Mitigating Dataset Harms Requires Stewardship: Lessons from 1000 Papers

Abstract

Machine learning datasets have elicited concerns about privacy, bias, and unethical applications, leading to the retraction of prominent datasets such as DukeMTMC, MS-Celeb-1M, and Tiny Images. In response, the machine learning community has called for higher ethical standards in dataset creation. To help inform these efforts, we studied three influential but ethically problematic face and person recognition datasets -- Labeled Faces in the Wild (LFW), MS-Celeb-1M, and DukeMTM -- by analyzing nearly 1000 papers that cite them. We found that the creation of derivative datasets and models, broader technological and social change, the lack of clarity of licenses, and dataset management practices can introduce a wide range of ethical concerns. We conclude by suggesting a distributed approach to harm mitigation that considers the entire life cycle of a dataset.

References in corpus (14)

• Semantics derived automatically from language corpora contain human-like biases
• Deep Facial Expression Recognition: A Survey
• Towards a Critical Race Methodology in Algorithmic Fairness
• AI Fairness 360: An Extensible Toolkit for Detecting, Understanding, and Mitigating Unwanted Algorithmic Bias
• Image Representations Learned With Unsupervised Pre-Training Contain Human-like Biases
• Racial categories in machine learning
• A Study of Face Obfuscation in ImageNet
• Bringing the People Back In: Contesting Benchmark Machine Learning Datasets
• Large image datasets: A pyrrhic win for computer vision?
• Algorithmic Injustices: Towards a Relational Ethics
• About Face: A Survey of Facial Recognition Evaluation
• Towards Standardization of Data Licenses: The Montreal Data License
• An Ethical Highlighter for People-Centric Dataset Creation
• Like a Researcher Stating Broader Impact For the Very First Time