In an attempt to fix Twitter’s problematic automated image-cropping tool, its engineers said the software will in future simply rely less on AI algorithms.
When previewing pictures on the social media platform, Twitter automatically crops and resizes the image to match your screen size, be a smartphone display, PC monitor, etc. Twitter uses computer-vision software to decide which part of the pic to focus on, and it tends to home in on women’s chests or those with lighter skin. There are times where it will pick someone with darker skin over a lighter-skinned person, though generally, it seems to prefer women’s chests and lighter skin.
Twitter promised to have a rethink after the apparent bias was highlighted last month. It now believes the best solution is to, well, use less machine learning.
Another reminder that bias, testing, diversity is needed in machine learning: Twitter’s image-crop AI may favor white men, women’s chests
“We are prioritizing work to decrease our reliance on ML-based image cropping by giving people more visibility and control over what their images will look like in a tweet,” Parag Agrawal and Dantley Davis, Twitter’s CTO and chief design officer, respectively, said last week.
“We’ve started exploring different options to see what will work best across the wide range of images people tweet every day. We hope that giving people more choices for image cropping and previewing what they’ll look like in the tweet composer may help reduce the risk of harm,” the pair added.
A Twitter spokesperson told The Register the biz has not decided how best to fix the problem yet: “We’re unable to share specifically how we’ll do it at this point. There’s definitely more to come, including additional technical details on our testing in the next few weeks.”
Twitter’s cropping algorithms compute a “saliency map” of an image, which identifies where the pixel values change the most, pixel to pixel, and thus where there is a high level of detail. This area is likely to be where people will look first, and thus where the image should be cropped. For example, given something like a photograph of a dog playing in a park, the tool should focus on the dog rather than a big patch of blue sky.
The social network’s techies insisted again they tested their software to ensure it wouldn’t pick White people over Black people, say, and found no clear bias favoring particular genders or race. “While our analyses to date haven’t shown racial or gender bias, we recognize that the way we automatically crop photos means there is a potential for harm. We should’ve done a better job of anticipating this possibility when we were first designing and building this product,” it admitted.
In short, the code may indeed not be biased, however there is at least a perception of bias in that users are growing increasingly aware of bias in machine-learning software, and Twitter’s system at times shows a bias when cropping people’s faces. Ergo, Twitter’s technology is deemed to be biased all the time by concerned netizens, and now Twitter has to address that concern – perhaps by ditching the whole automatic cropping entirely. ®