Identifying Statistical Bias in Dataset Replication

Authors: Logan Engstrom, Andrew Ilyas, Shibani Santurkar, Dimitris Tsipras, Jacob Steinhardt, Aleksander Madry

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

Reproducibility Variable Result LLM Response
Research Type Experimental We study Image Net-v2, a replication of the Image Net dataset on which models exhibit a significant (11-14%) drop in accuracy, even after controlling for selection frequency, a human-in-the-loop measure of data quality. We show that after remeasuring selection frequencies and correcting for statistical bias, only an estimated 3.6% 1.5% of the original 11.7% 1.0% accuracy drop remains unaccounted for.
Researcher Affiliation Academia 1MIT 2UC Berkeley. Correspondence to: Logan Engstrom <engstrom@mit.edu>.
Pseudocode Yes We provide further detail (including pseudocode) on the fitting process for pi(s(x); θ) in Appendix F.
Open Source Code Yes Code for our study is publicly available1. 1https://git.io/data-rep-analysis
Open Datasets Yes Image Net (Deng et al., 2009; Russakovsky et al., 2015) (which we also refer to as Image Net-v1 or v1) is one of the most widely used datasets in computer vision.
Dataset Splits No The paper refers to pre-existing datasets like ImageNet and ImageNet-v2, and their respective 'test sets', but does not explicitly provide details about training, validation, and test splits (e.g., percentages or sample counts for each split).
Hardware Specification No The paper does not provide any specific details about the hardware (e.g., GPU/CPU models, memory) used for running the experiments.
Software Dependencies No The paper refers to using 'Amazon Mechanical Turk (MTurk)' and mentions 'PyTorch' in a reference's title, but does not provide specific version numbers for any software dependencies used in their experiments.
Experiment Setup Yes In these tasks, MTurk annotators were shown grids of 48 images at a time, each corresponding to an Image Net class. ... Each image was seen by 40 distinct annotators... We opt to use mixtures of beta distributions as the family pi( ; θ) ... a cubic spline