In this thesis, we address a fundamental challenge in modern deep learning: uncertainty estimation. While deep neural networks have achieved remarkable success across diverse and impactful applications-from robotics to large language models and advanced information retrieval-their ability to assess the reliability of predictions remains limited. This gap presents a significant challenge as these systems are increasingly deployed in high-stakes, real-world scenarios. As reliance on machine learning grows, so does the need for reliable and robust models capable of adapting to uncertainty. Despite its importance, uncertainty estimation in deep learning still faces significant challenges, including scalability, efficiency, and adaptability.

First, we address one of the most significant challenges of current deep learning methods: computational complexity during training and inference. One of the most popular, robust, and effective methods for deep learning uncertainty estimation-deep ensembling-suffers from both issues, making it impractical for many applications. To tackle training complexity, we propose Masksembles, which allows training only a single model while enabling ensembling during inference. This approach reduces training costs while preserving uncertainty estimation quality. Masksembles improves efficiency and provides a seamless interpolation between MC-Dropout and Deep Ensembles, combining their strengths in a computationally efficient framework. We demonstrate the effectiveness of Masksembles in a synthetic crowd counting setup, where models trained on synthetic data often struggle to generalize to real-world images due to domain shifts. Using Masksembles, we train a model on both labeled synthetic and unlabeled real images through an uncertainty-based pseudo-labeling procedure. This approach achieves robust cross-domain adaptation, outperforming state-of-the-art methods while keeping inference overhead minimal.

Furthermore, we introduce the idempotence property of neural networks for uncertainty estimation, proposing ZigZag, a sampling-free method that is efficient, generic, and produces state-of-the-art uncertainty estimates. ZigZag trains the network to produce consistent outputs with and without additional prediction information, using the difference as a measure of uncertainty. It performs on par with deep ensembles while being significantly more computationally efficient. Building on ZigZag, we propose Idempotent Test-Time Training (ITTT), a domain-agnostic framework for addressing distribution shifts. By using ZigZag's uncertainty score as a test-time training loss, ITTT aligns representations with the training distribution during inference, enhancing model performance. It is versatile across diverse tasks and works seamlessly with any model, including MLPs, CNNs, and GNNs-unlike existing test-time training methods.

Finally, we propose an uncertainty estimation method for iterative architectures, leveraging the convergence rates of successive outputs. This approach achieves state-of-the-art uncertainty estimation quality, enabling Bayesian Optimization to effectively explore beyond training distributions for aerodynamic shape optimization and providing efficient out-of-distribution detection for road detection in aerial images.