Different Dimension Me - Comment #6,076,712

You are viewing a single comment's thread.

You can also view the context of this comment.

drossbots
drossbots

in reply to Jill

This isn't surprising to anyone that understands Machine learning algorithms to a degree. There's a misconception that AI is super intelligent and logical, when in reality it's as stupid as any computer, it can just process huge amounts of data for pattern matching on a large scale. If it lacks data, or is given crap data "which is like, 95% of data" you get shit like this.

It's unavoidable with these kinds of programs; garbage in, garbage out.

+12
๐“ข๐–Ž๐–“๐–†๐–Š๐–‘ ๐“ฒ๐“ผ ๐“ฝ๐“ฑ๐“ฎ ๐“’๐“”๐“ž ๐“ธ๐“ฏ ๐“—๐“ธ๐“ป๐“ท๐”‚
๐“ข๐–Ž๐–“๐–†๐–Š๐–‘ ๐“ฒ๐“ผ ๐“ฝ๐“ฑ๐“ฎ ๐“’๐“”๐“ž ๐“ธ๐“ฏ ๐“—๐“ธ๐“ป๐“ท๐”‚

in reply to Mand'alor

The algorithm changes every man into a generic bishounen. Keyword being "generic".
Generic bishounen are white, therefore every man parsed through this app will be parsed as white.

Technical side: AI are trained on existing images/patterns, and since black/tan bishounen drawn by actual artists are only about 0.01% of the total whole, the AI had never trained to draw them, and the skin tone of the source has no neural connections to any of the appropriate categories.

This is why introducing bias modifiers and conditional categorization (make it learn to recognize separate traits and tell it what to do with them) is important in AI learning algorithms.

+5
mattwo
mattwo

in reply to Jill

"If companies in the US that are mocked for being "woke" like Google can have issues"

Like when Disney shot the live-action Mulan in front of a Uyghur concentration camp and the lead actress, Crystal Liu/Liu Yifei supported the Hong Kong police, cut ties with Adidas after they pulled out of China because of concerns regarding aforementioned concentration camps and yet still somehow managed to also piss off the Chinese nationalists by identifying as "Asian" instead of Chinese?

+1

Hey! You must login or signup first!