Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in [1]
A General Representation-Based Approach to Multi-Source Domain Adaptation
Authors: Ignavier Ng, Yan Li, Zijian Li, Yujia Zheng, Guangyi Chen, Kun Zhang
ICML 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | 6. Experiments We show the effectiveness of our method compared with existing ones on widely used datasets in domain adaptation. Further details and empirical studies can be found in Appendix E. 6.1. Datasets and Baselines Datasets. We validate our method on two wellknown benchmarks for domain adaptation: Office Home (Venkateswara et al., 2017) and PACS (Li et al., 2017). ... 6.2. Numerical Results The results for Office-Home and PACS datasets are provided in Table 1. ... 6.3. Ablation Study To evaluate the effectiveness of our special design to capture θ in the target domain, we design two model variants: (1) GAMA-vae: we remove all the VAE related losses; (2) GAMA-theta: we remove the losses brought by θ-related VAEs: Lvae,Y , Lvae,Zch , we also remove Lch, and LY as some items for calculating these losses are related to θ. Experiment results on the Office-Home dataset are shown in Table 2. |
| Researcher Affiliation | Academia | 1Carnegie Mellon University 2Mohamed bin Zayed University of Artificial Intelligence. |
| Pseudocode | No | The paper describes the 'Domain Adaptation Approach' in Section 5 with detailed textual descriptions and equations for the model architecture and loss functions, but it does not present a clearly labeled 'Algorithm' or 'Pseudocode' block with structured, numbered steps. |
| Open Source Code | No | The paper does not contain any explicit statement about releasing code for the described methodology, nor does it provide a direct link to a code repository. It mentions 'Further details and empirical studies can be found in Appendix E.', but this appendix details model specifics and experimental settings rather than code access. |
| Open Datasets | Yes | 6.1. Datasets and Baselines Datasets. We validate our method on two wellknown benchmarks for domain adaptation: Office Home (Venkateswara et al., 2017) and PACS (Li et al., 2017). |
| Dataset Splits | No | In each dataset, a single domain is designated as the target, and the remaining domains serve as sources. For Office Home, we extract features using a pretrained Res Net50, then apply MLP-based VAEs alongside a classifier. Meanwhile, for PACS, we employ Res Net18 as the backbone and similarly integrate MLP-based VAEs and a classifier. All metrics are computed by averaging over three random seeds. |
| Hardware Specification | Yes | Computing resources and efficiency. We train our model using a NVIDIA A100-SXM4-40GB GPU. For the Office-Home dataset, the batch size is set to 32, and the model is trained for 70 epochs, which takes approximately 160 minutes. The peak memory usage is around 35 GB. The majority of the computational cost comes from the Res Net-50 backbone, as we only add several lightweight MLP layers after it. For the PACS dataset, the batch size is set to 32, and the model is trained for 70 epochs, each epoch has 200 steps, which takes approximately 32 minutes. The peak memory usage is around 11 GB. |
| Software Dependencies | No | The model details describe the use of 'Res Net-50', 'MLP-based VAEs', 'VAE framework', 'Gumbel-Softmax VAE', and 'Res Net18' as architectural components and frameworks. However, the paper does not specify version numbers for any software libraries, programming languages (e.g., Python, PyTorch, TensorFlow), or operating systems used in the implementation. |
| Experiment Setup | Yes | Computing resources and efficiency. ... For the Office-Home dataset, the batch size is set to 32, and the model is trained for 70 epochs... For the PACS dataset, the batch size is set to 32, and the model is trained for 70 epochs, each epoch has 200 steps... Table 5: Hyperparameters for Office-Home (Ar, Cl, Pr, Rw) and PACS (P, A, C, S) datasets. Parameter Office-Home PACS Ar Cl Pr Rw P A C S λ1 2 10 3 6 10 4 3 10 3 6 10 4 7 10 4 3 10 3 5 10 3 9 10 3 λ2 4 10 4 2 10 4 3 10 3 1 10 4 4 10 4 1 10 4 1 10 4 1 10 3 λ3 1 10 4 8 10 4 1 10 3 4 10 3 5 10 4 2 10 3 2 10 4 7 10 4 λ4 5 10 3 6 10 4 7 10 3 1 10 3 1 10 3 2 10 4 4 10 4 5 10 3 λ5 2 10 3 4 10 4 3 10 4 4 10 3 4 10 3 9 10 4 4 10 3 5 10 3 |