Google develops new measurement of skin tone to reduce product bias

Alphabet’s Google told Reuters this week that it is developing an alternative to the industry-standard method for classifying skin tones.

At issue is a six-color scale known as the Fitzpatrick Skin Type (FST), which dermatologists have been using since the 1970s. Technology companies are now using this to classify people and measure whether products such as facial recognition systems and smartwatch heart rate sensors work equally well across skin tones.

Critics say the FST, which includes four categories of “white” skin and one category of “black” and “brown”, ignores diversity among people of color.U.S. Department of Homeland Security Researcher, During Federal Technical Standards Meeting Last October, the FST did not adequately represent the color range of diverse populations and recommended abandoning the FST to assess face recognition.

In response to Reuters questions about FST GoogleSaid for the first time, and ahead of its peers, that it was quietly pursuing better measures.

“We are working on more comprehensive alternatives that may help in product development and work with scientific and medical professionals and groups that work with the color community,” the company details. Refused to provide. For effort.

The controversy is part of a larger calculation of racism and diversity in high-tech industries where the workforce is whiter than sectors like finance. Ensuring that technology works well not only for all skin tones, but also for different ages and genders is often the case. artificial intelligence Expand to delicate and regulated areas such as (AI), healthcare and law enforcement.

Companies are aware that their products may be defective for groups that are underestimated in survey and test data. The concern with FST is that its limited scale for darker skin can lead to techniques that work for, for example, golden skin, but fail for espresso red tones.

Many types of products offer a much richer palette than FST. Crayola launched 24 skin tone crayons last year, and Mattel’s Barbie Fashionista doll covers nine tones this year.

This issue is far from academic for Google. When the company announced in February that some Android smartphone cameras could measure pulse rate with your fingertips. Said On average, measurements are 1.8% error, regardless of whether the user’s skin is light or dark.

Later company Gave a similar guarantee Its skin type does not significantly affect the results of the ability to filter backgrounds in Meet video conferencing or the results of future web tools for identifying informally dubbed skin conditions. Dam Assist..

These conclusions were drawn from testing with a 6-tone FST.

Starting point

Dr. Thomas Fitzpatrick, late dermatologist at Harvard University Invented Scale to personal UV treatment of psoriasis, an itchy skin condition. He grouped the skin of “white” people as Roman numerals I to IV by asking how much sunburn or sunburn occurred in the sun for a particular period of time.

Ten years later, Type V appeared on “brown” skin and VI appeared on “black”. This scale is part of the US regulation for testing sunscreen products and continues to be a common dermatological standard for assessing patients’ cancer risk and more.

Some dermatologists say that this scale is an inadequate and worn-out scale for care and is often confused with race and ethnicity.

Dr. Susan Taylor, a dermatologist at the University of Pennsylvania who founded the Skin of Color Society in 2004 to facilitate research in a community that has reached its limits, said: “Looking at my skin tone and saying I’m Type V hurt me.”

Until recently, tech companies were indifferent. Unicode, an industry group that oversees emoji, mentioned FST in 2014 as the basis for adopting five skin tones other than yellow. say it The scale was “no negative association”.

2018 study titled “Gender Shades” Face analysis system found People with dark skin and more often misgender, popular using FST to evaluate AI. The study described FST as a “starting point,” but scientists from a similar study, published later, told Reuters that they used scales for consistency.

“As a first step towards a relatively immature market, it serves the purpose of helping us identify danger signals,” said Inioluwa Deborah Raji, a Mozilla Fellow focused on AI auditing. I did.

In April Research AI test to detect deepfake, Facebook Researchers write the FST as “apparently not embracing the variety of brown and black skin tones.” Still, they released 3,000 videos used to evaluate AI systems and tagged them with FST based on the ratings of eight human evaluators.

The judgment of the evaluator is the main focus.Last year, facial recognition software startup AnyVision presented a celebrity Example To evaluators: Former baseball great Derek Jeter as Type IV, model Tyra Banks V, rapper 50 Cent VI.

AnyVision told Reuters that it agreed with Google’s decision to reconsider the use of FST, saying Facebook could accept better measures.

Microsoft And smartwatch makers Apple And Garmin If you work with health-related sensors, see FST.

However, the use of FST provides a “false guarantee” for reading heart rate from a dark-skinned smartwatch, a clinician at the University of California, San Diego, inspired by Black Lives Matter’s social equality movement. It may help. I have written It was published in Sleep magazine last year.

Microsoft has acknowledged a flaw in FST. Apple said it uses a variety of means to test humans beyond skin color. Garmin said he believed the measurements were reliable for extensive testing.

Victor Casale, who founded the makeup company Mob Beauty and helped Crayola with a new crayon, developed 40 shades for the foundation, each about 3% different from the next shade, or so that most adults can distinguish it. I said that I did.

The color accuracy of electronics suggests that technical standards should have 12 to 18 tones, he added, “you can’t have only six.”

© Thomson Reuters 2021 Google develops new measurement of skin tone to reduce product bias

Back to top button