Transfer Learning with Imprecise Observations: Theory and Algorithms
- Publication Type:
- Thesis
- Issue Date:
- 2024
Open Access
Copyright Clearance Process
- Recently Added
- In Progress
- Open Access
This item is open access.
Transfer learning aims to leverage previously-acquired knowledge from domains with abundant labels (i.e., source domains) to help train a classifier or predictor for the domain with insufficient labels (i.e., target domain). Most existing methods share a common assumption that the observations in the source and target domains are precise. Unfortunately, precise observations are often unavailable in real-world scenarios.
In this research, we consider a new, realistic problem called transfer learning with imprecise observations (TLIMO), where the source or target domains only contain imprecise observations. To develop new theories and construct algorithms for addressing TLIMO problem in various real-world scenarios, this thesis intends to address four orthogonal problems: 1) How to construct a theoretical foundation for imprecise data analysis and handle a simple problem called multi-class classification with imprecise observations (MCIMO); 2) How to handle TLIMO problem in single-source domain scenario; 3) How to handle the multi-source transfer learning problem when the instances in the source or target domains are imprecise; and 4) How to handle the universal domain adaptation (UniDA) problem when the instances in the source or target domains are imprecise.
To address Problem 1), this thesis develops a theoretical foundation for imprecise data analysis based on fuzzy random variables and provides a theoretical analysis of MCIMO problem (Chapter 3). This theoretical analysis ensures that we can always train a fuzzy classifier with high classification accuracy when infinite imprecise instances can be collected. Two new frameworks are constructed for addressing MCIMO problem (Chapters 3 and 4).
To address Problem 2), we extend the theoretical analysis of MCIMO to develop theory for the TLIMO problem. This theory derives a generalization bound to guide model construction. A novel transfer learning approach is then proposed to transfer knowledge from a single-source domain to a single-target domain with imprecise data (Chapter 5).
To address Problem 3), Chapter 6 presents two domain adaptation models to transfer knowledge from multiple source domains with crisp-valued data to a single-target domain with imprecise data.
In Chapter 7, we develop a novel dynamic reweighted loss learning strategy to tackle an unsolved problem in transfer learning, where the distribution discrepancy between the source domain and the target domain in different categories may be significantly different.
In summary, this thesis not only contributes to the theory of transfer learning when the observations are imprecise but also proposes a set of effective algorithms for different transfer learning scenarios.
Please use this identifier to cite or link to this item: