Learning Debiased Classifier with Biased Committee
Authors: Nayeong Kim, SEHYUN HWANG, Sungsoo Ahn, Jaesik Park, Suha Kwak
NeurIPS 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | On five real-world datasets, our method outperforms prior arts using no spurious attribute label like ours and even surpasses those relying on bias labels occasionally. Our code is available at https://github.com/nayeong-v-kim/LWBC. |
| Researcher Affiliation | Academia | Pohang University of Science and Technology (POSTECH), South Korea {nayeong.kim, sehyun03, sungsoo.ahn, jaesik.park, suha.kwak}@postech.ac.kr |
| Pseudocode | Yes | Algorithm 1 Learning a debiased classifier with a biased committee |
| Open Source Code | Yes | Our code is available at https://github.com/nayeong-v-kim/LWBC. |
| Open Datasets | Yes | Celeb A. Celeb A [31] is a dataset for face recognition where each sample is labeled with 40 attributes. Image Net-9. Image Net-9 [20] is a subset of Image Net [35] containing nine super-classes. Image Net-A. Image Net-A [17] contains real-world images misclassified by an Image Net-trained Res Net 50 [15]. BAR. The Biased Action Recognition (BAR) dataset [32] is a real-world image dataset intentionally designed to exhibit spurious correlations between human action and place on its images. NICO. NICO [16] is a real-world dataset for simulating out-of-distribution image classification scenarios. |
| Dataset Splits | Yes | Following the setting adopted by Bahng et al. [3], we conduct experiments with 54,600 training images and 2,100 validation images. In our setting, we use 10% of the original BAR training set as validation and set the bias-conflicting ratio of the training set to 1%. The validation and test sets consist of 7 seen context classes and 3 unseen context classes per object class. |
| Hardware Specification | No | The paper does not explicitly describe the specific hardware used (e.g., GPU model, CPU type) for running its experiments. |
| Software Dependencies | No | The paper does not provide specific version numbers for software components or libraries (e.g., PyTorch version, CUDA version) needed for replication. |
| Experiment Setup | Yes | We set the batch size to {64, 64, 128, 256}, learning rate to {1e-3, 1e-3, 1e-4, 6e-3}, the size of the committee m to {30, 30, 30, 40}, the size of subset Sl to {10, 10, 80, 300}, λ to {0.9, 0.6, 0.6, 0.6}, and τ to {1, 1, 1, 2.5}, respectively for {BAR, NICO, Imagenet-9, Celeb A}, and α to 0.02. Note that we run LWBC on 3 random seeds and report the average and the standard deviation. |