AGRO: Adversarial discovery of error-prone Groups for Robust Optimization

Authors: Bhargavi Paranjape, Pradeep Dasigi, Vivek Srikumar, Luke Zettlemoyer, Hannaneh Hajishirzi

ICLR 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental On the WILDS benchmark, AGRO results in 8% higher model performance on average on known worst-groups, compared to prior group discovery approaches used with G-DRO.
Researcher Affiliation Collaboration Bhargavi Paranjape1 Pradeep Dasigi2 Vivek Srikumar2,3 Luke Zettlemoyer1,4 Hannaneh Hajishirzi1,2 1University of Washington, 2Allen Institute of Artificial Intelligence, 3University of Utah 4Meta AI
Pseudocode Yes Algorithm 1 presents the pseudo-code for AGRO.
Open Source Code Yes Our code and models are public3. 3https://github.com/bhargaviparanjape/robust-transformers
Open Datasets Yes We evaluate performance on four datasets in language and vision from the popular WILDS benchmark (Koh et al., 2021) to study in-the-wild distribution shifts.
Dataset Splits Yes We follow the train/dev/test split in (Sagawa et al., 2019), which results in 206,175 training examples.
Hardware Specification No The paper mentions training models on 'GPU' but does not provide specific details such as GPU models, CPU models, or cloud computing instance specifications.
Software Dependencies No The paper mentions transformer-based models like DeBERTa-Base, BEIT, RoBERTa-base, and ViT, but does not provide specific software versions for libraries, programming languages, or operating systems used in the experiments.
Experiment Setup Yes Hyperparameters For θ, we use transformer-based (Vaswani et al., 2017) pretrained models... Appendix A.4 and A.6.3 further detail hyperparameter tuning, particularly for AGRO-specific hyperparameters α and m.