| Summary: | This thesis explores the use of unsupervised methods to address the challenges of distribution shift under strong computational constraints in deep learning vision tasks. Through a series of contributions, it demonstrates the effectiveness of unsupervised learning in enabling efficient and adaptable models suitable for resource-constrained environments. The Diversified Dynamic Routing (DivDR) method showcases the potential of unsupervised clustering for efficient image recognition, while the Domain Partitioning Network (DoPaNet) tackles mode collapse in generative adversarial learning. A systematic unification of pixel-level alignment, feature-level alignment, and pseudo-labeling is proposed for cross-domain object detection, highlighting the power of unsupervised adaptation techniques. The thesis also extends its investigation to online continual learning with label delay, proposing a novel approach to address the challenges of evolving data distributions and delayed feedback. The findings have significant implications beyond vision tasks, with the potential to impact a wide range of real-world applications. As machine learning continues to evolve and tackle increasingly complex challenges, the importance of unsupervised learning in enabling robust, efficient, and adaptable models will only grow. This thesis represents a significant step forward in the development of unsupervised methods, paving the way for future machine learning applications.
|
|---|