explain their black box behavior has led to a recent surge of explainability methods. However, there are growing concerns that these explainability methods are not robust and trustworthy. In this paper, we perform the first robustness analysis of Neuron Explanation methods under a unified pipeline and show that these explanations can be significantly corrupted by random noises and well-designed perturbations added to the probing dataset. We find that even adding small random noise with a standard deviation of 0.02 can already change the assigned concepts of up to 28% neurons in the deeper layers. Furthermore, we devise a novel corruption algorithm and show that our algorithm can manipulate the concept of more than 80% neurons by poisoning less than 10\% of probing data. This raises the concern of trusting Neuron Explanation methods in real-life safety and fairness critical applications.

Melanoma is an incredibly dangerous form of skin cancer. In 2023 alone, the American Cancer Society estimates that about 7,990 Americans will die from new melanoma cases. Image analysis tools that automate the diagnosis of melanoma will improve dermatologists' diagnostic accuracy and should provide a significantly better detection of melanoma. In this paper, we present an application of modern deep learning architectures using the images of skin lesions in combination with patient-level contextual information to detect the presence of melanoma. More specifically, we used the latest variant of the residual network, the ResNeSt model, to leverage its deep architecture and channel-wise attention mechanisms for melanoma detection. In parallel, we used a multi-layer perceptron model to process patient-level data, and fused its output with the extracted features from the ResNeSt model. To compare and contrast performances, we trained a similar model using a smaller, more generic, convolutional neural network in addition to the larger ResNeSt network. After properly tuning all associated hyperparameters, our primary model obtained a balanced accuracy of 0.853 and a ROC-AUC score of 0.93. The split-attention network was shown to improve the total ROC-AUC by approximately 9%, suggesting that our model was able to successfully leverage the ResNeSt design to extract the most important features and most relevant cross-feature interactions for the detection of melanoma.

We introduce the problem of model-extraction attacks in cyber-physical systems in which an attacker attempts to estimate (or extract) the feedback controller of the system. Extracting (or estimating) the controller provides an unmatched edge to attackers since it allows them to predict the future control actions of the system and plan their attack accordingly. Hence, it is important to understand the ability of the attackers to perform such an attack. In this paper, we focus on the setting when a Deep Neural Network (DNN) controller is trained using Reinforcement Learning (RL) algorithms and is used to control a stochastic system. We play the role of the attacker that aims to estimate such an unknown DNN controller, and we propose a two-phase algorithm. In the first phase, also called the offline phase, the attacker uses side-channel information about the RL-reward function and the system dynamics to identify a set of candidate estimates of the unknown DNN. In the second phase, also called the online phase, the attacker observes the behavior of the unknown DNN and uses these observations to shortlist the set of final policy estimates. We provide theoretical analysis of the error between the unknown DNN and the estimated one. We also provide numerical results showing the effectiveness of the proposed algorithm.

In this paper we conduct transfer learning on the Faster R-CNN ResNet-50 FPN V2 object detection model from Benchmarking Detection Transfer Learning with Vision Transformers [12] to a new problem space: detecting what parts of a webpage can be clicked on. The model predicts the bounding boxes and class labels for what part of a website image can be clicked on, and what the result will be. We found that training the feature extraction layers of the model was detrimental to the learning (354% higher loss after 10 epochs), but that training the RPN layers was not detrimental. We additionally found that the most effective resizing for website imagery was to 224 x 224 pixels. Since the meta trainer we developed did not produce usable data, we experiment on using the data diversity approach in active learning to select the most representative images to train on. Compared to our control, the model trained on the active learning dataset scored 12.5% better on the validation data. However, with such a small dataset of 10 labeled images, the model was unable to generalize and did not predict any of the 15 objects on a labeled image outside the training set.

I propose four network architectures for improving security in machine learning model training where training data must remain secure from the model trainer and the model must be secure from the data owner. The architectures are numbered 1 through 4 with the first being the least secure and 4 being the most. Architecture 4 is the final iteration implemented in practice and facilitates end-to-end security for the Provider and one or more Consumer’s.Each architecture is comprised of multiple processes running concurrently and performing distinct tasks. The processes are deployed in docker containers with strictly defined networks to allow communications between processes only where necessary to uphold security objectives. Techniques such as attestation and cryptography are applied among various other technologies including FarmStack Trusted Connectors and Docker. All of which were employed to reduce data and application security threats further. The important contributions • Plausible architectures, which provide security to the Provider of the raw data and Con- sumer of that data, for training models in the agricultural industry. • A previously untested application of FarmStack Trusted Connectors. • The development of the Internal Data Handler (IDH), which replaced disk storage as an internal cryptographic file storage server. My results prove that of the range of architectures described in this paper, the only worth deploying are architectures 2 and 4. Architecture 2 has far from the same security guaran- tees as architecture 4 but is approximately 3.7 times faster than architecture 4 in its current state. Architecture 4, while slower, addresses all objectives of the threat model. There are many objectives in the threat model not addressed by Architecture 2. Hence, Architecture 4 is recommended over Architecture 2.

Automatic Plate Recognition plays an important roll in intelligent transportation systems. However, most license plate recognition methods work under restricted conditions like slow speed and good illumination. That is a restriction on industrial application. In this thesis, the constraints are relaxed by vanished points distortion-recovery method and denoising method. This thesis implements a license plate recognition method by morphological edge detection method and convolution neural network recognition method. The thesis is constructed contributes to several papers optimization methods. The proposed approach can be trained for recognition of country-specific license plates. More than 500 images are collected for training and over 300 images are collected for recognition test. This paper achieves 97.05% on license plate recognition for detecting total characters and numbers of the license plates.License plate recognition consists three parts, pre-processing image, locating license plate and identifying license numbers and characters. License plate location is important to obtain license images and plays a key role in identifying plates. The plate recognition has two major steps, character separation and identification. In this paper, machine learning method is applied for license plate recognition.

Many techniques for analyzing cardiovascular health rely on cardiac magnetic resonance images that have been segmented to identify various components of the heart. Manually segmenting these images is cumbersome and prone to variability, which calls for the development of accessible automation tools for cardiac researchers. In order to benefit the developing symbiosis between machine learning and medicine, such tools must be accurate, efficient and inferentially transparent. This paper introduces a U-Net-based pipeline for left ventricular segmentation of short-axis CMRs. The U-Net, a Fully Convolutional Network known for its success in biomedical image segmentation, is a natural candidate for our task. This work constitutes the core of a larger-scale project focused on improving human disease models through the acceleration of animal cardiac research. Accordingly, experiments discussed here leverage both human and animal data to explore the efficacy of image processing and model training strategies.This paper focuses on optimizing our U-Nets for resource constrained environments, and demonstrates that these models require only a fraction of their typical convolutional filters. This reduction affords efficiency with the added benefit of explainability by improving the practicality of visualizing learned features. Inference can also be further optimized by pruning trained models without any loss of accuracy or the need for retraining. Specifically, we show that U-Nets with less than 2% of their original parameters train in minutes on a single GPU and achieve Dice scores above 0.95 on multiple CMR datasets. Furthermore, inference on hundreds of images can be performed in seconds on a laptop.

Class Imbalance is one of the numerous problems that affect the performance of a classifier. Data is said to be imbalanced when one or more classes have much fewer instances compared to other classes. In the process of learning from data, having a low number of instances of one class, results in a poor model. The Model tends to be partial to the class having a large number of instances. The issue of Imbalanced data sets is always ongoing research and there are several solutions to cope up with it in the case of two-class classification problems. However, for the multi-class imbalanced data, not every traditional technique helps to solve the problem. In the case of multi-class imbalanced data, few classes can have an extremely low number of instances and few can be moderately low. Balancing the distribution in such cases is a challenge. It is important to make sure that a model can differentiate between all available classes in the data.Apart from imbalanced data, Concept drifts also make a difference in learning a model. Concept Drifts are very likely to occur. For instance, a model that is trained on summer shopping trends can no longer be suitable to predict winter shopping trends. This paper is an analysis of how a combination of existing solutions for the above-mentioned problems behave on few synthetic and real-world data.

This paper explores the frequency at which SQL/PHP posts on the website Stackoverflow.com contain code susceptible to SQL Injection, a common database vulnerability. Specifically, we analyze whether other users give notice of the vulnerability or provide an answer that is secure. The majority of questions analyzed were vulnerable to SQL Injection and were not corrected in their answers or brought to the attention of the original poster. To mitigate this, we present a machine learning bot which analyzes the poster’s code and alerts them of potential injection vulnerabilities, if necessary.

Selective pressure and molecular phenotype provide complimentary perspectives on functional properties of the human genome. In this dissertation, I develop computational methods for identifying collections of molecular phenotypes called functional classes, that are optimally informative about recent selective pressure in humans. Aggregating selective pressure across genomic positions within each functional class produces a score representing the probability that a position evincing a class-associated molecular phenotype is under selective pressure. A class’s score is interpreted as a measure of potential for fitness-influencing genomic function. Functional classes and attendant selective pressure scores are developed over the course of two papers. In the first paper, I investigate three ENCODE cell-types and develop a non-parametric representation of covariates from four genomic properties including DNase-seq, RNA-seq, chromatin state, and protein coding annotation. The resultant 624 classes and attendants scores, are shown to predict eQTL, transcription factor binding, and enhancers as well as or better than contemporary methods using high dimensional functional covariates, or selective constraint alone. The interpretation of the score as selective pressure is also shown to be consistent with previous measures of genome-wide selective pressure. In the second paper, I expand the cell-type cohort to 115 Epigenomic Roadmap cell-types and nine genomic properties including splicing, transcription factor binding, and small RNA-seq. Complexity constraints are developed to reduce the number of functional classes from more than 1.2 million possibilities to 61. The resultant functional classes, genomic segmentations, and positional scoring (FitCons2 scores) are used to detect small features including disease associated variation from HGMD and ClinVar clinical databases. FitCons2 scores are shown to have power comparable or superior to contemporary methods designed specifically to detect such features. Functional classes and scores are also shown to identify cell-type specific regulatory behavior of promoters and enhancers, while highlighting regulatory relationships between differing cell-types and developmental stages. I demonstrate how cell-type sensitivity in FitCons2 scores can be used to address an unsolved biological problem in characterizing transcription factor binding in craniofacial enhancers that are believed to differentiate neural development between humans, chimpanzees, and Neanderthals.

This thesis is an investigation into the performance ramifications of making specialized component reservations for machine learning workloads in virtualized environments. The reliance of machine learning on floating-point operations makes graphics processing units an important part of processing these workloads quickly. While virtualization is one of the most widely-used consolidation techniques used in data centers of all sizes, compatibility issues between graphics processing units and virtualization have slowed the adoption of virtualization for machine learning workloads. To that end, this paper discusses the motivations and history behind virtualization and the application-specific acceleration devices used and how they are applied to machine learning on various public and private computing platforms. This is followed by a presentation of an experimental framework for testing the impact of controlling for non-uniform memory access when running machine learning workloads. Using this framework, a series of experiments were performed and documented in this thesis that test multiple placement configurations for graphics processing units in a virtualized system and how throughput of data from the host system to the device was affected. Current virtualization platforms offer recommendations to use these settings, but do not talk about the specific impacts of implementing them. Based on the results of the experiments, configuration parameters and placement recommendations are presented along with information about how these settings can help optimize the machine learning pipeline and the potential pitfalls to their use.

RankBoost is a famous boosting algorithm for the ranking problem, which is intuitively essential in many settings, such as information retrieval and recommendation systems. Bipartite RankBoost is a faster and more efficient version of RankBoost since it can reduce its time complexity down from quadratic time to linear time. We usually implement Bipartite RankBoost in web search engines to rank files or web pages based on their different relevance in a given query. RankBoost+, an improved and tested version of RankBoost, is a well-designed boosting algorithm with a brand new error function and it not only has good empirical behavior but also shows excellent theoretical properties. In this paper, we will pointed out two flaws in the original Bipartite RankBoost first with examples and solid proofs, and such weak points are also shown in the original RankBoost algorithm. However, RankBoost+ successfully fixes those problems and even gives better performance in data analysis, so we will implement the basic concept and structure of RankBoost+ to improve the performance of Bipartite RankBoost.

Robustness and counterfactual bias are usually evaluated on a test dataset. However, are these evaluations robust? In other words, if a model is robust or unbiased on a test set, will the properties still hold under a slightly perturbed test set? In this paper, we propose a ``double perturbation'' framework to uncover model weaknesses beyond the test dataset. The framework first perturbs the test dataset to construct abundant natural sentences similar to the test data, and then diagnoses the prediction change regarding a single-word substitution. We apply this framework to study two perturbation-based approaches that are used to analyze models' robustness and counterfactual bias. (1) For robustness, we focus on synonym substitutions and identify vulnerable examples where prediction can be altered. Our proposed attack attains high success rates ($96.0\%$\textendash$99.8\%$) in finding vulnerable examples on both original and robustly trained CNNs and Transformers. (2) For counterfactual bias, we focus on substituting protected tokens (e.g., gender, race), and measure the shift of the \emph{expected} prediction. In the experiments, our method reveals the hidden model bias even if the test set is adversarially chosen.

With the rapid development of the worldwide network, faster Internet will attract more users. Cisco White Paper [1] predicted that the number of Internet users will grow from 51% to 66% of the global population. Also, video traffic will be 80% of all Internet traffic in 2022, which means there will be many more network requests in communications between client and server. The critical problem becomes how to use the video streaming requests to improve the quality of experience (QoE).

Coordination failure reduces match quality among employers and candidates in the job market, resulting in a large number of unfilled positions and/or unstable, short-term employment. Centralized job search engines provide a platform that connects directly employers with job-seekers. However, they require users to disclose a significant amount of personal data, i.e., build a user profile, in order to provide meaningful recommendations. In this paper, we present PrivateJobMatch -- a privacy-oriented deferred multi-match recommender system -- which generates stable pairings while requiring users to provide only a partial ranking of their preferences. PrivateJobMatch explores a series of adaptations of the game-theoretic Gale-Shapley deferred-acceptance algorithm which combine the flexibility of decentralized markets with the intelligence of centralized matching. We identify the shortcomings of the original algorithm when applied to a job market and propose novel solutions that rely on machine learning techniques. Experimental results on real and synthetic data confirm the benefits of the proposed algorithms across several quality measures. Over the past year, we have implemented a PrivateJobMatch prototype and deployed it in an active job market economy. Using the gathered real-user preference data, we find that the match-recommendations are superior to a typical decentralized job market---while requiring only a partial ranking of the user preferences.

Twitter has quickly become one of the most popular social media sites. It has 313 million monthly active users, and 500 million tweets are published daily. With the massive number of tweets, Twitter users share information about a location along with the temporal awareness. In this work, I focus on tweets where author of the tweets exclusively mentions a location in the tweet. Natural language processing systems can leverage wide range of information from the tweets to build applications like recommender systems that predict the location of the author. This kind of system can be used to increase the visibility of the targeted audience and can also provide recommendations interesting places to visit, hotels to stay, restaurants to eat, targeted on-line advertising, and co-traveler matching based on the temporal information extracted from a tweet. In this work I determine if the author of the tweet is present in the mentioned location of the tweet. I also determine if the author is present in the location before tweeting, while tweeting, or after tweeting. I introduce 5 temporal tags (before the tweet but > 24 hours; before the tweet but < 24 hours; during the tweet is posted; after the tweet is posted but < 24 hours; and after the tweet is posted but > 24 hours). The major contributions of this paper are: (1) creation of a corpus of 1062 tweets containing 1200 location named entities, containing annotations whether author of a tweet is or is not located in the location he tweets about with respect to 5 temporal tags; (2) detailed corpus analysis including real annotation examples and label distributions per temporal tag; (3) detailed inter-annotator agreements, including Cohen's kappa, Krippendorff's alpha and confusion matrices per temporal tag; (4) label distributions and analysis; and (5) supervised learning experiments, along with the results.

Breast cancer is the most common diagnosed cancer for the worldwide women. Due to the multiformity of the clinical behaviors, it is difficult to predict and diagnosed only with clinical information. In order to find out a better solution to make some prediction in the early stage, the genome wide analysis is introduced. In this paper, we got the dataset from The Cancer Genome Atlas (TCGA) database to find a best predictive machine learning model. Since the copy number variations (CNVs) is highly related with the breast cancer, CNVs is used as a fundamental indicator of each genome segmentation in the study. Based on the start and the end positions, the datasets can be sorted and reorganized into five grouping sets. We tested the predictive power of the Gradient Boosting Machine, Distributed Random Forest, XGBoost and Deep Learning Neural Network. With the different genome segmentation grouping dataset and different machine learning models, we finally found the Gradient Boost Machine is the most powerful model for this problem. It can finally reach AUC of 0.756799 after 15-fold cross validation trained with “merged” grouping dataset.

Computational creativity researchers interested in applying machine learning to computer composition often use the music of J.S. Bach to train their systems. Working with Bach, though, requires grappling with the conventions of tonal music, which can be difficult for computer systems to learn. In this paper, we propose and implement an alternate approach to composition and harmonization of chorales based on pitch-relative note encodings to avoid tonality altogether. We then evaluate our approach using a survey and expert analysis, and find that pitch-relative encodings do not significantly affect human-comparability, likability or creativity. However, an extension of this model that better addresses the criteria survey participants used to evaluate the music, such as instrument timbre and harmonic dissonance, still shows promise.

Educational Data Mining is a nascent, but rapidly growing field which has typically applied data mining techniques to educational data extracted from digital systems such as Learning Content Management Systems and Intelligent Tutoring Systems. The research thus far has been able to identify interesting patterns in the ways students learn when using these systems, but an analysis of students' ordinary problem-solving processes remains unexplored. In this work we apply data mining and machine learning techniques to a digital data set of students' ordinary, handwritten coursework in the context of a Mechanical Engineering course. This work makes four major contributions. It is the first, to our knowledge, study in which data mining and machine learning techniques have been applied to students' problem-solving processes in their ordinary learning environment, using pen and paper at home or in the classroom; because the data set is unique, we provide an in-depth description of the large digital collection of students' handwritten course work we have collected. Second, we investigate novel, discrete and numerical representations of students' handwritten coursework which characterize different aspects of students' ordinary problem-solving processes. Third, we identify patterns in these representations as well as correlations between these representations and course performance using various machine learning and data mining techniques. Last and most importantly, we present interesting conclusions that may be drawn from these patterns and correlations, which allow instructors of the course which was investigated to improve future offerings.