New Methods to Make AI Vision Systems Fairer

Computer vision systems are everywhere. They help classify and tag images on social media, detect objects and faces in images and videos, and highlight relevant elements in images. However, they are biased, and their accuracy decreases when it comes to images of black or brown people and women. Furthermore, there is another problem: researchers have found that the methods used to detect bias in these systems are biased themselves, as they categorize people into broad categories without fully considering the complexity of human diversity.

Sony and Meta researchers have proposed methods to measure bias in computer vision systems in two new papers, in order to capture human diversity more comprehensively. These papers will be presented at the ICCV computer vision conference in October. Developers can use these tools to examine the diversity of their datasets, helping to generate better and more diverse AI training data. These tools can also be used to measure the diversity of human images generated by generative AI.

Traditionally, skin color bias in computer vision has been measured using the Fitzpatrick scale, which measures from light to dark. Sony AI ethics researcher William Thong said that this scale was originally used to measure the degree of tanning in white skin, but later became widely used as a tool to determine race. It is used to measure bias in computer systems by comparing the accuracy of AI models on light-skinned and dark-skinned populations.

However, describing people's skin color with a one-dimensional scale is misleading, according to Sony AI ethics global lead Alice Xiang. By categorizing people into different groups based on this rough scale, researchers overlook biases that affect groups such as Asians, who are underrepresented in Western AI datasets and can belong to both light-skinned and dark-skinned categories. It also fails to account for the fact that people's skin color changes. For example, researchers point out that Asian skin becomes darker and more yellow as they age, while white skin becomes darker and redder.

Thong and Xiang's team has developed a tool - exclusively shared with MIT Technology Review - that expands the skin color scale to two dimensions, measuring both hue (from light to dark) and chroma (from red to yellow). Sony provides this tool for free online.

Thong said he was inspired by Brazilian artist Angélica Dass, whose work shows that people from similar backgrounds can have a wide range of skin colors. However, the idea of representing skin color comprehensively is not new. The cosmetics industry has been using the same technology for years.

"For anyone who has ever chosen a foundation shade... you know that it's not just about whether someone's skin is light or dark, but also whether it's warm or cool-toned," Xiang said.

Sony's work on skin color "provides an insight that has been a missing component that people have overlooked," said Guha Balakrishnan, an assistant professor at Rice University who has studied bias in computer vision models.

Measuring Bias

Currently, there is no standardized method for measuring bias in computer vision, making it difficult to compare different systems.

To make bias evaluation smoother, Meta has developed a new method for measuring fairness in computer vision models called Fairness in Computer Vision Evaluation (FACET), which can be used for a range of common tasks such as classification, detection, and segmentation. Meta AI researcher Laura Gustafson said FACET is the first fairness evaluation that includes multiple computer vision tasks and includes a broader range of fairness metrics than other bias tools.

To create FACET, Meta collected 32,000 human images and hired annotators from around the world to label them. Annotators were asked to label the images with 13 different visual attributes, such as perceived age, skin color, gender presentation, hair color, and texture. They were also asked to label individuals based on what they were doing or what profession they appeared to have, such as hairdresser, skateboarder, student, musician, or gymnast. The researchers said this increased the granularity and accuracy of bias evaluation.

Meta then used FACET to evaluate the performance of state-of-the-art vision models on different population groups; the results showed significant disparities. For example, the models performed better in detecting people with lighter skin, even if they had braided or curly hair.

Recruiting annotators with geographic diversity is important because people around the world have their own biases when evaluating images of others, said Angelina Wang, a Princeton University doctoral researcher who has studied bias in computer vision models.

Meta's decision to make its data available online for free will also benefit researchers. Annotated data is expensive, so it is only accessible to large tech companies on a large scale. "It's a welcome addition," said Balakrishnan.

However, Wang cautioned that it is wise to be realistic about how much impact these systems can have. They may bring about small improvements rather than a complete transformation of AI.

"I think we're far from capturing how humans express themselves, and maybe we never will," she said.