UC Riverside

Deep neural networks have demonstrated remarkable efficacy across a wide range of tasks, yet they face a significant limitation in their ability to adapt to distributional shifts. In contrast, humans possess an inherent adaptability, effortlessly adjusting to changes in data distributions and modifying task strategies to accommodate environmental variations. To fully harness the potential of deep learning models and enhance their practical applicability, it is crucial to impart robustness to distributional shifts. This dissertation addresses this need by presenting algorithms to empower deep learning models with the capacity to seamlessly navigate diverse forms of distributional shifts.

The dissertation encompasses four significant contributions. First, we explore the adaptation of a person re-identification model trained on labeled data from a single camera to other cameras in the network using only unlabeled data. By optimizing temporal consistency across frames in unlabeled videos, the model acquires generalizable representations. Second, we address the adaptation of 2D human pose estimation models to different imaging conditions, achieving adaptation through pre-trained models and unlabeled data from the target domain. Leveraging a pre-built human pose prior that captures plausible human poses, labeled data becomes unnecessary for the adaptation process.

Expanding the concept of adaptation beyond static tasks, we proceed to tackle sequential decision-making problems. It demonstrates how imitation learning can be executed when expert demonstrations originate from domains with distinct morphologies compared to the learning agent. By utilizing cyclic state transformation consistency and value function consistency, a transformation function is learned to render demonstrations comprehensible to the agent.

Finally, we shift focus towards adapting to user constraints, a critical aspect of deep learning model adaptability. It addresses the challenge of adapting multi-task models to changing user preferences by introducing a hypernetwork controller capable of dynamically modifying model architecture and weights without necessitating re-training.

By bridging the gap between human adaptability and the limitations of current models, this dissertation paves the way for deep learning to become more versatile and applicable in real-world scenarios, unlocking its full potential across various domains.