Bibtex
@article{valuesInML2021,
title={The Values Encoded in Machine Learning Research},
author={Birhane, Abeba and Kalluri, Pratyusha and Card, Dallas and Agnew, William and Dotan, Ravit and Bao, Michelle},
journal={arXiv preprint arXiv:},
year={2021}
}
Machine learning (ML) currently exerts an outsized influence on the world, increas-ingly affecting communities and institutional practices. It is therefore critical thatwe question vague conceptions of the field as value-neutral or universally beneficial,and investigate what specific values the field is advancing. In this paper, we presenta rigorous examination of the values of the field by quantitatively and qualitativelyanalysing 100 highly cited ML papers published at premier ML conferences, ICMLand NeurIPS. We annotate key features of papers which reveal their values: howthey justify their choice of project, which aspects they uplift, their considerationof potential negative consequences, and their institutional affiliations and fundingsources. We find that societal needs are typically very loosely connected to thechoice of project, if mentioned at all, and that consideration of negative conse-quences is extremely rare. We identify 67 values that are uplifted in these papers,and, of these, we find that papers most frequently justify and assess themselvesbased on performance, generalization, efficiency, researcher understanding, novelty,and building on previous work. We present extensive textual evidence and analysisof how these values are operationalized. Notably, we find that each of these topvalues is currently being defined and applied with assumptions and implicationsgenerally supporting the centralization of power. Finally, we find increasingly closeties between these highly cited papers and tech companies and elite universities.
- annotations_anon.xlsx: raw annotations
- code-for-appendix.zip: replication code for experiments in Appendix