Uncertainty-Aware Predictions in Real Time: A Missing Link in Edge Intelligence

This thesis presents a comprehensive exploration into lightweight uncertainty quantification (UQ) in deep learning. The cornerstone of this work lies in addressing various limitations associated with existing uncertainty-aware deep learning algorithms in terms of scalability, tractability, and statistical efficiency, particularly for resource-constrained applications such as in edge intelligence. The thesis begins with the introduction of "Lightweight, Uncertainty-Aware Conformalized Visual Odometry," a novel study detailing embedding of uncertainty estimation within visual odometry, a common computer vision task for autonomous navigation. The work settles on an optimal UQ method called Conformalized Joint Prediction (CJP). This technique is designed to be robust, generalizable, and undemanding, encouraging its potential utility across various tasks without imposing computational burdens. The thesis further advances with "Mutual Information-calibrated Conformal Feature Fusion for Uncertainty-Aware Multimodal 3D Object Detection at the Edge." This study proposes a method for fusing multimodal data streams (e.g., from RGB camera and LiDAR sensors) under a unified UQ framework, enhancing 3D object detection through improved uncertainty calibration techniques based on mutual information. The culmination of these studies is presented in "Echoes of Socratic Doubt: Embracing Uncertainty in Calibrated Evidential Reinforcement Learning," which combines CJP with evidential learning for a comprehensive uncertainty estimation method that synergizes their strengths and mitigates their shortcomings. Applied to deep reinforcement learning, this approach demonstrates the versatility and effectiveness of the proposed technique across both supervised and dynamic stochastic tasks, all while fitting within the constraints of edge computing. Overall, the thesis contributes to the field by offering novel methods for quantifying uncertainty in deep learning tasks and ensuring these advancements are accessible for applications constrained by computational resources. It lays groundwork for future research in uncertainty-aware computer vision, reinforcement learning, and other tasks, emphasizing the importance of reliable and computationally efficient methodologies.