Reducing Bias in AI-based Analysis of Visual Artworks
Zhuomin Zhang, Jia Li, David G. Stork, Elizabeth Mansfield, John E. Russell, Catherine E. Adams, James Z. Wang
Abstract
Empirical research in science and the humanities is vulnerable to bias which, by definition, implies incorrect or misleading findings. Artificial intelligence-based analysis of visual artworks is vulnerable to bias in ways specific to the domain. Works of art belong to a distinct cultural category that often prioritizes such characteristics as hand-craftsmanship, uniqueness, originality, and imaginative content; works of art are also responsive to diverse social and cultural contexts. Ascertaining which features of an artwork can be rightly ascribed to an objective “truth,” without which the concept of bias is not even relevant, is itself challenging. Incorporating expert knowledge into machine learning applications can help reduce bias in final estimates. We review several sources of bias that can occur across different stages of AI-based analysis, protocols and best practices for reducing bias, and approaches to measuring these biases. This systematic investigation of various types of bias can help researchers better understand bias, become aware of practical solutions, and ultimately cultivate the prudent adoption of AI-based approaches to artwork analysis.