Twitter investigating why its picture-cropping algorithm prefers White faces

Re:Twitter investigating why its picture-cropping algorithm ... by nacs on Monday September 21, 2020 @03:06PM (#24547151)

I think it’s pretty clear what’s happening here. It probably detects both the white (Mitch McConnell) and black (Barack Obama) faces as faces but has a higher confidence that Mitch’s face is a face so it picks the white face because it was trained more on white faces.

Definitely some training bias going on.

[ Parent | on HN ]

Re:Twitter investigating why its picture-cropping algorithm ... by tiago_o on Monday September 21, 2020 @10:51PM (#24550412)

Maybe training bias. Maybe confirmation bias.

The truth is, we don’t know. A couple of minor changes in the background and you get a counterexample.

Thinking about a traditional face detection method (e.g. Viola-Jones), facial features are isolated based on shape and contrast. I don’t think it’s controversial to say white faces provide better contrast. Maybe a perfectly equitable training set could help, but I doubt it would. Maybe they use a different method altogether.

The traction of the original tweet speaks volumes on how far identity politics has gone in tech and how easily people get swayed into inflammatory generalizations based on superficial, anecdotal and easily-manufactured evidence.

[ Parent | on HN ]