Re:Microsoft lays off an ethical AI team as it doubles down ...
by AstralStorm on Tuesday March 14, 2023 @08:48AM
(#35150615)
You can certainly try this kind of data augmentation strategy. Pretty sure it'll fail because of the text analysis still being biased.
You would need equalized amounts of text referring to various groups of people and topics too. That's much harder to augment.
Re:Microsoft lays off an ethical AI team as it doubles down ...
by pmarreck on Tuesday March 14, 2023 @07:36PM
(#35161134)
What's stopping someone from taking training text that mentions races by name (using any common name) and having it automatically changed to a different race for a new training sample?
honestly in order to make race an irrelevant or random factor, I believe the training data would have to be evenly divided between all known races... and possibly some blends of races? And I think that would be a good thing.
Could you synthesize such data by (ironically perhaps) using AI to change the race of a person in a given training photo?