PhD Oral Exam - Afshar Shamsi Jokandan, Information and Systems Engineering

When studying for a doctoral degree (PhD), candidates submit a thesis that provides a critical review of the current state of knowledge of the thesis subject as well as the student’s own contributions to the subject. The distinguishing criterion of doctoral graduate research is a significant and original contribution to knowledge.

Once accepted, the candidate presents the thesis orally. This oral exam is open to the public.

Abstract

Abstract Scalable Bayesian and Adaptive Learning under Distribution Shift Modern AI models often experience significant performance degradation when deployed in real-world environments, with predictions becoming less reliable and uncertainty estimates poorly calibrated. These challenges are particularly critical in safety-sensitive domains, where privacy restrictions, limited computational resources, and the absence of source data preclude conventional retraining. This thesis investigates key directions for improving the reliability of deep learning systems by examining uncertainty-aware learning, scalable Bayesian inference, and test-time adaptation (TTA), with the goal of enhancing both predictive accuracy and trustworthiness. The first contribution focuses on improving Monte Carlo Dropout (MCD), a widely used and easily adaptable uncertainty estimation technique for models with dropout regularization. Despite its simplicity, MCD often produces poorly calibrated predictions and unreliable uncertainty estimates. We develop an enhanced variant of MCD that improves the alignment between predictive uncertainty and model correctness, resulting in better-calibrated uncertainty estimates. To further address sensitivity to hyperparameter selection, we incorporate Bayesian optimization for automatic tuning, enabling more stable and consistent performance across validation settings. The second contribution introduces Bella (Bella), a scalable Bayesian posterior approximation framework for large neural networks. Bella attaches low-overhead trainable modules to pretrained models and employs particle-based variational inference to approximate the posterior distribution over model parameters efficiently. This design enables uncertainty-aware inference in high-capacity architectures while significantly reducing the computational and memory overhead associated with full Bayesian methods. In addition, the use of multiple parameter particles allows the model to capture diverse solutions, improving robustness under distributional variability. The third contribution addresses reliable adaptation under distribution shift in real-world deployment scenarios. We propose a Bayesian test-time adaptation (BTTA) framework that enables online adaptation without access to labeled data or source-domain samples. By updating a small subset of model parameters and incorporating particle-based diversification, the method explores multiple adaptation trajectories during inference, improving robustness while mitigating model drift and performance degradation during deployment. Finally, we investigate the problem of underspecification in unsupervised TTA, where multiple solutions can achieve similar optimization objectives but exhibit different generalization behavior. We introduce diversification strategies at the input, output, and parameter levels to regularize the adaptation process, reduce overconfidence, and improve stability. These strategies enhance robustness across a range of challenging scenarios, including corruption, label shift, and low-batch adaptation, reducing the tendency of existing methods to converge to suboptimal solutions. Taken together, these contributions advance the development of reliable AI systems by improving uncertainty estimation, enabling scalable Bayesian learning, and strengthening robustness under real-world deployment conditions.