Domain Generalization via Heckman-type Selection Models

Authors: Hyungu Kahng, Hyungrok Do, Judy Zhong

ICLR 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We also demonstrate its efficacy empirically through simulations and experiments on a set of benchmark datasets comparing with other well-known DG methods.
Researcher Affiliation Academia 1Korea University 2NYU School of Medicine hgkahng@korea.ac.kr, {hyungrok.do,judy.zhong}@nyulangone.org
Pseudocode Yes Algorithm 1 Two-Step Optimization for Heckman DG
Open Source Code Yes code available: https://github.com/hgkahng/domain-generalization-lightning
Open Datasets Yes To further demonstrate the effectiveness of Heckman DG on high-dimensional data regimes, we conducted experiments on four datasets from the WILDS benchmark (Koh et al., 2021): 1) CAMELYON17, 2) POVERTYMAP, 3) IWILDCAM, and 4) RXRX1.
Dataset Splits Yes Detailed descriptions of dataset statistics are presented in Table 5 of Appendix A.4. In the Domain row, the three numbers in parentheses denote the number of train, validation, and test domains. (e.g., CAMELYON17: 5 Hospitals (3, 1, 1))
Hardware Specification No The paper does not provide specific details on the hardware used for experiments, such as GPU models, CPU models, or detailed specifications of computing resources.
Software Dependencies No The paper mentions software components like 'Dense Net-121', 'Res Net-18-MS', 'Res Net-50', 'Adam', and 'SGD' but does not specify version numbers for these or other key software dependencies (e.g., Python, PyTorch, CUDA).
Experiment Setup Yes Details on training configurations of Heckman DG are provided in Table 6. This includes parameters such as 'Epochs', 'Batch Size', 'Learning Rate', 'Weight Decay', 'Image Net Weights', and 'Data Augmentation'.