Researchers Becky Desrosiers, Abner Casillas-Colon, George Shoriz, and Naomi Ohashi undertook a capstone project that evaluates Intel Labs' Open Model Zoo. Intel’s Open Model Zoo (OMZ) is an offering that allows regular people to use pre-trained AI models for their own purposes. The team intended to produce a tool or methodology that will help users identify and quantify bias in Intel’s models, with an emphasis on the context of the model.
This project is important because AI models can easily have bias inherent to their training, which can impact their performance and their impact on society, depending on for what purpose they are leveraged. Since OMZ is publicly available, the potential for applications is extensive, and so are the potential for consequences from biased training. Stakeholders could include anyone who employs the model, or anyone who could be affected by myriad applications of the model.
The project evaluated the face-detection-0200 model from Intel Labs’ Open Model Zoo for potential bias against protected characteristic(s). The team started by researching bias metrics, identifying useful datasets, and choosing a model that would be feasible to evaluate. Finally, the model’s inference on the dataset was investigated using the bias evaluation tool Aequitas to find moderate bias in the model’s performance. All code and materials are available at on GitHub.
Researchers: Becky Desrosiers, Abner Casillas-Colon, George Shoriz, Naomi Ohashi
Sponsors: Emanuel Moss, Elizabeth Watkins, Dawn Nafus
Advisors: Philip Waggoner