More about HKUST
Advances in Transfer Learning: Handling Domains with Set-Valued Outputs and Data Streams with Continuous Covariate Shifts and Noisy Samples
The Hong Kong University of Science and Technology Department of Computer Science and Engineering PhD Thesis Defence Title: "Advances in Transfer Learning: Handling Domains with Set-Valued Outputs and Data Streams with Continuous Covariate Shifts and Noisy Samples" By Mr. Xingzhi ZHOU Abstract: Transfer learning leverages pre-trained knowledge to address new tasks and is classified into two categories: inductive and transductive. Inductive transfer learning involves different tasks and limited labeled data in the target domain. In contrast, transductive transfer learning concerns tasks that are identical but lack labels in the target domain. Although promising, inductive methods often exhibit poor generalization due to scarce data. Meanwhile, transductive methods during inference (i.e., test-time adaptation) face instability caused by continuous covariate shifts, temporal correlations, and noisy data. This dissertation aims to address the above challenges, and its contributions are organized into three distinct studies. The first study addresses inductive transfer learning for traditional Chinese medicine (TCM) prescription prediction, a task with set- valued outputs challenged by limited labeled data and complex symptom-herb relationships. We propose TCM-FTP, a fine-tuning framework that adapts large language models (LLMs) using a curated dataset, DigestDS, derived from clinical documentation. TCM-FTP integrates low-rank adaptation (LoRA) to reduce computational demands and utilizes herb-order randomization for data augmentation, based on the set-valued output property of prescriptions. We empirically show that TCM-FTP significantly improves herb identification and dosage estimation. The second study explores test-time adaptation under continuous covariate shifts and temporal correlation. We introduce ResiTTA, a resilient test-time adaptation framework that mitigates overfitting by softly regularizing batch normalization statistics between source and target domains. To stabilize adaptation, ResiTTA maintains a low-entropy, class-balanced memory bank, enabling teacher-student self-training under an approximate i.i.d. assumption. Rigorous benchmarks demonstrate that ResiTTA consistently outperforms other baselines. The third study explores noisy test-time settings. We propose MoTTA, a model-pruning- based framework that identifies noisy samples via output difference under model pruning (ODP), offering greater robustness than prediction-based filtering. Additionally, MoTTA introduces flatness-aware entropy minimization (FlatEM), guiding optimization toward flatter loss landscapes through zeroth- and first-order constraints. Empirical evaluations confirm that MoTTA achieves superior adaptation performance under noisy conditions. Date: Monday, 25 August 2025 Time: 3:00pm - 5:00pm Venue: Room 2408 Lifts 17/18 Chairman: Prof. Guohua CHEN (CBE) Committee Members: Prof. Nevin ZHANG (Supervisor) Dr. Brian MAK Prof. Dit-Yan YEUNG Dr. Yuan LIU (ISD) Prof. William Kwok Wai CHUENG (HKBU)