| Summary: | This thesis focuses on the crucial challenge of enhancing the resilience of deep learn- ing models against natural corruptions. Although deep learning models have the potential to bring about significant advancements in various fields, they are susceptible to failure when faced with scenarios that differ from their training data, such as noise, blur, weather changes, and digital artifacts. Such failures can have serious implications for human safety, such as accidents caused by self-driving cars. To tackle this issue, this research investigates the correlation between robustness and entropy, and proposes a novel Gaussian adversarial training method to enhance the corruption resilience of image classification models. Furthermore, to expand the understanding of robust deep learning from images to videos, this thesis establishes a large-scale benchmark for assessing video classification robustness, and conducts a comprehensive study using state-of-the-art deep learning models and techniques. The study reveals that techniques for improving video model robustness have been under-explored. Consequently, this research explores approaches involving diverse data augmentations and consistency regularizations. Lastly, inspired by the temporal coherence nature of videos, a test-time optimization technique is proposed to enhance efficiency and effectiveness. The findings of this research bear significant implications for the development of safe and reliable AI systems, paving the way for the widespread deployment of these technologies in real-world applications.
|
|---|