announced Twitter Platform announces the results of an open competition for finding algorithmic bias in its image cropping system.
The company disabled automatic image cropping in March after experiments last year by Twitter users indicated that it favored white faces over black ones.
The platform then launched an algorithm reward program to try to analyze the problem more closely.
The competition confirmed these previous results. Show entry that occupied first place The cropping algorithm favors skinny, young faces with fair skin tone and smooth skin texture with typical feminine facial features.
Show the two entries The second Andthe third that the system was biased against people with white or gray hair, suggesting discrimination based on age, and favored English over Arabic text in pictures.
and in Power point For these findings at DEF CON 29, Roman Choudhury, director of the platform’s META team (which studies machine learning ethics, transparency, and accountability), praised participants for demonstrating the realistic effects of algorithmic bias.
“When we think about the biases in our models, it’s not just about academic or empirical,” Chaudhry said. But how that also works with the way we think about society.
The first place winner in the competition, and the winner of the $3,500 prize, was Bogdan Kulinic. Graduate student at EPFL, a research university in Switzerland.
Kolinic used an artificial intelligence program called StyleGAN2 to create a large number of realistic faces. Which varied according to skin color and female facial features versus male facial features and thinness. He then fed these variables into the platform’s image-cropping algorithm to find your preference.
As Kolinic notes in his summary, these algorithmic biases amplify the biases in society. It also removes those who do not meet the algorithm’s preferences regarding body weight, age and skin color.
Read also: The obsession with automatic image cropping haunts Twitter
Twitter’s image cropping algorithm is biased
Another contestant, Vincenzo de Sico, demonstrated that the image cropping algorithm Help yourself Also emojis with lighter skin tones over emojis with darker skin tones.
Reveal the entry that won the third place, submitted by Roya Pakzad, that biases extend to written features as well. Pakzad’s work compared comics using English and Arabic text, showing that the image cropping algorithm regularly highlights the English text.
Although the results of the Twitter bias competition may seem frustrating, underscoring the pervasive nature of societal bias in algorithmic systems. But it also shows how tech companies can combat these issues by opening their systems to external scrutiny.
“The ability for people to dive deep into a certain type of bias is something that teams in companies don’t have the luxury to do,” Chaudhry said.
Patrick Hall, one of the judges in the Twitter competition, emphasized that such biases exist in all AI systems.
The researcher in artificial intelligence who works in the field of algorithmic discrimination added: Companies need to work proactively to find them.
Read also: Twitter wants to make you able to remove certain followers