But when racial information is correlated with the output, to decorate with race, you destroy the input. This is most obvious with a face dataset, but is true with anything race correlated: credit scores, where you live, etc. If you're willing to destroy the training data so it no longer resembles real world information, you might as well just not use it in the first place.

That's what the ethicists say: don't use facial recognition models. Don't work on them. Don't research them. They cannot be both unbiased and useful. And in general, there's few to no uses that are ethical, period.

Well, the ethicists just don't understand the models, then. For instance, there are a bunch of measurements you can take of faces to identify people, if you were doing it manually. Things like pupillary distance, canthal tilt, nose width, etc.

Some of these correlate with race. But only part of the information correlates with race, not all of it. It is, in principle, possible to remove the information that identifies race without destroying the information that identifies the individual. It is true that part of an individual's essential characteristics are their racial characteristics, but it is not true that the only way to identify an individual is their racial characteristics. For instance, there is no way that i'm aware of to infer race from fingerprints, but you can absolutely identify a person by their fingerprints. So, the question is, can we extract a facial fingerprint that identifies a person, but not their race? I think the answer is almost certainly yes, and it is going to be up to a clever model design to do it. But essentially it would look like a GAN where the adversarial component is constantly trying to predict race, while the Generative component is trying to trick the race classifier without tricking the person-identifier.

> Well, the ethicists just don't understand the models, then. For instance, there are a bunch of measurements you can take of faces to identify people, if you were doing it manually. Things like pupillary distance, canthal tilt, nose width, etc.

Or perhaps they understand that this won't work in practice.

That could be. But afaik it hasn't been attempted yet.