Learning from Failure: De-biasing Classifier from Biased Classifier

Authors: Junhyun Nam, Hyuntak Cha, Sungsoo Ahn, Jaeho Lee, Jinwoo Shin

NeurIPS 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive experiments demonstrate that our method significantly improves the training of network against various types of biases in both synthetic and real-world datasets. Surprisingly, our framework even occasionally outperforms the debiasing methods requiring explicit supervision of the spuriously correlated attributes.
Researcher Affiliation Academia 1School of Electrical Engineering, KAIST 2Graduate School of AI, KAIST
Pseudocode Yes Algorithm 1 Learning from Failure
Open Source Code No The paper mentions 'BAR is publicly available1' with a footnote linking to 'https://github.com/alinlab/BAR', which is specified as a dataset. There is no explicit statement or link provided for the open-source code of the Lf F methodology itself.
Open Datasets Yes We show the effectiveness of Lf F on various biased datasets, including Colored MNIST [14, 18] with color bias, Corrupted CIFAR-10 [12] with texture bias, and Celeb A [19] with gender bias. In addition, we newly construct a real-world dataset, coined biased action recognition (BAR), to resolve the lack of realistic evaluation benchmark for debiasing schemes. BAR is publicly available1, and a detailed description of BAR is in Appendix B.
Dataset Splits No The paper describes the construction of 'unbiased evaluation set' and 'bias-conflicting evaluation set' and varies the ratio of bias-aligned samples in the 'training dataset', but it does not provide specific training/validation/test splits (e.g., percentages or counts) for model training and hyperparameter tuning processes.
Hardware Specification No The paper does not provide specific details about the hardware (e.g., GPU models, CPU types, or memory) used for running the experiments.
Software Dependencies No The paper does not specify any software dependencies with version numbers (e.g., specific versions of deep learning frameworks or libraries).
Experiment Setup No The paper mentions hyperparameters like 'q' for GCE loss and 'learning rate η, number of iterations T' in Algorithm 1. It also states 'We provide a detailed description of datasets we considered in Appendix A, B and experimental details in Appendix C.' in Section 4. However, specific numerical values for these hyperparameters and detailed training configurations are not provided in the main text of the paper.