Mourelatos, Zissimos PChavare, Sudeep DhananjayLatcha, MichaelPandey, VijitashwaDrignei, Dorin2026-07-172026-07-172025-01-01https://hdl.handle.net/10323/22155Machine learning has witnessed widespread adoption across various domains, bringing about transformative changes in decision-making, trend prediction, task automation, and personalized experiences. Despite the remarkable predictive capabilities of machine learning models, the associated uncertainty in their predictions remains a critical concern. Uncertainty estimation plays a pivotal role in ensuring robust decision-making, going beyond mere outcome prediction to quantify the models confidence and potential error. This research first presents a review of existing uncertainty quantification techniques in machine learning, including Monte Carlo dropout and ensemble methods, highlighting their advantages in addressing uncertainty as well as their limitations. The dissertation is centered on advancing methods for model uncertainty quantification (UQ) in machine learning. In particular, two novel approaches are proposed to address critical gaps in the existing body of work. The first method operates within the framework of Neural Network Ensembles, wherein multiple neural networks are trained using identical hyperparameters, with variations introduced through different subsets of the training data. Ensemble predictions are subsequently generated, and a novel adaptive Gaussian Process Regression (GPR) model is developed to estimate the associated uncertainty bounds. The adaptive GPR model offers accurate and efficient estimation of uncertainty, making it highly suitable for real-time applications involving either scalar outputs or time-series data. Through this development, the dissertation aims to contribute practical and computationally viable solutions to the field of uncertainty quantification, enhancing the reliability of decision-making in complex and dynamic environments. The second method, while also based on Neural Network Ensembles, is fundamentally different in its objective. After training an ensemble of neural networks, a “Most Probable Model” is constructed to provide a fast and accurate estimation of the mean prediction. Unlike the first method, this approach does not quantify prediction uncertainty or provide confidence bounds. Instead, it offers a significant reduction in computational overhead associated with ensemble averaging, making it particularly advantageous in scenarios where rapid mean prediction is critical. In specific problem domains—further elaborated in subsequent chapters—this method may represent a practical alternative when full uncertainty quantification is either infeasible or unnecessary. Both methods proposed in this work are highly dependent on the accuracy and coverage of the ensemble models. If the ensemble does not capture the ground truth, these methods become ineffective. To mitigate this limitation, an autoencoder-based “Trust Index” is introduced to flag out-of-distribution inputs, providing a warning when the input data fall outside the range of the training set. All the aforementioned methods require training multiple models to form an ensemble of predictions, which can be computationally expensive and time-consuming. To address this limitation, a novel approach based on Conditional Variational Autoencoders (CVAEs) is proposed, wherein synthetic data are generated to emulate the ensemble predictions. This strategy effectively reduces the number of models that need to be trained while maintaining comparable predictive diversity and accuracy.Collectively, these contributions aim to advance the current capabilities in machine learning by introducing accurate uncertainty estimation methods alongside efficient mean prediction strategies, offering both theoretical insights and practical benefits for real-world applications.Artificial IntelligenceEnsembleMachine LearningNeural NetworksUncertainty QuantificationVAE