15 Jun 2023 — A team of researchers from NUS Presidential Young Professor Reza Shokri’s lab has been honoured with the Best Paper Award at the sixth annual ACM Conference on Fairness, Accountability, and Transparency (ACM FAccT) held in Chicago. The awarded paper entitled “On The Impact of Machine Learning Randomness on Group Fairness,” addresses the variability in statistical measures for group fairness in machine learning and investigates the role of different sources of randomness in training neural networks.
The research team, consisting of MComp graduate Prakhar Ganesh, PhD students Hongyan Chang and Martin Strobel, and Prof. Reza Shokri, delved into the high variance exhibited by group fairness measures, rendering them unreliable for empirical evaluations. They focused on uncovering the factors contributing to this variance and explored ways to control group-level accuracy without compromising overall model performance.
Their investigation revealed that the dominant source of randomness affecting fairness measures is the stochasticity of data order during training. By reshuffling the data during training, significant changes in fairness were observed, even between consecutive epochs within a single run. The team demonstrated that minorities are particularly vulnerable to shifting model behavior, resulting in higher prediction uncertainty for under-represented groups. Consequently, any statistical fairness measure based on model predictions is impacted by the disparate prediction uncertainty between groups.
Additionally, the researchers highlighted the immediate impact of data order on fairness, finding that a model’s fairness score is heavily influenced by recent gradient updates, regardless of preceding training. They proposed the use of fairness variance across epochs within a single training run as a proxy for studying fairness variance across multiple runs, significantly reducing computational requirements. Moreover, the team presented a custom data ordering technique that effectively improves model fairness within a single training epoch.
The ACM FAccT conference provides a platform for researchers and practitioners from diverse fields such as computer science, social science, law, policy, and ethics to explore fairness, accountability, transparency, and ethical considerations in the context of artificial intelligence, machine learning, and data science. It serves as an avenue for presenting research papers, fostering discussions, and addressing the challenges associated with responsible AI and ML usage.
The Best Paper Award received by the researchers from Prof. Reza Shokri’s lab underscores their valuable contributions to understanding and mitigating the risks of machine learning. Their findings have practical implications for enhancing model fairness and reducing computational requirements in training processes.
Assistant Professor Yair Zick: Ethics in Artificial Intelligence