In the pursuit of research in remote areas, robots may be employed to deploy sensor networks. These robots need a method of classifying a surface to determine if it is a suitable installation site. Developing surface classification models manually requires significant time and detracts from the goal of automating systems. We create a system that automatically collects the data using an Unmanned Aerial Vehicle (UAV), extracts features, trains a large number of classifiers, selects the best classifier, and programs the UAV with that classifier. We design this system with user configurable parameters for choosing a high accuracy, efficient classifier. In support of this system, we also develop an algorithm for evaluating the effectiveness of individual features as indicators of the variable of interest. Motivating our work is a prior project that manually developed a surface classifier using an accelerometer; we replicate those results with our new automated system and improve on those results, providing a four-surface classifier with a 75% classification rate and a hard/soft classifier with a 100% classification rate. We further verify our system through a field experiment that collects and classifies new data, proving its end-to-end functionality. The general form of our system provides a valuable tool for automation of classifier creation and is released as an open-source tool.

There is an increasing interest in using deep neural networks in various downstream natural language processing tasks. Such models are commonly used as black boxes, meaning that their decision-making is difficult to interpret. In order to build trust in models, it is crucial to analyse their inner workings which lead to predictions. The need to interpret natural language processing models has induced research on linguistically-informed interpretability. This field revolves around choosing specific linguistic phenomena and inspecting models' capability to capture them without being explicitly trained for it.  This thesis project contributes to the field by assessing the ability of BERT-style models to learn subject number in Lithuanian and English. The experiments revolve around designing diagnostic classifiers which are used to determine if the models are capable of learning this particular linguistic phenomenon. The results show that BERT-style models are capable of implicitly learning the number of a subject both in Lithuanian and English. However, this seems to be harder in Lithuanian, as diagnostic classifiers show a lower accuracy. The study observes that the accuracy of logistic regression diagnostic classifiers fluctuates to a large extent. Fully connected neural network classifiers outperform logistic regression classifiers.

The decision tree is a well-known methodology for classification and regression. In this dissertation, we focus on the minimization of the misclassification rate for decision tree classifiers. We derive the necessary equations that provide the optimal tree prediction, the estimated risk of the tree's prediction, and the reliability of the tree's risk estimation. We carry out an extensive analysis of the application of Lidstone's law of succession for the estimation of the class probabilities. In contrast to existing research, we not only compute the expected values of the risks but also calculate the corresponding reliability of the risk (measured by standard deviations). We also provide an explicit expression of the k-norm estimation for the tree's misclassification rate that combines both the expected value and the reliability. Furthermore, our proposed and proven theorem on k-norm estimation suggests an efficient pruning algorithm that has a clear theoretical interpretation, is easily implemented, and does not require a validation set. Our experiments show that our proposed pruning algorithm produces accurate trees quickly that compares very favorably with two other well-known pruning algorithms, CCP of CART and EBP of C4.5. Finally, our work provides a deeper understanding of decision trees.

This study investigates E-banking Operational Risk Assessment (ORA) to enable the development of a new ORA framework and methodology. The general view is that E-banking systems have modified some of the traditional banking risks, particularly Operational Risk (OR) as suggested by the Basel Committee on Banking Supervision in 2003. In addition, recent E-banking financial losses together with risk management principles and standards raise the need for an effective ORA methodology and framework in the context of E-banking. Moreover, evaluation tools and / or methods for ORA are highly subjective, are still in their infant stages, and have not yet reached a consensus. Therefore, it is essential to develop valid and reliable methods for effective ORA and evaluations. The main contribution of this thesis is to apply Fuzzy Inference System (FIS) and Tree Augmented Naïve Bayes (TAN) classifier as standard tools for identifying OR, and measuring OR exposure level. In addition, a new ORA methodology is proposed which consists of four major steps: a risk model, assessment approach, analysis approach and a risk assessment process. Further, a new ORA framework and measurement metrics are proposed with six factors: frequency of triggering event, effectiveness of avoidance barriers, frequency of undesirable operational state, effectiveness of recovery barriers before the risk outcome, approximate cost for Undesirable Operational State (UOS) occurrence, and severity of the risk outcome. The study results were reported based on surveys conducted with Nigerian senior banking officers and banking customers. The study revealed that the framework and assessment tools gave good predictions for risk learning and inference in such systems. Thus, results obtained can be considered promising and useful for both E-banking system adopters and future researchers in this area.

State and federal agencies spend considerable time and resources to enhance and create habitat for wildlife. Understanding how target and non-target species respond to these efforts can help direct the allocation of limited conservation resources. However, monitoring species response to habitat management comes with several logistical challenges that are exacerbated as the area of geographic focus increases. I used autonomous recording units (ARUs) to mitigate these challenges when assessing Eastern Whip-poor-will (Antrostomus vociferus) response to forest management. I deployed 1,265 ARUs across managed and unmanaged public and private forests from western North Carolina to southern Maine. I then applied a machine learned classifier to all recordings to create whip-poor-will daily detection histories for each survey location. I used detection data and generalized linear models to examine regional, landscape, and site factors that influenced whip-poor-will occurrence. Whip-poor-wills were detected at 399 (35%) survey locations. At the regional scale, occupancy decreased with latitude and elevation. At the landscape scale, occupancy was negatively associated with the amount of impervious cover within 500m, and was positively associated with the amount of oak forest and evergreen forest cover within 1,750m. Additionally, whip-poor-will occupancy exhibited a quadratic relationship with the amount of shrub/scrub cover within 1,500m. At the site-level, occupancy was negatively associated with increased basal area and exhibited a quadratic relationship with woody stem density. Whip-poor-will populations can benefit from the implementation of forestry practices that create and sustain early successional forests within forested landscapes, especially those dominated by oak forest types. The use of ARUs helped overcome several challenges associated with intensive broad-scale monitoring efforts for a species with a limited survey window, but also presented new challenges associated with data management, storage, and analyses.

In this thesis, we focus on building reliable hybrid human-machine classifiers to be deployed in cost-sensitive classification tasks. The objective is to assess ML quality in hybrid classification contexts and design the appropriate metrics, thereby knowing whether we can trust the model predictions and identifying the subset of items on which the model is well-calibrated and trustworthy. We start by discussing the key concepts, research questions, challenges, and architecture to design and implement an effective hybrid classification service. We then present a deeper investigation of each service component along with our solutions and results. We mainly contribute to cost-sensitive hybrid classification, selective classification, model calibration, and active learning. We highlight the importance of model calibration in hybrid classification services and propose novel approaches to improve the calibration of human-machine classifiers. In addition, we argue that the current accuracy-based metrics are misaligned with the actual value of machine learning models and propose a novel metric ``value". We further test the performance of SOTA machine learning models in NLP tasks with a cost-sensitive hybrid classification context. We show that the performance of the SOTA models in cost-sensitive tasks significantly drops when we evaluate them according to value rather than accuracy. Finally, we investigate the quality of hybrid classifiers in the active learning scenarios. We review the existing active learning strategies, evaluate their effectiveness, and propose a novel value-aware active learning strategy to improve the performance of selective classifiers in the active learning of cost-sensitive tasks.

No description available.

Vehicle Action Intention Prediction plays a more and more crucial role in automated driving and traffic safety. It allows automated vehicles to comprehend the other road participants’ current actions, and foresee the upcoming actions, which can significantly reduce the likelihood of traffic accidents, so as to enhance overall road safety. Meanwhile, by anticipating other vehicles’ movements on the road, the ego vehicle can plan its velocity and trajectory in advance, and make more smooth and finer adjustments during the whole driving process, contributing to a more safe and efficient traffic. Furthermore, the intention prediction enables vehicles to respond proactively rather than reactively in traditional ADAS (Advanced Driver Assistance Systems), such as AEB (Automatic Emergency Braking), which facilitates a more preventive and early intervention approach to traffic safety. In normal conditions, traffic behavior is controlled by traffic rules. This thesis explores vehicle behavior using intention prediction models in scenarios where there are no traffic rules. At hand, we have a unique dataset containing vehicle trajectories, collected from 2 cameras installed overhead on a 1-lane narrowing street, where the vehicles need to negotiate their right of way. After pre-processing these data to form specific input structures, we use different classifier models including both traditional methods and deep learning methods to make vehicle action intention predictions. The data was organized in 3-second windows and contained vehicle position and distance to the center of the intersection along with the speed of both vehicles. We compared and evaluated the model performances and found that MLPs (Multi-Layer Perceptrons) and LSTM (Long Short Term Memory) yield the best performance. Furthermore, a feature selection method and features’ importance analysis are also applied to explore which variables influence the model most in order to explain the internal principle of the prediction model. It was found that close to the narrowing street the first and last samples of the position and distance in the time window and the last sample of the speed of both vehicles were found to influence the model performance the most. Further away from the narrowing street, the first and last samples of the position of the vehicle have a higher influence on the model.

In this thesis, we introduce a radial basis function network approach to solve a color image classification problem in a real time industrial application. Radial basis function networks are employed to classify the images of finished wooden parts in terms of their color and species. Other classification methods are also examined in this work. The minimum distance classifiers are presented since they have been employed by the previous research. We give brief definitions about color space, color texture, color quantization, color classification methods. We also give an intensive review of radial basis functions, regularization theory, regularized radial basis function networks, and generalized radial basis function networks. The centers of the radial basis functions are calculated by the k-means clustering algorithm. We examine the k-means algorithm in terms of starting criteria, the movement rule, and the updating rule. The dilations of the radial basis functions are calculated using a statistical method. Learning classifier systems are also employed to solve the same classification problem. Learning classifier systems learn the training samples completely whereas they are not successful to classify the test samples. Finally, we present some simulation results for both radial basis function network method and learning classifier systems method. A comparison is given between the results of each method. The results show that the best classification method examined in this work is the radial basis function network method. / Master of Science

Several ready-to-use supervised classifiers perform predictively well in large-sample cases, but generally, the same cannot be expected when transitioning to high-dimensional settings. This can be explained by the classical supervised theory that has not been developed within high-dimensional spaces, giving several classifiers a hard combat against the curse of dimensionality. A rise in parsimonious classification procedures, particularly techniques incorporating feature selectors, can be observed. It can be interpreted as a two-step procedure: allowing an arbitrary selector to obtain a feature subset independent of a ready-to-use model and subsequently classify unlabelled instances within the selected subset. Modeling the two-step procedure is often heavy in motivation, and theoretical and algorithmic descriptions are frequently overlooked. In this thesis, we aim to describe the theoretical and algorithmic framework when employing a feature selector as a pre-processing step for Support Vector Machine and assess its validity in high-dimensional settings. The validity of the proposed classifier is evaluated based on predictive performance through a comparative study with a state-of-the-art algorithm designed for advanced learning tasks. The chosen algorithm effectively employs feature relevance during training, making it suitable for high-dimensional settings. The results suggest that the proposed classifier performs predicatively superior to the Support Vector Machine in lower input dimensions; however, a high rate of convergence towards a performance comparable to the Support Vector Machine tends to emerge for input dimensions beyond a certain threshold. Additionally, the thesis could not conclude any strict superior performance between the chosen state-of-the-art algorithm and the proposed classifier. Nonetheless, the state-of-the-art algorithm imposes a more balanced performance across both labels.