Date of Award
5-31-2026
Document Type
Open Access Dissertation
Degree Name
Doctor of Philosophy (PhD)
Department
Computer Science
First Advisor
Xiaohui Liang
Abstract
Modern foundation models are typically trained with large-scale data to ensure good performance. However, certain tasks cannot scale with large amounts of data due to cost and practical constraints, resulting in limited performance. To address this, in this dissertation, I introduce model-task alignment, a general methodology for making a foundation model work well with tasks with limited data. The methodology comprises two parts: aligning downstream tasks with foundation models and aligning foundation models with downstream tasks. To study and validate this methodology, I first focus on speech-based dementia detection, a representative task with limited data, and then extend to general settings. Moreover, I highlight the importance of interpretability in the model-task alignment process, ensuring the model learns something meaningful from limited data rather than merely faking learning. Specifically, I explore interpretability from two aspects: the intermediate process and the training/inference dynamics. At last, I discuss potential future directions in acculturated learning with limited data, automated model-task alignment, and interpretability-driven model development.
Recommended Citation
Zhu, Youxiang, "Towards Interpretable Transfer Learning with Limited Data" (2026). Graduate Doctoral Dissertations. 1165.
https://scholarworks.umb.edu/doctoral_dissertations/1165
Comments
Free and open access to this Campus Access Thesis is made available to the UMass Boston community by ScholarWorks at UMass Boston. Those not on campus and those without a UMass Boston campus username and password may gain access to this thesis through Interlibrary Loan. If you have a UMass Boston campus username and password and would like to download this work from off-campus, click on the “Off-Campus Users” button.