ImageNet-X: Understanding Model Mistakes with Factor of Variation Annotations

Authors: Badr Youbi Idrissi, Diane Bouchacourt, Randall Balestriero, Ivan Evtimov, Caner Hazirbas, Nicolas Ballas, Pascal Vincent, Michal Drozdzal, David Lopez-Paz, Mark Ibrahim

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

Reproducibility Variable Result LLM Response
Research Type Experimental Equipped with Image Net-X, we investigate 2,200 current recognition models and study the types of mistakes as a function of model s (1) architecture e.g. transformer vs. convolutional , (2) learning paradigm e.g. supervised vs. self-supervised , and (3) training procedures e.g. data augmentation.
Researcher Affiliation Industry Fundamental AI Research (FAIR), Meta AI {byoubi,marksibrahim}@meta.com
Pseudocode No No pseudocode or algorithm blocks were found in the paper.
Open Source Code Yes We release all the Image Net-X annotations along with an open-source toolkit to probe existing or new models failure types. The data and code are available at https://facebookresearch.github.io/imagenetx/site/home.
Open Datasets Yes To address this need, we introduce Image Net-X a set of sixteen human annotations of factors such as pose, background, or lighting for the entire Image Net-1k validation set as well as a random subset of 12k training images.
Dataset Splits Yes Image Net-X contains human annotations for each of the 50,000 images in the validation set of the Image Net dataset and 12,000 random sample from the training set.
Hardware Specification No No specific hardware details (e.g., GPU/CPU models, memory amounts, or detailed computer specifications) used for running experiments were provided.
Software Dependencies No The paper mentions data preprocessing using 'Pandas and Numpy, both freely available Python packages', but does not provide specific version numbers for these or other software dependencies.
Experiment Setup Yes Each run across all policies share the exact same optimizer (SGD), weight-decay (1e-5), mini-batch size (512), number of epochs (80), and data ordering through training.