Computer science

Recent developments in Artificial Intelligence (AI), particularly Large Language Models (LLMs), have provided powerful tools for Natural Language Processing (NLP) tasks like sentiment analysis. However, their fine-tuning and deployment present challenges, specifically in terms of computational efficiency and high training costs. To address these challenges, this work applies optimization techniques such as Quantized Low-Rank Adaptation (QLoRA) for parameter-efficient fine-tuning, followed by Generalized Post-Training Quantization (GPTQ) on the Llama 3.1 LLM. To evaluate these optimizations, we apply the model to a practical task: hate speech detection, using a curated dataset comprising of X (formerly Twitter) posts. Overall, the optimized model achieved a 67% reduction in size along with significant improvements in classification accuracy and inference speed compared to the base model.

Author Keywords: Generalized Post-Training Quantization, Large Language Models, Low-Rank Adaptation, Parameter-Efficient Fine-Tuning, Quantized Low-Rank Adaptation

This research presented an AI-driven methodology for precise COVID-19 detectionin medical diagnostic imaging using chest X-ray images. The primary focus was on developing an optimized model using convolutional neural networks (CNNs) and leveraging transfer learning as a low-level feature extractor method. Significant attention was also so given to data transformation and enhancement techniques to improve the information content and distinguishability of non-linearities. The proposed methodology enabled the flexibility to apply multiple models within the framework and identify the most suitable model for the specific task at hand. By emphasizing state-of-the-art CNN models and employing a strategic exploration-exploitation trade-off, this study identified a robust model with heightened accuracy. The results demonstrated a model accuracy of 90.11%, a sensitivity of 91.16%, and a precision of 89.19%, highlighting the model's effectiveness in accurately identifying both true positive and true negative test cases.

Author Keywords: Automatic Bayesian Optimization, Convolution Neural Networks, COVID- 19 Detection, Deep Learning, Medical Diagnostic Imaging, Transfer Learning

This thesis presents a comprehensive investigation into machine learning approaches forlandmine detection using thermal imagery. It addresses both classification and precise lo- calization challenges that are integral for humanitarian demining operations. The research encompasses two complementary methodological frameworks: comparative evaluation of traditional machine learning versus deep learning approaches, then followed by an imple- mentation of hyperparameter optimization for enhanced safety performance. The foundation of the study demonstrates that traditional machine learning methods achieve competitive classification performance. Conventional models achieved significant performance with the Random Forest and RestNet50 respectively scoring accuracies of 91.88% and 94.29%, though struggle to achieve >10% when tasked with classification and localization. Expanding on this foundation, we addresses this gap through multi-task learn- ing frameworks that simultaneously optimize for both detection and precise localization. Through systematic hyperparameter tuning across 64 configurations, the optimized multi- task approach achieves 90% detection accuracy with 92% precision while providing precise bounding box localization, representing a 37.5% reduction in false negatives. These find- ings demonstrate that while traditional machine learning offers computational efficiency for basic detection, multi-task deep learning frameworks provide significant performance gains when requiring precise spatial localization, which is an important requirement in demining operations.

Author Keywords: computer-vision, demining, humanitarian, Landmines, Multi-Task Learning, YOLO

Video games have always been intrinsically linked with the technology available for the progress of the medium. With improvements in technology correlating directly to improvements in video games, this has recently not been the case. One recent technology video games have not fully leveraged is Cloud technology. This Thesis investigates a potential solution for video games to leverage Cloud technology. The methodology compares the relative performance of a Local, Cloud and a proposed Hybrid Model of video games. We find when comparing the results of the relative performance of the Local, Cloud and Hybrid Models that there is potential in a Hybrid technology for increased performance in Cloud gaming as well as increasing stability in overall game play.

Author Keywords: cloud, cloud gaming, streaming, video game

Predicting vehicle speed at critical road segments, such as pedestrian crossings during left-turn maneuvers at signalized intersections, is essential for improving traffic safety and supporting autonomous driving systems. This thesis presents a novel two-stage hybrid deep learning framework enhanced with reinforcement learning to forecast vehicle left-turn speed at pedestrian crossings.

Using a multivariate time series dataset of vehicle speed and acceleration, the final three seconds of data are intentionally removed to simulate real-world decision-making prior to reaching pedestrian crossings. In stage one, a Convolutional Neural Network (CNN) imputes the removed values. Stage two uses the imputed data to forecast speed, combining Temporal Convolutional Networks (TCNs) and Long Short-Term Memory (LSTM) networks as feature extractors, followed by a Random Forest Regressor (RFR) for robust speed predictions.

Reinforcement learning is employed to dynamically adjusts the TCN's dilation rate, improving temporal pattern capture. Experimental results show the proposed framework outperforms standalone, hybrid, and state-of-the-art models.

Author Keywords: Data Imputation, Dynamic Dilation, Left-Turn Maneuver, Reinforcement Learning, Temporal Convolutional Network, Time series Forecasting

This thesis presents an approach for automated grading of UML Use Case diagrams. Many software engineering courses require students to learn how to model the behavioural features of a problem domain or an object-oriented design in the form of a use case diagram. Because assessing UML assignments is a time-consuming and labor-intensive operation, there is a need for an automated grading strategy that may help instructors by speeding up the grading process while also maintaining uniformity and fairness in large classrooms. The effectiveness of this automated grading approach was assessed by applying it to two real-world assignments. We demonstrate how the result is similar to manual grading, which was less than 7% on average; and when we applied some strategies, such as configuring settings and using multiple solutions, the average differences were even lower. Also, the grading methods and the tool are proposed and empirically validated.

Author Keywords: Automated Grading, Compare Models, Use Case

Video games have always been intrinsically linked with the technology available for the progress of the medium. With improvements in technology correlating directly to improvements in video games, this has recently not been the case. One recent technology video games have not fully leveraged is Cloud technology. This Thesis investigates a potential solution for video games to leverage Cloud technology. The methodology compares the relative performance of a Local, Cloud and a proposed Hybrid Model of video games. We find when comparing the results of the relative performance of the Local, Cloud and Hybrid Models that there is potential in a Hybrid technology for increased performance in Cloud gaming as well as increasing stability in overall game play.

Author Keywords: cloud, cloud gaming, streaming, video game

In this thesis, we present a framework to model user object-level request patterns in the World Wide Web.This framework consists of three sub-models: one for file access, one for Web pages, and one for storage sites. Web Pages are modelled to be made up of different types and sizes of objects, which are characterized by way of categories.

We developed a discrete event simulation to investigate the performance of systems that utilize our model.Using this simulation, we established parameters that produce a wide range of conditions that serve as a basis for generating a variety of user request patterns. We demonstrated that with our framework, we can affect the mean response time (our performance metric of choice) by varying the composition of Web pages using our categories. To further test our framework, it was applied to a Web caching system, for which our results showed improved mean response time and server load.

Author Keywords: discrete event simulation (DES), Internet, performance modelling, Web caching, World Wide Web

This thesis is part of an industry project which collaborates with an aviation technology company on pilot performance assessment. In this project, we propose utilizing the pilots' training data to develop a model that can recognize the pilots' activity patterns for evaluation. The data will present as a time series, representing a pilot's actions during maneuvers. In this thesis, the main contribution is focusing on a multivariate time series dataset, including preprocessing and transformation. The main difficulties in time series classification is the data sequence of the time dimension. In this thesis, I developed an algorithm which formats time series data into equal length data.

Three classification and two transformation methods were used. In total, there are six models for comparison. The initial accuracy was 40%. By optimization through resampling, we increased the accuracy to 60%.

Author Keywords: Data Mining, K-NN, Machine Learning, Multivariate Time Series Classification, Time Series Forest

In this work, various aspects of neural networks, pre-trained with denoising autoencoders (DAE) are explored. To saturate neurons more quickly for feature learning in DAE, an activation function that offers higher gradients is introduced. Moreover, the introduction of sparsity functions applied to the hidden layer representations is studied. More importantly, a technique that swaps the activation functions of fully trained DAE to logistic functions is studied, networks trained using this technique are reffered to as SPAF-networks. For evaluation, the popular MNIST dataset as well as all \(3\) sub-datasets of the Chars74k dataset are used for classification purposes. The SPAF-network is also analyzed for the features it learns with a logistic, ReLU and a custom activation function. Lastly future roadmap is proposed for enhancements to the SPAF-network.

Author Keywords: Artificial Neural Network, AutoEncoder, Machine Learning, Neural Networks, SPAF network, Unsupervised Learning