Twitter planned to alter its existing machine learning (ML) based image cropping system to eliminate racial bias.
It plans to rely less on ML for mage cropping, and bring more transparency around the process.
"Bias in ML systems is an industry-wide issue," Twitter said.
The company believes there is potential harm in the way the system automatically crops photos. The micro-blogging site plans to give people more visibility and control over what their images will look like in a Tweet.
It will soon follow “what you see is what you get” principle of design. This will allow users to preview what they will look like in the Tweet composer. The photo that they see in the composer is how they will appear in the Tweet.
Twitter's measure comes after recent conversation around their photo cropping methods.
"We’ve been reviewing the way we test for bias in our systems and discussing ways we can improve how we display images on Twitter," it said.
It tested the existing ML system that decides how to crop images before bringing it to the platform. The image cropping system relies on saliency, which predicts where people might look first.
For this, it tested pairwise preference between demographic groups, White-Black, White-Indian, White-Asian and male-female.
In each trial, it combined two faces into the same image and computed the saliency map over the image. It repeated this 200 times for each pair of demographic categories and evaluated the frequency of preferring one over the other.
It is also planning to open-source the analysis to let others also help.
