Uniformly Distributed Feature Representations for Fair and Robust Learning
Authors: Kiran Krishnamachari, See-Kiong Ng, Chuan-Sheng Foo
TMLR 2024 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Our experiments on multiple datasets for sub-population shifts and domain generalization show that URM improves the generalization of deep neural networks without requiring knowledge of groups or domains during training. URM is competitive with the best existing methods designed for these tasks and can also be easily combined with them for improved performance. |
| Researcher Affiliation | Academia | Kiran Krishnamachari EMAIL Institute for Infocomm Research (I 2R), A*STAR, Singapore School of Computing, National University of Singapore, Singapore See-Kiong Ng EMAIL Institute of Data Science, National University of Singapore, Singapore School of Computing, National University of Singapore, Singapore Chuan-Sheng Foo EMAIL Institute for Infocomm Research (I 2R), A*STAR, Singapore Centre for Frontier AI Research (CFAR), A*STAR, Singapore |
| Pseudocode | No | The paper describes methods and objectives mathematically, but does not present any structured pseudocode or algorithm blocks. |
| Open Source Code | Yes | Code is available at https://github.com/kiranchari/Uniform Risk Minimization. |
| Open Datasets | Yes | For group robustness, we evaluate on Waterbirds (Wah et al., 2011), Celeb A (Liu et al., 2015), Civil Comments (Borkan et al., 2019) and Multi NLI (Williams et al., 2018). For domain generalization (DG), we benchmark on Colored MNIST (Arjovsky et al., 2020), Rotated MNIST (Ghifary et al., 2015), PACS (Li et al., 2017), VLCS (Fang et al., 2013), Office Home (Venkateswara et al., 2017) and Terra Incognita (Beery et al., 2018). |
| Dataset Splits | Yes | For domain generalization experiments, our experimental setup followed Gulrajani & Lopez-Paz (2021) using the leave-one-domain-out cross-validation scheme. |
| Hardware Specification | Yes | All experiments were run on a computer with 8 A100 GPUs. |
| Software Dependencies | No | The paper mentions software components like ResNet50, BERT, UMAP, PyTorch (implicitly through references), and optimizers (SGD, AdamW), but does not provide specific version numbers for any of them. |
| Experiment Setup | Yes | For both tasks, we searched over sixteen random hyper-parameter combinations and applied adversarial distribution matching to the penultimate layer before the linear classification head. Hyper-parameter search ranges for URM are included in the Appendix (Table 4). Table 4: Search ranges for URM-specific hyper-parameters. Other training parameters ranges are based on (Yang et al., 2023). Method Parameter Random Distribution lambda λ Uniform(0, 0.2) generator output Random Choice([Tan H, Re LU]) number of layers in discriminator Random Choice([1, 2, 3]) learning rate (for image or tabular datasets) 10Uniform( 5, 3) optimizer (for image or tabular datasets) Random Choice([SGD]) learning rate (text datasets) 10Uniform( 6, 5) optimizer (text datasets) Random Choice([Adam W]) |