Unknown Domain Inconsistency Minimization for Domain Generalization
Authors: Seungjae Shin, HeeSun Bae, Byeonghu Na, Yoon-Yeong Kim, Il-chul Moon
ICLR 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | In an empirical aspect, UDIM consistently outperforms SAM variants across multiple DG benchmark datasets. Notably, UDIM shows statistically significant improvements in scenarios with more restrictive domain information, underscoring UDIM s generalization capability in unseen domains. |
| Researcher Affiliation | Collaboration | Seungjae Shin 1, Hee Sun Bae 1, Byeonghu Na1, Yoon-Yeong Kim2 & Il-Chul Moon1,3 1Department of Industrial and Systems Engineering, KAIST 2Department of Statistics, University of Seoul, 3summary.ai |
| Pseudocode | Yes | Algorithm of UDIM is in Appendix C. ... Algorithm 1: Training algorithm of UDIM w/ SAM |
| Open Source Code | Yes | Our code is available at https://github.com/SJShin-AI/UDIM. |
| Open Datasets | Yes | First, we conducted evaluation on CIFAR10-C (Hendrycks & Dietterich, 2019), a synthetic dataset that emulates various domains by applying several synthetic corruptions to CIFAR-10 (Krizhevsky et al., 2009). Furthermore, we extend our evaluation to real-world datasets with multiple domains, namely PACS (Li et al., 2017), Office Home (Venkateswara et al., 2017), and Domain Net (Peng et al., 2019). |
| Dataset Splits | Yes | We get the test performance whose accuracy for source validation dataset is best. |
| Hardware Specification | No | The paper mentions using ResNet-18 and ResNet-50 models and the Adam optimizer but does not specify the hardware (e.g., GPU models, CPU types) used for the experiments. |
| Software Dependencies | No | We utilize Back PACK (Dangel et al., 2020), which provides the faster computation of per-sample gradients. ... and use Adam (Kingma & Ba, 2014) optimizer basically. While Back PACK and Adam are mentioned, specific version numbers for these or other key software components (like Python, PyTorch) are not provided. |
| Experiment Setup | Yes | learning rate is set as 3 10 5 following Wang et al. (2023). ... we use batch size as 32 for PACS, Office Home, and Domain Net and 64 for CIFAR-10-C. ... we trained for total of 5,000 iterations. For Domain Net, we trained for 15,000 iterations. For CIFAR-10, since it usually trains for 100 epochs, we translate it to iterations, which becomes total of 781 * 100 = 78,100 iterations. |