Twitter To Pay Hackers To Tackle Bias in Its Image Cropping Algorithm
Updated: Aug 21, 2021
Ever noticed your images getting auto-cropped on Twitter and ending up with the subject's head cropped off or one of your friends cropped out of the photo? Well, Twitter's trying to address that by paying people to identify algorithmic biases in how photos are cropped.
In September 2020, Programmer Tony Arcieri demonstrated the algorithmic bias when he tweeted an experiment that featured images of former US President Barack Obama and US Senator Mitch McConnell. On several occasions, Arcieri saw that the Twitter algorithm chose to show McConnell’s face over Obama’s, hinting at racial bias.
It was only when Arcieri inverted the images’ colours that Twitter displayed both Obama and McConnell on the tweet.
Rumman Chowdhury, the Director of Twitter’s ML (Machine Learning) Ethics, Transparency & Accountability (META) team, responded to users’ backlash against the algorithmic bias by announcing that the company conducted tests revealing flaws in the saliency algorithm.
The saliency algorithm was responsible for cropping images on the platform from 2018 until May 2021 when Twitter junked the automated cropping system, admitting that it unintentionally favours white individuals over persons of colour as well as males over females.
Despite making that change, Twitter wants to give programmers a chance to isolate bias in its saliency algorithm by launching the tech scene’s first algorithmic bias bounty competition. In a blog post, Chowdhury and Twitter META Product Manager Jutta Williams explained that the company was inspired by hacking and research initiatives that improved the security industry.
The announcement emphasised Twitter’s desire to spot issues related to ML ethics in order for better algorithms to be implemented in the future.
Credit: Worawee Meepian via Shuttershock
“With this challenge we aim to set a precedent at Twitter, and in the industry, for proactive and collective identification of algorithmic harms. For this challenge, we are re-sharing our saliency model and the code used to generate a crop of an image given a predicted maximally salient point and asking participants to build their own assessment. Successful entries will consider both quantitative and qualitative methods in their approach,” wrote Chowdhury and Williams.
The winning team will receive US$3,500 (approximately S$4732) in cash prizes, while the first runner-up will be given US$1,000 (approximately S$1352). The third place will take home US$500 (approximately S$676). Winners for the Most Innovative and Most Generalisable categories will be awarded US$1,000.
Twitter clarified that the most generalisable entry will be a solution that can be applied to "the most types of algorithms".
To join the contest, interested competitors must sign up with a HackerOne account and submit an application before 6 August 2021, 11:59pm PT. Details about the competition can be found on HackerOne’s site as well.
Written by Sophia Lopez