Missingness Bias in Model Debugging

Authors: Saachi Jain, Hadi Salman, Eric Wong, Pengchuan Zhang, Vibhav Vineet, Sai Vemprala, Aleksander Madry

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

Reproducibility Variable Result LLM Response
Research Type Experimental To systematically measure the impacts of missingness bias, we iteratively remove subregions from the input and analyze the types of mistakes that our models make. See Appendix A for experimental details. We perform an extensive study across various: architectures (Appendix C.3), missingness approximations (Appendix C.4), subregion sizes (Appendix C.5), subregion shapes: patches vs superpixels (Appendix C.6), and datasets (Appendix E).
Researcher Affiliation Collaboration 1Massachusetts Institute of Technology 2Microsoft Research
Pseudocode No The paper describes experimental procedures and methods in paragraph text, but does not include any explicitly labeled pseudocode or algorithm blocks.
Open Source Code Yes Our code is available at https://github.com/madrylab/missingness.
Open Datasets Yes We train our models on Image Net (Russakovsky et al., 2015), with a custom (research, non-commercial) license, as found here https://paperswithcode.com/dataset/ imagenet.
Dataset Splits Yes For all experiments in this paper, we consider 10,000 image subset of the original Image Net validation set (we take every 5th image).
Hardware Specification Yes For Image Net, we train our models on 4 V100 GPUs each, and training took around 12 hours for Res Net-18 and Vi T-T, and around 20 hours for Res Net50 and Vi T-S.
Software Dependencies No The paper does not explicitly list specific software dependencies with version numbers (e.g., Python, PyTorch, CUDA versions).
Experiment Setup Yes For Res Nets, we train using SGD with batch size of 512, momentum of 0.9, and weight decay of 1e-4. We train for 90 epochs with an initial learning rate of 0.1 that drops by a factor of 10 every 30 epochs.