According to a new research by Sony AI, skin color bias is still prevalent on online platforms, despite the introduction of Google’s Monk Skin Tone (MST) Scale over a year ago. Developed in collaboration with Harvard sociologist Ellis Monk, the MST Scale aimed to offer a more inclusive perception of skin color with its ten standard colors, surpassing the limitations of the six-color Fitzpatrick scale. However, the scale still falls short in addressing discrimination stemming from the skin tones used in AI algorithms.
Sony’s research paper highlights additional layers of bias related to apparent skin color within computer vision datasets and models. It emphasizes that AI datasheets and model cards are insufficient in addressing discrimination against under-represented groups. Computer vision, a subset of AI, focuses on enabling computers to interpret and identify the world through images and videos.
Uncovering Biases in AI Algorithms
The study reveals that existing skin color scales impact how AI classifies people and emotions. Altering skin color to be lighter or redder can result in AI classifying non-smiling individuals as smiling, and vice versa. The study argues that AI classifiers also tend to perceive individuals with lighter skin tones as more feminine and those with redder skin hues as more smiley. Sony’s global head of AI ethics, Alice Xiang, notes that this indicates undetected bias in AI algorithms.
The researchers found that AI models exhibit bias not only for skin tone but also for skin hue. William Thong, a computer vision research scientist at Sony AI, explains that individuals with light red skin tones tend to perform better under the algorithm, while those with dark yellow skin tones have lower performance.
The researchers propose that collecting skin tone annotations can assist AI developers in identifying bias in various applications, such as facial recognition, image captioning, person detection, dermatology skin image analysis, face reconstruction, and deep fake detection. By doing so, bias can be recognized and eliminated.
Although Google has implemented the MST Scale in its products, search results on Google Images still use light to dark options based on the scale’s ten color shades. Google Photos has also faced criticism for mislabeling pictures of black people and generating stereotypical images.
Introducing Sony’s ‘Hue Angle’ in Skin Color Classification
To address these issues, Sony is introducing the ‘Hue Angle,’ a new dimension of computer vision that focuses on skin color ranging from red to yellow. Unlike the Fitzpatrick scale that categorizes skin color types, the ‘Hue Angle’ offers a multidimensional measure of skin color. It quantifies the extent to which common image datasets are biased toward light-red skin tones and under-represent dark-yellow skin tones, and how generative models trained on these datasets perpetuate the biases.
The ‘Hue Angle’ tool aims to assess diversity standards in existing data, mitigate skin tone bias in AI datasets and models, and potentially integrate with new AI projects to monitor skin color diversity during data collection and representation. The researchers draw inspiration from personal experiences in buying cosmetics and the works of Brazilian photographer Angelica Dass, who highlighted the diversity of human skin color in her project called Humanae.
While Dass’s project couldn’t analyze everyone, her work, along with Google’s and Sony’s efforts, has the potential to contribute to a more inclusive internet.
Denial of responsibility! Vigour Times is an automatic aggregator of Global media. In each content, the hyperlink to the primary source is specified. All trademarks belong to their rightful owners, and all materials to their authors. For any complaint, please reach us at – [email protected]. We will take necessary action within 24 hours.
