If you look at this single instance, independent of the world around it, it looks ridiculous.

You need to look at it in the context of a long[1] list[2] of works[3] detailing this exact problem, and the frustration that an leading expert in the field still doesn't take it seriously, despite the real-world consequences of that attitude.

1) https://www.pnas.org/content/117/14/7684

2) http://proceedings.mlr.press/v81/buolamwini18a.html?mod=arti...

3) https://www.liebertpub.com/doi/full/10.1089/big.2016.0047

Why not pick all those other hills to die on, then?

If you want to talk about statistical methods being a horrible idea for policing, I'm with you. Probably most machine learning experts would be, too, just like most software engineers don't trust electronic voting. Instead, people chose "dataset bias affects machine learning models", the most basic Machine Learning 101 fact ever, to try and debate against.

It's like the woke people seek out the weakest possible ground because it generates in-group opposition that they can fight with and out-woke. Stronger ground could have national consensus, way beyond bay area consensus.

Actually, that's Gebru's point. Gebru was pointing out frustration that LeCun was focusing on this narrow question of statistical bias, rather than the broader social issues you're getting at.

Your comment, among the sea of others, got through and made some sense to me. My follow-up questions are:

Does that negate the truthfulness of LeCun's response? What would Gebru rather he do or say?

I feel like I can't tell what the desired form of action is from the article, and I'm unsure what good it does for removing bias from ML to attack prominent figures (on twitter let alone anywhere public).

> Instead, people chose "dataset bias affects machine learning models", the most basic Machine Learning 101 fact ever, to try and debate against.

No one disagrees with that statement. The point that the ethicists are making is that that isn't the whole story, and that focusing only on the lowest hanging fruit (dataset bias) is ultimately harmful since it shifts focus away from the other, deeper problems.

I'm not clear about how Yann does not take it seriously? He even admitted respect for Gebru's work. He's just making a technical claim that this is due to data bias. Whether that's true, who knows, but I really don't see how he is being callous in all of this.

I'm not sure I understand what the issue is at all.

To agree with your point, LeCun did many things which indicate he takes this seriously:

1. Responded with a detailed summary of his views.

2. Said he was open to changing his mind in the face of convincing arguments; indicated a desire to learn.

3. Asked people to stop attacking his opponent, in response to unfair attacks on her.

4. Only a week ago, sent multiple tweets out about the problems with a paper which demonstrated racial bias. https://medium.com/@CoalitionForCriticalTechnology/abolish-t...

There's something in between not taking someone seriously and agreeing with their argument.

The problem is, the "Lions of AI" do take them seriously - they just aren't beating their drums about it and making it the top thing they shout about.

All of the same concerns are discussed in their works about representational data and biased algorithms etc...

"Garbage In, Garbage Out" apparently dates back to 1957. I remember the expression from the very first computer book I read in the 80s.

The idea that this concept, one almost as old as computers and practically a theorem of computation, would somehow be racist or trying to paper over racism is ridiculous.

Narrowly scoped to GIGO, the concept can be traced more generally back to Charles Babbage in the 1800s.

> On two occasions I have been asked, 'Pray, Mr. Babbage, if you put into the machine wrong figures, will the right answers come out?' I am not able rightly to apprehend the kind of confusion of ideas that could provoke such a question.

I couldn’t tell you when the discussion about bias and unreliability in ML training sets started, not my field of expertise.

GIGO, which in this context would translate to something like "using a bad dataset means there is a problem with the data, not that the tool is racist". I'm not an expert, but I think the argument here is it's far far too easy to have a "garbage" dataset with these tools. I.e. one that only functions correctly for some racial groups (like this one does). That's only racist if you then try to use it in a situation where it comes in contact with multiple racial groups - like, virtually anywhere in the real world.

I think you're right that this algorithm doesn't do a good job of reconstructing high res faces from pixelated ones, simply because it can't—there simply isn't enough information in the input, and there is a huge set of real people (from many races) that could map to a given pixelated image.

But I do think that when someone who is white (or another race) says "here's a cool thing, check it out" and someone who is black (or another race, different from the first) tries it and realizes it doesn't work for them because of their race, there is a bit of inadvertent racism there. There is an inadvertent exclusion of that second person. I think this happens a lot and in some cases is unavoidable (e.g. language is intertwined with race, and there is a vast amount of information—literature, news, art—that will be inaccessible to you depending on which language(s) you speak), but it is a problem, and it deserves fixing when possible.

The implications in this case are currently small—it's a demo. But there is a risk that these biases could make it into real systems and have real impact on people's lives, and to combat that we need more awareness of this effect in the ML community and more research into possible solutions.