Mitigating AI Bias Through Astronomical User-inspired Science

Regents Of The University Of Michigan - Ann Arbor, Ann Arbor MI

Investigators

Abstract

Recently there has been an explosion in the use of artificial intelligence (AI) techniques within astronomy. Astronomy is not immune to human and systemic biases, which go beyond representative datasets and agnostic machine learning algorithms. As emphasized in Special 1270 from the National Institutes of Standards and Technology, researchers should operationalize Trustworthy and Responsible AI and create new norms on how AI is built. This team of investigators will define these norms by identifying, characterizing, and mitigating gold standard labeling bias in galaxy shape classifications. These biases can cause downstream effects in (semi) supervised AI models. The goal of this project is to solve gold label bias mitigation by linking the effort to an interesting astronomical scientific use case. As an additional assessment, they will apply their techniques to medical imaging data to improve AI-assisted tumor detection. The researchers will train an AI system to predict the spatio-spectral properties of the stellar light inside galaxies from imaging data alone. This system will map resolved stellar properties like stellar mass, age, or metallicity to galaxy internal shape properties like bars, spiral arms, and central bulges. They aim to achieve an order of magnitude increase in the sample sizes of resolved stellar properties used for galaxy evolution studies. Other researchers can use this model to infer pixel-level properties that fall well below the typical signal-to-noise thresholds required for expensive spectroscopic studies. A critical component of this research program will be the identification, quantification, and mitigation of gold standard label bias, which they have shown to affect current human-based morphology datasets. In addition to the astronomical data, the team will utilize the screening mammography breast cancer detection dataset from the recent classification KAGGLE competition sponsored by the Radiological Society of North America. The PI will collaborate with staff and researchers in the Michigan Institute for Data and AI in Society (MIDAS) to identify additional datasets that can be studied to assess and mitigate labeling bias. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.