Spread Spurious Attribute: Improving Worst-group Accuracy with Spurious Attribute Estimation

Authors: Junhyun Nam, Jaehyung Kim, Jaeho Lee, Jinwoo Shin

ICLR 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Our experiments on various benchmark datasets show that our algorithm consistently outperforms the baseline methods using the same number of group-labeled samples.
Researcher Affiliation Academia Junhyun Nam1, Jaehyung Kim1, Jaeho Lee2 , Jinwoo Shin1 1KAIST, 2POSTECH {junhyun.nam,jaehyungkim,jinwoos}@kaist.ac.kr jaeho.lee@postech.ac.kr
Pseudocode Yes Algorithm 1 Spread Spurious Attribute
Open Source Code Yes Also, we provide our source code as a part of the open-to-public supplementary materials.
Open Datasets Yes Waterbirds (Sagawa et al., 2020)... Caltech-UCSD Birds dataset (Wah et al., 2011) with landscapes from Places (Zhou et al., 2017)., Celeb A (Liu ets al., 2015), Multi NLI (Williams et al., 2018), Civil Comments-WILDS (Borkan et al., 2019; Koh et al., 2021), CIFAR-10 (Krizhevsky et al., 2009).
Dataset Splits Yes For all datasets, we use the validation split of the dataset as the group-labeled set. and We use D L, D U to train the spurious attribute predictor that make prediction on D U, and validate the model with D L.
Hardware Specification Yes In Table 10, we provide the time required for the pseudo-labeling phase and the robust training phase on a single Nvidia Titan XP for each dataset.
Software Dependencies No The paper mentions software like torchvision and huggingface implementations, and optimizers like SGD and Adam W, but does not provide specific version numbers for these software components.
Experiment Setup Yes For Waterbirds and Celeb A, we tuned the learning rate over {1e3, 1e-4, 1e-5} and ℓ2 regularization over {1e-1, 1e-4}. We used SGD optimizer with momentum 0.9 and batch size 64. In pseudo-labeling phase, we train the spurious attribute predictor 1k iterations for Waterbirds and 45k iterations for Celeb A.