[ad_1]
Last fall, Canada Student Colin Madland noticed TwitterHis automatic cropping algorithm continuously selects his face from the photos of the two, instead of his darker colleague, and then displays it in the tweet.This The plot triggers allegations The prejudice of a large number of Twitter users published elongated photos to see if AI would choose white faces instead of blacks, or focus AI on women’s breasts.
At the time, a Twitter spokesperson said algorithm Found before going live in 2018 No evidence of racial or gender bias.Now, the biggest analysis of AI so far is Found the opposite: Twitter’s algorithm favors white people more than white people. The evaluation also found that the AI used to predict the most interesting part of the photo does not focus on the female body on the female face.
Previous test Via twitter And researcher Vinay Prabhu It involves hundreds of images or less. The analysis released Wednesday by Twitter research scientists was based on 10,000 pairs of images from different population groups to test the proponents of the algorithm.
When the algorithm was shown photos of two demographic groups, the researchers found a bias. In the end, the algorithm will choose a person’s face to appear on the Twitter timeline, and some groups on the platform are better represented than others. When the researchers entered a photo of black and white women into the system, the algorithm chose to show white women 64% of the time, while blacks only showed 36% of the time. This is the largest gap among all demographic groups. in. For images of white women and white women, the algorithm shows that woman 62% of the time. For images of white and black women, the algorithm shows that white women spend 57% of the time.
On May 5, Twitter canceled the method of cropping the image of a single photo published using the Twitter smartphone app. Dantley Davis, the chief design officer of Twitter, agreed with this method. since Algorithm controversy broke out last fall.Changes cause people to post Tall photos And sign the end “Surprise” tweet.
The so-called saliency algorithm is still used on Twitter.com, as well as algorithms for cropping multi-image tweets and creating image thumbnails. A Twitter spokesperson said that photos that are too tall or too wide are now being cropped, and the company plans to stop using the algorithm on Twitter. The saliency algorithm is trained by tracking what people see when viewing images.
Other websites, including Facebook and Instagram, used Automatic cropping based on AI. Facebook did not respond to a request for comment.
Unfortunately, allegations of gender and racial prejudice in computer vision systems are quite common. Google recently detailed how to improve the way the dark-skinned Android camera works.Last week group Find algorithm observation Use AI to tag cartoon illustrations on iPhone, and image tag AI that describes people with dark skin as “animals”. An Apple spokesperson declined to comment.
Regardless of the results of the fairness measurement, Twitter researchers said that algorithmic decision-making can keep users away from choices and have a profound impact, especially for marginalized people.
In the newly published study, Twitter researchers said that they found no evidence that photo cropping algorithms prefer women’s bodies to women’s faces. To determine this, they provided the algorithm with 100 randomly selected portraits of women who were identified as women, and found that only three of the centered bodies were on the face. Researchers believe that this is caused by badges or jersey numbers on people’s chests. To conduct this research, the researchers used photos from the WikiCeleb dataset. The identity characteristics of the people in the photo are from Wikidata.
[ad_2]
Source link
