AUSTIN (KXAN) — Researchers at the University of Texas at Austin are leading the way when it comes to honing the precision of artificial intelligence.
UT Machine Learning Lab held a public lecture titled “AI for Accurate and Fair Imaging” on Friday.
Researchers from the lab’s Institute for Foundations of Machine Learning (IFML) worked to improve the algorithm that in 2020 produced an internet-famous image of former President Barack Obama, dubbed “White Obama.”
The AI that was supposed to enhance a pixelated, low-resolution photo of the 44th president turned him into a white man instead.
“Even though it looked like a good image, a high-resolution, realistic image of a person, it had a bias,” said IFML co-director Alex Dimakis.
Dimakis and his team were able to improve the technology with good results.
They dug into the initial data used to train the algorithm (made up mostly of white celebrities), but found that wasn’t the problem. The problem was the way the algorithm was built.
“The obsession with getting the answer right tends to amplify even a small bias in the dataset,” Dimakis told KXAN.
“We saw, for example, a turban in one of the images,” he said, referring to another set of enhanced photos. “[The turban] was built like hair.
“It can be very discriminatory,” said IFML director Adam Klivans.
Klivans added that advances in this area of AI could have benefits beyond race and gender; it could also help with medical imaging, giving doctors better images to review.
“If you have an MRI or CT scan, and it’s noisy,” he said, “this technology is great for that.”
Suggest a fix