OK, so I'm going to throw my teeny-tiny little lecturer-in-a-remote-land hat in the ring here. Quite understandably, people like @timnitGebru are extremely tired of explaining social-power-and-AI things over and over again to people like @ylecun but...
I, a real life person of colour among other things, almost *never* get tired of 'splaining things and arguing picayune points. I can't claim to completely reflect Timnit's stance but I can take my own stab at it, so feel free to argue with me, I like arguing.
So here goes: the point, to me, is that neither the data nor the models in machine/deep learning or AI or whatever are amenable to a technical bias fix, because the question is ill-posed and they are the wrong target.
The problem is that *task* itself and the *context* that enables the task. We will NEVER come up with an "unbiased" face depixelizer, because the act of performing such operations on human data IS THE BIAS.
Almost always, these complex problems have two twin conjoined kernels at their core: gaze and objectification.
Gaze ≈ who is looking down into the microscope
Objectification ≈ how the specimens got affixed to the slides!
Objectification ≈ how the specimens got affixed to the slides!
In real life, we are not all affixed to the slide in the same way. The *consequences* of misidentification differ for different people, in a systematic way. It's not the same for me to give up my data for face boarding at an airport compared to Yann LeCun.
The person looking through the microscope disposes of samples differently -- that is the power of gaze. The ways in which *I* can be misidentified have a different probability distribution of consequences than the ways in which LeCun can be misidentified.
No amount of data debiasing or model engineering/expansion can fix this. Even if LeCun and I have the *same* probability of misidentification, the consequences of gaze relative to my Moment of objectification are just different.
Furthermore, this turns the theoretical possibility of debiasing (in the sense of giving everyone an equal chance of misidentification) into a practical impossibility.
Because to collect the data, you must do so under unequal conditions. The data is extracted from the more-consequentially-objectified group itself under a state of relatively greater suspicion and punitive consequence.
Now does this mean we cannot necessarily *ever* build NLP applications based on training models collected over mass data?
No, it does not necessarily mean that! It may just mean we cannot do so without explicit awareness of gaze and objectification.
But it turns out that even *that* is an unacknowledged sticking point.
And with that, I think I have adequately 'splained the issue in my own terms that are hopefully acceptable to everyone?
HAHAHA
HAHAHA
BTW, I actually wrote something in the context of usability research 17 years ago that is very compatible with the current debate, recently posted here: https://arxiv.org/abs/1904.03495