Twitter promises cash prizes for identifying bugs in the image cropping algorithm

The Twitter company announced a competition, which invited the participants to find errors in the image cropping algorithm, which could lead to the fact that the cropped image is “harmful”.

To understand what kind of harm we are talking about, it should be remembered that automatic cropping is applied to images on which there may be people. In the case of an incorrect algorithm, the person in the picture may be shown in a derogatory or stereotyped form, incorrectly recognized or excluded from the list, erased, etc. Figuratively speaking, the algorithm will show bias.

Twitter promises cash prizes for identifying bugs in the image cropping algorithm

Back in April, the company said it would investigate the potential “unintended harm” that its algorithms create, starting with the image cropping algorithm. This algorithm, in use since 2018, attempts to focus on the most interesting parts of images when generating a thumbnail preview image. Some users have criticized Twitter’s automatic cropping, arguing that the algorithm focuses on fair-skinned people.

“In May, we shared our approach to detecting bias in our salience algorithm (also known as the image cropping algorithm), and we made our code available so others can reproduce our work,” said the Twitter blog. “We want to take it one step further by inviting and encouraging the community to help identify the potential harm of this algorithm beyond what we have identified ourselves.”

There are prizes for finding bugs: $ 3,500 for first place, $ 1,000 for second place, and $ 500 for third place. In addition, $ 1000 is due for the most innovative approach and another $ 1000 for the most general one, that is, affecting the most types of algorithms.

The harm assessment criteria are given on the competition page. Note that this takes into account intentional and unintentional harm, as well as how large the affected user group is.

The winners will be announced on August 8th.

Source Link

You may also like