Optimal Transport Theory for Privacy Preserving Domain Adaption with Different Label Sets

Project: Research

Project Details

Project Description

1. Developing bridging theory to connect OT and LwLL.
Addressing the above challenges and answering the relevant research questions, this proposal aims to develop a novel OT-based theory to encourage the leverage of OT theory in the context of learning with limited and fewer labels. More specifically, our motivation and incentive are to base on the matching and clustering views of OT to break the symmetry of optimal transport distance between target data distribution and the mixture distribution of target data distribution and either source data distribution or prototypes of source classes. By breaking the symmetry, we create interactions between a target example and its closest neighborhoods to naturally and optimally compress their representations together. This naturally helps to increase inter-distance and decrease intra-distance which is certainly helpful for partial DA, open-set DA, and universal DA settings. More specifically, inspired by the imitation learning viewpoint, we develop OT based imitation learning framework which can be applied to a mixed variety of problems.
2. Privacy Domain Adaptation for Different Label Sets.
By breaking the symmetry between the target data distribution and source-class prototypes which encompass information of each source class, our approach can be naturally applied to the source-free DA setting. We also investigate different approaches to embody source classes by meaningful prototypes which are efficient for transfer learning.
StatusNot started