MIT researchers allegedly came up with a way to reduce bias in machine learning. Something called "Partial Attribute Decorrelation" which seems to be training a neural net on the "biasing" attributes and then "decorrelating" that metric with the full one. I don't know what that means exactly--a naive "subtraction" of the coefficients in one neural net from another would seem prone to weird behaviors.
Anyhow, I started casting about for the original paper, and found that site after site had copied the press release verbatim, or after only changing a word here or there. However, some got creative.
One sentence from the original read "The researchers’ solution, called Partial Attribute Decorrelation (PARADE), involves training the model to learn a separate similarity metric for a sensitive attribute, like skin tone, and then decorrelating the skin tone similarity metric from the targeted similarity metric."
"The researchers’ answer, referred to as Partial Attribute Decorrelation (PARADE), entails coaching the mannequin to be taught a separate similarity metric for a delicate attribute, like pores and skin tone, after which decorrelating the pores and skin tone similarity metric from the focused similarity metric."
I assume an algorithm was used to substitute equivalent words here and there. For some definitions of equivalent.
Found the paper and will be reading it--some cruft on MIT's web page obscured it.
No comments:
Post a Comment