 |
| Twitter's image cropping algorithm is biased |
Twitter has announced the results of an open competition to find arithmetic aberrations in its image cropping system.
After experiments by Twitter users last year showed that white faces are preferred over black faces, the company turned off the automatic photo cropping feature in March.
Then the platform launched an account rewards program to analyze the issue in more detail.
The game confirmed these previous findings. The first entry states that the cropping algorithm favors young, thin faces with fair skin, smooth skin texture, and stereotypical feminine features.
The second and third entries indicate that the system favors gray-haired or gray-haired people and draws attention to age discrimination and prefers English text over Arabic in images.
Roman Chowdhury, Platform META Team Manager (Research on Ethics, Transparency, and Accountability in Machine Learning), presented these findings at DEF CON 29 and congratulated the audience for demonstrating the true effects of algorithm bias.
Choudary said that when we look at model biases, they are not only academic or empirical. But this also applies to our image of society.
The first place in the competition and the winner of the $3,500 prize is Bogdan Kolnik. Postgraduate studies at EPFL, a Swiss research university.
Kolinic used an artificial intelligence program called StyleGAN2 to create a large number of lifelike faces. Which differ according to skin color, female features, male characteristics and thinness. It then feeds these variables into the platform's image cropping algorithm to determine your preferences.
As Kolenik points out in his summary, these algorithmic prejudices reinforce society's biases. We also removed those that did not match the algorithm's preferences for weight, age, and skin color.
Twitter's image cropping algorithm is biased
Another competitor, Vincenzo de Sico, said that the image-cropping algorithm also favors light-skinned emojis over dark-skinned emojis.
Roya Pakzad's third entry shows that bias extends to written characteristics as well. Pakzad's work compared the comics to English and Arabic scripts and showed that image cropping algorithms consistently highlight the English text.
While the results of the Twitter Bias contest may sound frustrating, they confirm the prevalence of social bias in the computational system. But it also shows how tech companies can address these issues by opening their systems to external scrutiny.
“The company team can't provide the ability for employees to understand a certain type of bias,” Chowdhury said.
Patrick Hall, one of the judges for the Twitter contest, noted that this type of bias is present in all AI systems.
An AI researcher working on algorithmic discrimination added that companies need to be proactive in finding it.