UCLA

This thesis presents three significant contributions to the field of machine learning, with a focus on Variational Autoencoders (VAEs), energy-based models, and education simulations. Firstly, we demonstrate the ability to impose substantial structure on the latent space of VAEs, enabling out-of-distribution data generation, structural hypothesis testing, and the production of augmentations in the latent space. These findings give us new ways to structure and interpret the latent space, creating robustness and explainability. Secondly, we identify a state-of-the-art defense technique using the unsupervised learning approach of energy-based models. This technique effectively defends against several poisoning techniques without requiring excessive additional training time or significantly reducing test accuracy. Lastly, we have developed a simulation for educational purposes that aims to model and comprehend the interactions between humans and machines. This simulation, built on causal information, provides insights into the design of practical educational experiments and highlights the challenges associated with implementing a dynamic Intelligent Tutoring System (ITS) in an educational context. Interestingly, our simulation reveals that heuristic methods continue to perform on par with deep learning techniques in the presence of unknown subpopulation distributions and hidden student states. This suggests that despite the rapid advancements in deep learning, heuristic methods retain their effectiveness in certain scenarios.These findings open new avenues for the application of machine learning techniques and provide a solid foundation for future research in these areas.