Meta's open-source FACET tool for assessing racial and gender bias in AI models

IT House reported on September 2 that in order to alleviate the systemic bias problem in many current computer vision models against women and people of color, recently launched a new AI tool called FACET for identification. Racial and gender bias in computer vision systems.

The FACET tool is currently trained on 30,000 images, including images of 50,000 people. It has especially enhanced the perception of gender and skin color and can be used to evaluate computer vision models on various features.

After training, the FACET tool can answer complex questions. For example, when the subject is identified as male, it can further identify skateboarders, as well as light or dark skin

Meta used FACET to evaluate the DINOv2 model and SEERv2 model developed by the company, as well as OpenAI's OpenCLIP model. Overall, OpenCLIP performed better than other models in terms of gender, while DINOv performed better in age and skin color. Good judgment.

Open sourcing

FACET will help researchers perform similar benchmarking to understand biases in their own models and monitor the impact of mitigation measures taken to address equity issues. IT House attaches the Meta press release address here, and interested users can read it in depth. Open-sourcing FACET will help researchers conduct similar benchmarks to understand the bias present in their models and monitor the impact of adopting fair resolution measures. IT House provides a link to Meta’s press release here. Interested users can read it in depth

The above is the detailed content of Meta's open-source FACET tool for assessing racial and gender bias in AI models. For more information, please follow other related articles on the PHP Chinese website!