Latent Covariate Shift: Unlocking Partial Identifiability for Multi-Source Domain Adaptation
Authors: Yuhang Liu, Zhen Zhang, Dong Gong, Mingming Gong, Biwei Huang, Anton van den Hengel, Kun Zhang, Javen Qinfeng Shi
TMLR 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | The proposed approach showcases exceptional performance and efficacy on both simulated and real-world datasets. Section 6 is titled 'Experiments' and includes subsections like 'Experiments on Synthetic Data', 'Experiments on Real Resampled PACS data', 'Experiments on Real Terra Incognita data', and 'Ablation studies'. |
| Researcher Affiliation | Academia | Yuhang Liu EMAIL Australian Institute for Machine Learning The University of Adelaide; Zhen Zhang EMAIL Australian Institute for Machine Learning The University of Adelaide; Dong Gong EMAIL School of Computer Science and Engineering The University of New South Wales; Mingming Gong EMAIL School of Mathematics and Statistics The University of Melbourne; Biwei Huang EMAIL Halicioğlu Data Science Institute University of California San Diego; Anton van den Hengel EMAIL Australian Institute for Machine Learning The University of Adelaide; Kun Zhang EMAIL Department of Philosophy Carnegie Mellon University; Javen Qinfeng Shi EMAIL Australian Institute for Machine Learning The University of Adelaide |
| Pseudocode | No | The paper describes the methods in prose and mathematical equations. There are no figures or sections explicitly labeled 'Pseudocode' or 'Algorithm'. |
| Open Source Code | Yes | The code is available at: https://sites.google.com/view/yuhangliu/projects |
| Open Datasets | Yes | There exist datasets, such as PACS (Li et al., 2017) and Office-Home (Venkateswara et al., 2017), commonly used for evaluating MSDA under previous paradigms. ... We further evaluate the proposed i LCC-LCS on Terra Incognita dataset proposed in (Beery et al., 2018) used for evaluation for domain generalization. |
| Dataset Splits | Yes | For synthetic data, the paper states: 'We use the first 4 segments as source domains, and the last segment as the target domain.' For real-world data, it mentions using 'four domains from the original data, L28, L43, L46, and L7' for Terra Incognita, and resampled PACS datasets based on DKL. The experiment tables (e.g., Tables 1-4) show results for individual domains as targets, implying a multi-source domain adaptation evaluation setup where some domains are sources and others are targets. |
| Hardware Specification | No | The paper mentions using a ResNet-18 backbone and training VAE on high-resolution images, but no specific hardware (GPU models, CPU types, memory) used for running the experiments is detailed. |
| Software Dependencies | No | The paper mentions using a ResNet-18 backbone and concepts like Variational Autoencoder, but does not provide specific version numbers for any software dependencies (e.g., Python, PyTorch, CUDA, specific libraries). |
| Experiment Setup | Yes | For synthetic data, the paper specifies network architectures: 'encoder, e.g. 3-layer fully connected network with 30 hidden nodes for each layer, and a decoder, e.g. 3-layer fully connected network with 30 hidden nodes for each layer. We use a 3-layer fully connected network with 30 hidden nodes for the prior model.' It also provides hyperparameters: 'we set β = 1 and γ = 0 to remove the heuristic constraints, and we set λ = 1e 2.' For real datasets, it states: 'All methods used the same network backbone, Res Net-18 pre-trained on Image Net. ... We then use 2-layer fully connected networks as the VAE encoder and decoder, use 2-layer fully connected network for the prior model, and use 2-layer fully connected network to transfer nc to zc. For hyper-parameters, we set β = 4, γ = 0.1, λ = 1e 4 for the proposed method on all datasets.' |