Robust Crowdsourcing: A Proactive Approach to Ensuring Fairness in Machine Learning
Abstract
The rise of Machine Learning (ML) has been made possible with the advent of crowdsourcing platforms (e.g., Amazon Mechanical Turk, Figure Eight), which are increasingly being used by ML researchers for creating large-scale labeled datasets for their applications. This use of crowdsourcing for generating ML training data has served researchers/practitioners (and society) reasonably well so far, as evidenced by the spectacular successes of ML systems in computer vision, natural language processing, and other such domains. However, as ML enters into highly sensitive domains (e.g., fake news detection, military intelligence, etc.), different subjective biases held by non-expert human crowd-workers adversely impact the quality/accuracy of their annotations/labels, and these biases are amplified in outputs of ML algorithms trained on such data. This problem affects every sensitive real-world domain where the correct label (or answer) is subjective (as opposed to the objectivity in labeling images as cats or dogs). Prior work on handling systemic biases in datasets has resulted in a rich literature on bias-aware ML algorithms in the field of Fairness, Accountability and Transparency in ML (or FAT-ML), but most of this prior work represents a reactive response to this problem, i.e., a core assumption in prior work is that the provided dataset suffers from bias, and methods are developed to reactively address this bias (e.g., by post-processing the outputs of standard ML algorithms). The goal of the proposed research is to develop proactive approaches for handling systemic biases in ML datasets by tackling the following question: Can we build better crowdsourcing systems which are robust to subjective biases of human crowd-workers such that datasets derived from such systems are bias-free (thereby leading to unbiased ML models)? Our proposed robust crowdsourcing systems (RCS) will ensure that ML algorithms trained in the future do not have to worry about bias in their training. More specifically, we propose to achieve this goal by designing optimal ways of incentivizing (or dissuading) human crowd-workers so that it is in their best interests (from a utilitarian perspective) to not let their personal biases affect their annotation process, thereby resulting in bias-free data. While the proposed research can be categorized as FAT-ML research, it represents a fundamental and novel departure from the prevalent thinking in the field (that of reactive bias mitigation), and opens up a new paradigm of research for the FAT-ML community. We propose to answer these questions by developing algorithmic solutions for fundamental problems in different crowdsourcing scenarios. First, we propose to conduct research in developing game-theory based RCS by modeling the interaction between job requesters (leaders) and crowd-workers (followers) on crowdsourcing platforms as a Stackelberg game. We intend to provide novel game formulations for these problems, define appropriate notions of equilibria in these games, and provide efficient algorithms for computation of such equilibria. Second, we propose to develop RCS based on simultaneous move games. Finally, we propose evaluating the effectiveness of our RCS via theoretical characterization, simulation experiments, and human subject experiments on real-world crowdsourcing platforms. The insights obtained from our evaluation will be used to lead to the development of even more practical and effective RCS.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Jan 04, 2021
- Source ID
- W911NF2110047
Entities
People
- Amulya Yadav
Organizations
- Army Contracting Command
- Pennsylvania State University
- United States Army