On Harmonizing Implicit Subpopulations
Authors: Feng Hong, Jiangchao Yao, Yueming Lyu, Zhihan Zhou, Ivor Tsang, Ya Zhang, Yanfeng Wang
ICLR 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments on various benchmark datasets show the effectiveness of SHE. The code is available. |
| Researcher Affiliation | Academia | 1Cooperative Medianet Innovation Center, Shanghai Jiao Tong University 2Shanghai Artificial Intelligence Laboratory 3CFAR, Agency for Science, Technology and Research 4IHPC, Agency for Science, Technology and Research 5Nanyang Technological University {feng.hong, Sunarker, zhihanzhou, ya zhang, wangyanfeng}@sjtu.edu.cn {Lyu Yueming, ivor tsang}@cfar.a-star.edu.sg |
| Pseudocode | No | The paper does not contain any clearly labeled pseudocode or algorithm blocks. |
| Open Source Code | Yes | The code is available. |
| Open Datasets | Yes | We evaluate our SHE on COCO (Lin et al., 2014), CIFAR-100 (Krizhevsky et al., 2009), and tiered Image Net (Ren et al., 2018). |
| Dataset Splits | Yes | For COCO, we follow the ALT-protocol (Tang et al., 2022) to conduct subpopulation-imbalanced training set and balanced test set. For CIFAR-100, we take the 20 superclasses as classification targets and generate subpopulation imbalances by sampling in the subclasses of each superclass. Following Cui et al. (2019), we use the exponential sampling with imbalance ratio IR {20, 50, 100}, where IR = maxs S (xi,yi,si) D 1(si=s) mins S (xi,yi,si) D 1(si=s) . For tiered Image Net, we take the 34 superclasses as classification targets and generate subpopulation imbalances by imbalanced sampling in 10 subclasses of each superclass with the imbalance ratio IR = 100. ... When the group information is unknown for all methods, we utilize the overall accuracy on a validation set that shares the same distribution as the training set as the selection metric. |
| Hardware Specification | Yes | All the experiments are conducted on NVIDIA Ge Force RTX 3090s with Python 3.7.10 and Pytorch 1.13.1. |
| Software Dependencies | Yes | All the experiments are conducted on NVIDIA Ge Force RTX 3090s with Python 3.7.10 and Pytorch 1.13.1. |
| Experiment Setup | Yes | We use 18-layer Res Net as the backbone. The standard data augmentations are applied as in Cubuk et al. (2020). The mini-batch size is set to 256 and all the methods are trained using SGD with momentum of 0.9 and weight decay of 0.005 as the optimizer. The pre-defined K is set to 4 if not specifically stated and the hyper-parameter β in Eq. (2) is set to 1.0. The initial learning rate is set to 0.1. We train the model for 200 epochs with the cosine learning-rate scheduling. |