this post was submitted on 28 Nov 2023
617 points (97.2% liked)
Technology
34769 readers
119 users here now
This is the official technology community of Lemmy.ml for all news related to creation and use of technology, and to facilitate civil, meaningful discussion around it.
Ask in DM before posting product reviews or ads. All such posts otherwise are subject to removal.
Rules:
1: All Lemmy rules apply
2: Do not post low effort posts
3: NEVER post naziped*gore stuff
4: Always post article URLs or their archived version URLs as sources, NOT screenshots. Help the blind users.
5: personal rants of Big Tech CEOs like Elon Musk are unwelcome (does not include posts about their companies affecting wide range of people)
6: no advertisement posts unless verified as legitimate and non-exploitative/non-consumerist
7: crypto related posts, unless essential, are disallowed
founded 5 years ago
MODERATORS
you are viewing a single comment's thread
view the rest of the comments
view the rest of the comments
They are not talking about the training process, to combat racial bias on the training process, they insert words on the prompt, like for example "racially ambiguous". For some reason, this time the AI weighted the inserted promt too much that it made Homer from the Caribbean.
They literally say they do this "to combat the racial bias in its training data"
And like I said, this makes no fucking sense.
If your training processes, specifically your training data, has biases, inserting key words does not fix that issue. It literally does nothing to actually combat it. It might hide issues if the data model has sufficient training to do the job with the inserted key words, but that is not a fix, nor combating the issue. It is a cheap hack that does not address the underlying training issues.
congratulations you stumbled upon the reason this is a bad idea all by yourself
all it took was a bit of actually-reading-the-original-post
?
My position was always that this is a bad idea.
the point of the original post is that artificially fixing a bias in training data post-training is a bad idea because it ends up in weird scenarios like this one
your comment is saying that the original post is dumb and betrays a lack of knowledge because artificially fixing a bias in training data post-training would obviously only result in weird scenarios like this one
i don't know what your aim is here
You started your initial rant based on a misunderstanding of what was actually said. Stumbling into the correct answer != knowing what you’re reacting to
Yes. The training data has a bias, and they are using a cheap hack (prompt manipulation) to try to patch it.
Any training data almost certainly has biases. For awhile, if you asked for pictures of people eating waffles or fried chicken they'd very likely be black.
Most of the pictures I tried of kid-type characters were blue eyed.
Then people review the output and say "hey this might still racist, so they tweak things to "diversity" the output. This is likely the result of that, where they've "fixed" one "problem" and created another.
Behold, Homer in brownface. D'oh!
So the issue is not that they don't have diverse training data, the issue is that not all things get equal representation. So their trained model will have biases to produce a white person when you ask generically for a "person". To prevent it from always spitting out a white person when someone prompts the model for a generic person, they inject additional words into the prompt, like "racially ambiguous". Therefore it occasionally encourages/forces more diversity in the results. The issue is that these models are too complex for these kinds of approaches to work seamlessly.