Understanding and Robustifying Sub-domain Alignment for Domain Adaptation
Authors: Yiling Liu, Juncheng Dong, Ziyang Jiang, Ahmed Aloui, Keyu Li, Michael Hunter Klein, Vahid Tarokh, David Carlson
TMLR 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Empirical experiments across various benchmarks validate our theoretical insights, prove the necessity for the proposed adaptation strategy, and demonstrate the algorithm s competitiveness in handling label shift. Empirical experiments across various benchmarks validate our theoretical insights, prove the necessity for the proposed adaptation strategy, and demonstrate the algorithm s competitiveness in handling label shift. In this section, we verify our theoretical results and assess DARSA s efficacy through real-world experiments. We begin by empirically confirming the superiority of the sub-domain-based generalization bound (Theorem 4.10) in Section 6.1. Then, we verify that the assumptions for Theorem 4.10 are empirically satisfied on real-world datasets (details in Appendix B). Next, we demonstrate the vital role of subdomain weight re-balancing in Section 6.2 and show DARSA s robustness to minor weight estimation discrepancies. Lastly, given that our theoretical analysis guarantees that DARSA should have competitive performance in scenarios where the number of classes is not overwhelming, we evaluate DARSA on real-world datasets with this property. Comparing with other state-of-the-art UDA baselines, we validate our theoretical analysis and demonstrate DARSA s effectiveness in real-world applications, including those in medical settings. |
| Researcher Affiliation | Collaboration | Yiling Liuú EMAIL Program in Computational Biology and Bioinformatics, Duke University Juncheng Dongú EMAIL Department of Electrical and Computer Engineering, Duke University Ziyang Jiang EMAIL Meta Platforms, Inc. Ahmed Aloui EMAIL Department of Electrical and Computer Engineering, Duke University Keyu Li EMAIL Department of Electrical and Computer Engineering, Duke University Michael Hunter Klein EMAIL Department of Electrical and Computer Engineering, Duke University Vahid Tarokh EMAIL Department of Electrical and Computer Engineering, Duke University David Carlson EMAIL Department of Civil and Environmental Engineering, Duke University Department of Biostatistics and Bioinformatics, Duke University |
| Pseudocode | Yes | The pseudo-code for DARSA can be found in Appendix D. |
| Open Source Code | Yes | The code to replicate all experiments is available at: https://github.com/yilingmialiu/DARSA_repo |
| Open Datasets | Yes | Experiments on the Digits Datasets. In our Digits datasets experiments, we evaluate our performance across four datasets: MNIST (M) (Le Cun et al., 1998), MNIST-M (MM) (Ganin et al., 2016), USPS (U), and SVHN (S), all modified to induce label distribution shifts. Experiments on the TST Dataset. We use the Tail Suspension Test (TST) dataset (Gallagher et al., 2017) of local field potentials (LFPs) from 26 mice with two genetic backgrounds: Clock19 (a bipolar disorder model) and wildtype. This dataset is publicly available (Carlson et al., 2023). Our study involves two domain adaptation tasks, predicting the current condition home cage (HC), open field (OF), or tailsuspension (TS) from one genotype to the other. Experiments on the Vis DA-2017 Dataset. We further evaluate DARSA on the large-scale Vis DA-2017 dataset (Peng et al., 2017), a challenging synthetic-to-real benchmark with 12 categories. The MNIST, BSDS500, USPS, SVHN, and Vis DA-2017 datasets are publicly available with an open-access license. The Tail Suspension Test (TST) dataset (Gallagher et al., 2017) is available to download at https://research.repository.duke.edu/concern/datasets/zc77sr31x?locale=en for free under a Creative Commons BYNC Attribution-Non Commercial 4.0 International license. |
| Dataset Splits | No | The paper describes how datasets were modified to induce label distribution shifts and mentions that 'For comprehensive details, refer to Appendix F, G, and H'. However, the main text does not provide specific percentages, absolute sample counts, or explicit train/test/validation split ratios for any of the datasets used, nor does it cite predefined standard splits with sufficient detail in the provided main content. |
| Hardware Specification | Yes | The experiments are conducted on a computer cluster equipped with a NVIDIA Ge Force RTX 2080 Ti that has a memory capacity of 11019Mi B. |
| Software Dependencies | No | The paper mentions that the code to replicate all experiments is available at a GitHub repository, but it does not explicitly list any specific software dependencies with version numbers (e.g., Python, PyTorch versions) in the main text. |
| Experiment Setup | Yes | Table 4: Ablation study results. The table presents different configurations and their corresponding prediction accuracy (%) across four experimental setups. Experiment Y D a c Accuracy 0.4 0.35 0.9 1 96.0 ... The objective function of DARSA is defined as follows: E Y LY + DLD + LC, where LY , LD, LC are losses described below with relative weights given by Y and D. |