In general I take these types of tropes to be cloaked (politically correct) way of saying "fuck x group".
Dataset bias is only one way this occurs, though. Even if you include more colored faces, you still may be taking pictures with a camera that doesn't capture as much contrast. Or even if you retrain, you may have chosen model structure to optimize behavior in a biased dataset. Or... if you're considering the cost of misidentification by a facial recognition system used by, say, police, you need to be sure that you take a perspective that applies to society as a whole and not to your own interactions with police.
It's perfectly reasonable to say "AI is worsening the experience of minorities in various ways" and to differ with responses that assume with slight additional care in curating datasets that the problem will be completely solved.
In France, since you have to be able to explain in plain words the algorithm that was used to help to make an administrative decision, it indirectly makes these black box learning systems illegal.
(And yet they probably already have been used by the police?)
Now, the naive response is that the engineer that implemented a system should take responsibility for any harm done by the system - and I guess that if these engineers could lose their license to use a computer and face jail time, they might also start to put pressure upstream on researchers to provide better systems ?
That statement is contained in a paper I linked that dates from 1980. I'm pretty sure it's factually correct, because I'm trying to make nothing but a demonstration that an idée reçue, that ML systems are only based due to their data, is incorrect.
There's really no cloaked message behind it, and I'm not sure why one would think there is (hence the disclaimer). I think it's a pretty uncharitable interpretation.
That's reasonable. I may have taken the initial text in the wrong context. Thanks for clarifying.
To expand on my prior comment of "Fuck x group", perhaps the tone of prior tweets by Timnit Gebru short-cut me mentally, namely this one: https://twitter.com/timnitGebru/status/1274809418475950080
> Even amidst of world wide protests people don’t hear our voices and try to learn from us, they assume they’re experts in everything. Let us lead her and you follow. Just listen. And learn from scholars like @ruha9 We even brought her to your house, your conference.
It strikes me as very unprofessional, and frankly I believe her anti-white bias shone through in that instance.
Racism is racism, even for "good" purposes.
Personally, I don't think either indicate racism. I see the link you tweeted as an excessively abrasive response at a dismissive and reductive understanding of an issue, bias in learning system, despite research old and new. At the same time, I see LeCun's dismissal as a genuinely held position in error.
To see the first as racist but the second as okay seems to be a case of being much more charitable to one side than the other, though I understand how it may look this way.
This is a consequence of the fact that it's not possible to do generalization, and thus learning, without bias to begin with. What an AI researcher does when deciding which architecture is essentially tuning the bias in the generalization system in order to produce better results, which is not in and of itself bad, and is indeed unavoidable.
By bias, I mean really anything that means that the function g() is used over the function h(). In practice this can be anything from the seed for the random values, to the choice of activation function, to the architecture of the network, and so on.
If the architectures are selected for their performance on dataset y, then it is possible that even when trained on dataset x, they retain some of the bias of dataset y, because their generalization bias, which is architectural, was tuned for dataset y. This is, admittedly, a theoretical result, and it's not clear that it applies in every case, but it is a solid point.
For that reason the point that benchmarking against non-biased datasets is important was made, and thus that there is more to bias than the final dataset used for training during deployment. Therefore, it's not only a responsibility of the engineering community, but also of the scientific community.
I've posted it somewhere here already, but here's a nice paper that explains the point that there is more to bias than simply the dataset : http://www.cs.cmu.edu/~tom/pubs/NeedForBias_1980.pdf
Interestingly enough, this is true of all learning systems, including humans, and is a proof of why it's impossible to be completely unbiased in really anything.