SelecMix: Debiased Learning by Contradicting-pair Sampling
Authors: Inwoo Hwang, Sangjun Lee, Yunhyeok Kwak, Seong Joon Oh, Damien Teney, Jin-Hwa Kim, Byoung-Tak Zhang
NeurIPS 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experiments on standard benchmarks demonstrate the effectiveness of the method, in particular when label noise complicates the identification of bias-conflicting examples. We evaluate our method on standard benchmarks for debiasing. Experimental results suggest that Selec Mix consistently outperforms prior methods, especially when bias-conflicting samples are scarce. |
| Researcher Affiliation | Collaboration | Inwoo Hwang1 Sangjun Lee1 Yunhyeok Kwak1 Seong Joon Oh3 Damien Teney4 Jin-Hwa Kim 12 Byoung-Tak Zhang 1 1AI Institute, Seoul National University 2NAVER AI Lab 3University of Tübingen 4Idiap Research Institute |
| Pseudocode | Yes | The pseudo-code of the proposed algorithm is presented in Alg. 1 and Alg. 2. |
| Open Source Code | No | Did you include the code, data, and instructions needed to reproduce the main experimental results (either in the supplemental material or as a URL)? [No] |
| Open Datasets | Yes | Datasets. The Colored MNIST is a modified MNIST [19]... The Corrupted CIFAR10 is constructed by applying different types of corruptions to the corresponding objects in the original CIFAR-10 [18] dataset... The Biased FFHQ (BFFHQ) [20] is constructed based on the real-world dataset FFHQ [12]... All datasets are available in the official repository of DFA [20]. |
| Dataset Splits | No | The paper mentions that it evaluates unbiased accuracy on the test set of the dataset and specifies the ratio of bias-conflicting samples in the training set. However, it does not explicitly describe a validation dataset split (percentages, counts, or methodology) for hyperparameter tuning or early stopping. |
| Hardware Specification | Yes | All experiments are conducted on NVIDIA A100 GPUs. |
| Software Dependencies | Yes | All experiments are implemented using PyTorch (1.10.0). |
| Experiment Setup | Yes | We train models for 200 epochs (Colored MNIST and Corrupted CIFAR-10) and 100 epochs (BFFHQ) with SGD optimizer (momentum=0.9, weight decay=1e-4). The initial learning rate is 0.1, and it is annealed by 0.1 at 100 and 150 epochs (Colored MNIST and Corrupted CIFAR-10), and 50 and 80 epochs (BFFHQ). The batch size is 128. The temperature hyperparameter of the contrastive loss is 0.07. |