Twitter is trying to find out the problem with its AI and why its photo-preview algorithm seems to be favoring White faces over Black

A couple of users recently made an interesting revelation, but it could have made Twitter fodder for racial discrimination activists. Luckily, it was soon found out that there is a problem with the neural network that crops photo previews.

It so happened that a White man posted on Twitter about how Zoom kept removing the head of one of his Black colleagues in a call when he tried to use virtual backgrounds. This prompted him and several other users to find out that not only Zoom but Twitter is also doing the same thing. If a photo containing a Blackface and a Whiteface is put for a preview, Twitter’s preview algorithms would pick up the Whiteface most of the time. So, before Twitter would start getting accused of using or incorporating racial bias within its systems, it was demonstrated through several un-official experiments that the problem is not based on any kind of bias, but is a neural network issue that Twitter needs to work on.

During this informal testing, users also discovered that Twitter’s preview algorithm favors non-Black cartoon characters too.

This was an unprecedented finding for Twitter’s machine learning researchers, and they explained the whole situation in a blog post that previously, they used face detection and facial recognition technology to focus the view on the most prominent face they could find in a picture. But this approach has its limitations because not every image has a face in it, so what will the system recognize then? In some cases, the detector would miss faces even if they were there in a picture, and sometimes, it would detect faces when there were none. In the case of photos without faces, the detector would focus on the view in the center of the image, and that would also result in awkwardly cropped photo previews.

So, it seems that this is what is happening with Twitter’s neural network as well as Zoom’s. The chief design officer of Twitter, Dantley Davis also tweeted that the company was looking into this matter and is trying to resolve the issue with its neural network. Davis conducted some unscientific experiments with images on Twitter himself in unofficial posts.

Liz Kelley from the communications team of Twitter also posted that the company looked into the matter to find out any evidence of racial or gender bias, but there is nothing like that going on. In fact, it is just a problem with the technology that needs to be fixed, and Twitter teams are on it. She also tweeted that Twitter would open-source its work so others can also review and replicate if they want.

Twitter’s chief technology officer, Parag Agarwal finds this whole unofficial and informal testing by users quite fascinating and he said that the company’s neural network model surely needs continuous improvement and through such informal tests, the company is open to learning more from such experiments.




Read next: Twitter seems to be testing ways for users to request account verification directly

Featured Photo: Anadolu Agency via Getty Images

No comments:

Post a Comment