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 inherent adaptability, effortlessly adjusting to changes in data distributions and modifying task strategies to accommodate environmental variations without any external supervision. This adaptability has inspired the field of unsupervised domain adaptation (UDA).Most existing UDA methods rely on access to the source data on which the model was initially trained during the adaptation phase. However, a more practical scenario involves situations where only the trained model is available, rather than the source data. This approach mirrors human learning more closely, as humans do not use previous data directly; instead, their brains are pre-trained on source data and apply this knowledge to new situations. Based on this observation, the field of source-free domain adaptation has emerged. In source-free domain adaptation, only the pre-trained source models and new target data are used during adaptation to a new environment. This dissertation encompasses five significant contributions to this emerging field. First, we explore a scenario where we leverage multiple pretrained source models, each trained on different domains, during the adaptation phase without using source data. We develop an algorithm that effectively combines these models such that the most correlated source model with respect to the target data receives the highest weight, while the least correlated one receives the lowest weight. This approach ensures maximum knowledge transfer from all sources, resulting in final adaptation performance that surpasses any individual source model. This algorithm is designed for a static target distribution, where all target data are available during adaptation and do not change over time. Expanding on this approach, we next consider a scenario where the target data is time-varying and arrives in a streaming fashion. This dynamic setting requires continual adaptation as new data becomes available, presenting unique challenges compared to the static scenario. We then explore two applications of these approaches:(i) adapting to target data from a modality different than the sources, and(ii) adding a new source model to the ensemble of source models with reliance on only a few labeled target data points. Finally, we focus on another emerging field of research called unlearning, where the trained model must forget certain data it has seen during training to meet user privacy concerns. Unlike existing approaches that require access to all training data during the unlearning process, we address this in a source-free manner, needing only the data to be forgotten during the unlearning procedure.