thinking abt once when a discussion abt bias in ML came up and an ML engineer was like "i don't see how there's a problem with ML or the model, it's just an issue with the training data"https://twitter.com/nicolaskb/status/1244921742486917120 …
-
Amusingly this is actually a problem with the ML model (Probably? Not totally what percentage of under-paid human annotators would classify image 1 as not a gun.)
nope that's not how this works. it's not like the training data contains tons of handheld thermometer pics that are getting mislabeled, it's far more likely that due to training data biases the model has builtin associations between dark skin + guns.https://twitter.com/bjnagel/status/1245300089226174465 …
-
-
I just mean that whenever a supervised learning model forecast disagrees with what a training label would hypothetically be then it's a model error, because it's learning an association of "dark skin->guns" that doesn't generalize outside the training set.
-
yeah i see how it's distinct from the amazon example below but i'd argue it's still a training set issue. can't say if the model is overfitting here — maybe it's learning everything it's fed at the correct level but w/o data bias we wouldn't have told it to learn this at all
- 3 more replies
New conversation -
-
-
The original tweet isn't a super compelling example to me on the racial-bias hypothesis (too many confounding factors), but this response with tweaking and cropping the image is.
Thanks. Twitter will use this to make your timeline better. UndoUndo
-
-
-
I really like this example. I remember seeing a demo of bias in sentiment algs where they swapped ethnicities in a simple sentence, and the sentiment was absolutely not symmetric. These same problems are silently embedded in our entire digital economy. Tech is so far from neutral
Thanks. Twitter will use this to make your timeline better. UndoUndo
-
Loading seems to be taking a while.
Twitter may be over capacity or experiencing a momentary hiccup. Try again or visit Twitter Status for more information.