Twitter consistently centers image previews on whiter faces(twitter.com)
twitter.com
Twitter consistently centers image previews on whiter faces
https://twitter.com/bascule/status/1307440596668182528
16 comments
Perfect example of how not proactively making an algorithm anti-racist can yield results that end up being racist. If you train a machine learning algorithm on photos with predominantly white faces, your algorithm is going to think people with lighter skin are more important thank people with darker skin.
Why are people downvoting this? The parent is correct. If you don't test your models adequately then they'll have holes like this one.
HN doesn't like talking about race or numerous other "controversial" topics. This story is already flagged off the front page.
> HN doesn't like talking about race or numerous other "controversial" topics.
There was plenty of discussions on HN when GitHub was renaming master/slave terminology [0] and when other open source projects were doing the same. Essentially they break the HN guidelines when they 'like' if it is a company / person they 'like'.
On the other hand, when a topic attacks their 'narrative' or attacks their side of the argument from someone they don't like, they downvote, flag and censor and make sure it is never seen. Even if it is factual and has evidence.
[0] https://news.ycombinator.com/item?id=23518123
There was plenty of discussions on HN when GitHub was renaming master/slave terminology [0] and when other open source projects were doing the same. Essentially they break the HN guidelines when they 'like' if it is a company / person they 'like'.
On the other hand, when a topic attacks their 'narrative' or attacks their side of the argument from someone they don't like, they downvote, flag and censor and make sure it is never seen. Even if it is factual and has evidence.
[0] https://news.ycombinator.com/item?id=23518123
I find this on hacker news by time site. These discussions become inflated by politics, so are moderated.
Because we are not children. We understand the idea training data can affect the results.
OP has no evidence for the incendiary accusation and Twitter denies it.
https://twitter.com/ZehanWang/status/1307461285811032066?s=2...
OP has no evidence for the incendiary accusation and Twitter denies it.
https://twitter.com/ZehanWang/status/1307461285811032066?s=2...
This assessment sounds correct. The silver lining here is that engineers will pay more attention to algorithm biases in the future.
rvz(1)
Wow.
(nothing else to say)
(nothing else to say)
In case anyone else did not realize this and is as confused as I was, apparently you can attach more than one photo to a tweet, and Twitter will then show previews of all of them in a gallery. The previews are formed by taking some part of the original image. Clicking on any of the previews opens a view of the corresponding full image.
I knew about twitter taking a part of an image and using that for a preview, but had not realized that there could be more than one image in a tweet.
I knew about twitter taking a part of an image and using that for a preview, but had not realized that there could be more than one image in a tweet.
Consistent in the title is misleading because this is essentially n=1. The Twitter user did try to transform the image in order to produce a different result but we don’t know if there is something not race based in Mitch’s face that the algorithm prefer’s over Obama’s. We also don’t know if the user has cherry picked this example.
Why is this flagged?