TL;DR
This paper introduces a graduate-level course on AI Ethics (FACT-AI) that uniquely leverages reproducibility as its core teaching mechanism. By having students re-implement existing FACT-AI research papers, the course aims to foster a deep understanding of ethical AI principles and the practical challenges of ensuring reliable, transparent research.
What Problem Does This Paper Address?
As AI systems become increasingly influential, there's a growing demand for researchers and developers to consider the ethical implications of their work, particularly concerning fairness, accountability, confidentiality, and transparency (FACT). However, teaching these concepts effectively in a technical AI program presents a challenge. Traditional approaches might focus on theory, but a hands-on method is needed to truly grasp the nuances. Furthermore, the AI research community itself faces a significant "reproducibility crisis," where many published results are difficult or impossible to verify, undermining trust and hindering progress. This paper addresses how to teach FACT-AI principles while simultaneously tackling the critical issue of research reproducibility.
Methods
The core of the FACT-AI course is a group project where students are tasked with reproducing existing FACT-AI algorithms from top AI conferences. This involves:
- Re-implementation: Students select a FACT-AI paper and re-implement its algorithms from scratch. If code is available, they are encouraged to extend the method.
- Reproducibility Report: Students write a detailed report documenting their re-implementation process, including challenges encountered, experimental setups, and a critical analysis of the original paper's claims and experimental evidence.
- Code Submission: Students provide their re-implemented code, often in Jupyter notebooks, along with instructions for setting up the environment.
- Presentations: Groups present their findings and the reproducibility of their work.
The course also includes lectures covering FACT-AI topics and algorithmic harm, as well as paper discussion sessions where students critically analyze seminal FACT-AI papers.
Over two iterations, the course evolved:
- First Iteration (2019-2020): Focused on creating an open-source GitHub repository of student-implemented FACT-AI algorithms.
- Second Iteration (2020-2021): Encouraged students to submit their reproducibility reports to the Machine Learning Reproducibility Challenge (MLRC), leading to publications in the ReScience journal.
Datasets
The paper does not explicitly list specific datasets used by the students for their re-implementation projects. However, a key criterion for selecting papers for the project is that "At least one dataset in the paper is publicly available." This ensures students have access to the necessary data to reproduce the original experiments.
Key Insights and Ideas
- Reproducibility as a Pedagogical Tool: The paper strongly advocates for using reproducibility not just as a research ideal, but as a powerful teaching mechanism for complex ethical AI concepts. By grappling with the practicalities of re-implementation, students gain a deeper, more critical understanding of FACT principles.
- Bridging Theory and Practice: The hands-on nature of the project allows students to move beyond theoretical discussions of fairness or transparency and directly confront the challenges of implementing and verifying these concepts in real-world algorithms.
- The "Reproducibility Crisis" as a Learning Opportunity: The paper highlights that the very difficulty of reproducing research is a valuable lesson in itself. Students learn to critically assess the trustworthiness of published work and understand the importance of rigorous methodology.
- Community Contribution: By contributing to open-source repositories or submitting to challenges like MLRC, students gain a sense of purpose and contribute to the broader AI research community, enhancing their professional development.
- Adaptability to Online Learning: The paper discusses the challenges and adaptations made when the course transitioned to an online format due to the COVID-19 pandemic, offering insights for remote teaching of similar technical courses.
Key Results
- Student Engagement: The reproducibility project was a significant driver of student satisfaction, with high scores reported for academic challenge, supervision, and workload.
- MLRC Success: In the second iteration, 9 out of 30 student groups had their reproducibility reports accepted for publication in the ReScience journal through the MLRC, demonstrating the high quality of their work and the effectiveness of the course's approach.
- Qualitative Feedback: Students reported that the course provided "good insights into the trustworthiness of research papers," was an "eye-opening" experience, and fostered a more critical approach to academic literature.
- MLRC Reviewer Feedback: Reviews from the MLRC often highlighted that student reports revealed "dark spots of the original paper," provided "strong reproducibility reports," and offered "good insights on how the experiments in the original paper actually work."
Limitations and Future Work
- Technical Focus: While the course excels at teaching the technical aspects of FACT-AI, the authors acknowledge that incorporating more non-technical perspectives (e.g., legal, philosophical) would be beneficial. Future work includes exploring guest lectures from outside computer science, workshop sessions, and broader impact statements in reports.
- Initial Logistical Issues: In the first iteration, some students struggled with the standard procedures of writing and submitting a paper, such as understanding the purpose of an introduction or meeting deadlines. Future iterations will include explicit instruction and examples on these aspects.
- Quality of Reproducibility Examples: The authors plan to provide more examples of high-quality reproducibility papers to better guide students on what constitutes excellent work.
- Online Learning Challenges: While the course adapted to online delivery, the lack of physical interaction posed challenges, particularly in facilitating group discussions and project coordination. The authors suggest that returning to in-person classes would likely improve these aspects.
Summary updated