PAIR AI Explorables | Is the problem in the data? Examples on Fairness, Diversity, and Bias.

In the recurring debate about bias in Machine Learning models, there is a growing argument saying that "the problem is not in the data", often citing the influence of various choices like loss functions or network architecture. In this video, we take a look at PAIR's AI Explorables through the lens of whether or not the bias problem is a data problem.

OUTLINE:
0:00 - Intro & Overview
1:45 - Recap: Bias in ML
4:25 - AI Explorables
5:40 - Measuring Fairness Explorable
11:00 - Hidden Bias Explorable
16:10 - Measuring Diversity Explorable
23:00 - Conclusion & Comments

Links:

If you want to support me, the best thing to do is to share out the content :)

If you want to support me financially (completely optional and voluntary, but a lot of people have asked for this):
Bitcoin (BTC): bc1q49lsw3q325tr58ygf8sudx2dqfguclvngvy2cq
Ethereum (ETH): 0x7ad3513E3B8f66799f507Aa7874b1B0eBC7F85e2
Litecoin (LTC): LQW2TRyKYetVC8WjFkhpPhtpbDM4Vw7r9m
Monero (XMR): 4ACL8AGrEo5hAir8A9CeVrW8pEauWvnp1WnSDZxW7tziCDLhZAGsgzhRQABDnFy8yuM9fWJDviJPHKRjV4FWt19CJZN9D4n

[AICoffeeBreak]

So many problems (including the debates about bias) would vanish if statistics were more intuitive to humans. It is surprising to see how much misunderstandings come from us humans being so bad at statistics; but by no means I say this to blame or shame: Statistics is not particularly intuitive to us. Just imagine in comparison how natural it is for us to throw a ball. On the other hand, every time statistics comes into play (see elections or COVID-19), it immediately gets controversial. A a lot of concurrent interpretations about just one histogram join the (often political) fight; with the actual statistical explanation being often the least credible because it is the least intuitive. Imagine this controversy and multiply it with all the "statistically spoiled places" when a ML system acts on the world. I do not expect this bias discussion and controversy to ever end (unless all humans suddenly become "native in statistics"). Therefore it is even more important to discuss and educate about this, as done e.g. in this video. Thanks for it!

[MachineLearningStreetTalk]

18:00 "Wish engines -- why use a search engine when I already know what I want to come out?" 😂

[alonamaloh]

In the last scenario, the problem is not even in the data. The data is fine. The problem is that they don't like it.

[robby]

Great video, just a quick point. Even if you choose to represent the data as you found it, e.g. in a search engine, you're already doing social engineering. The effects on society start when you decide to implement a model, not when you start to think about whether it can cause harm.

[nomenec]

In statistics/mathematics, the correct word for "she is beautiful" being more common than "he is beautiful" is CORRELATION not bias. A source of misunderstanding is that to many lay folk and activists "bias" = "correlation we don't like".

[mobkiller]

In regards to the point of fairness and "wish machines", I feel like it shouldn't be my job as an ML practitioner to socially engineer the world, simply by using the same argument they used in their previous examples about biases. My wish for society might be totally different from what is "good" or "bad" in the future. Pressuring systems or individuals to make these choices will also lead to negative feedback loops in the future.
Obviously there are mechanisms like divers teams try to compensate for that, but even that falls short when we compare the diversity available in certain regions of the world e.g. teams form the US have way better chances to create divers teams (since they inherently live in a more divers society) than teams from Europe or places like India / Pakistan, where there is overall fewer possibilities to create divers teams.
I don't really see a solution to that, since people from different cultural backgrounds and norms will also have different understandings of diversity and their ideal society.
All in all, I topics like this just needs to start with a broader societal discussion, rather than being discussed by experts of a field whose views are also not reflective of reality. Otherwise we will just go on and create other negative feedback loops, while not solving the inherent problem.

[mizupof]

First, thanks for introducing this website, I didn't know it. Then, this is a great video. Very interesting to follow from start to end!

[scottmiller]

Not only is it important to recognize (as you rightly have) that statistical bias is not the same as polico-socilogical bias (and hence the latter should be called something else to avoid confusion, although "fairness" already has an explicit definition in interpretable and explainable models, and hence still might not be a good replacement term), but that statistical bias is sometimes desirable. Biased models often generalize better than unbiased models. The BLUE (best linear unbiased estimator) is often ordinary least squares (OLS) (see Gauss-Markov theorem for the exact conditions under which this holds), but the biased LASSO can often beat OLS error in predictions, interpretability, generalization, and data efficiency. LASSO is biased in the statistical sense.

Whether or not this is biased in the polico-sociological sense depends on your policy goals, which is outside the scope of statistics. Politicians do not view the ambiguity of "bias" as a bug, but as a feature, since it lets them use it in whichever sense gives them a political advantage at any given moment, and they can trust that the public usually don't know the difference, and will quickly forget what the politician said mere moments ago, often in the same sentence. In the beginning, this was just ignorance on the politicians part, but now ambiguity in the sense of "bias" in statistics and politics is being knowingly exploited to manipulate public opinion. It's the same reason that proposed law changes in US congress have names that are often unrelated to, or the exact opposite of, the impact they have when implemented as a law.

[TheEbbemonster]

You are an insightful man Dr Kilcher! Since I discovered your channel I have started watching your videos instead of cat videos, and I have learned a lot. You are really good at logically breaking down the problems and bringing more sense than emotions into the discussion.

[jean-baptistedelabroise]

if they continue on this slope, they will use adversarial neural networks to generate 'unbiased' outputs for their wish engine

[pierrehulot]

thank you for this video, great summary of the fairness problem. do you want to represent the reality (or the data) or your reality ?

[CaptchaSamurai]

22:49 Using Yannic's argument we could say "Why Porsche is making commercials!? Do they really think people will spend 100k based on 30s clip?"

As far as I know the presumed mechanisms behind biases are more subtle [1]. More like commercials or constant suggestions. Result of exposure are presumed to be blind spots. Thus one, under the influences of exposure, can not go meta to ask whether sth is okay, bc this is outside one's thoughtable thoughts.

Thank you for the video anyway! The rest was clear and representative. I appreciate your willingness to understand all sides. I'm also very happy, that the term "explorables" (by Nicky Case, I think) went mainstream!

[DamianReloaded]

21:26 Consider you had 2 types of information that you regarded as equally important, A and B, but type A was only 10% as common as type B for which there are more samples. Would you want your search results to only show you results where only 10% of the information presented to you is related to type A? If you consider both types A and B equally important, you want them to be equally available to you regardless of the amount of data available to show. So if you have 10 results of type A and 90 of type B, the search engine can decide to show you the 10 results of type A on the first page to make sure you don't miss them. (instead of giving you 1 type A result every 9 of type B on average). EDIT: I reflect this is akin to spaces for handicapped people, who are usually a minority but we aspire to have them well represented when it comes to bathroom availability or public transport.

[jabowery]

I'm not sure if this comment was deleted by Yannic or by YouTube, or if I have a false memory of having posted it, but it is of rather crucial importance to drawing proper distinctions so, at the risk of being rude I'm re-writing it in a different form that will hopefully persist:

Taken to the limit of including the set of _all_ data in one's modeling efforts, avoiding bias "in the data" is indistinguishable from avoiding bias "in the model" of the data, as long as the model selection criterion is unbiased. The reason for this is fairly straight forward: If you have an "evidence-based" opinion of what constitutes "bias", then that evidence is, by definition, present in "the set of _all_ data in one's modeling efforts." This results in a model that includes a model of what constitutes "bias", so that one's model detects, and corrects for, bias in the rest of the data.

On the other hand, if one's opinion of what constitutes "bias" is _not_ "evidence-based", then one is advocating a mere subjective opinion or making value judgments -- unhinged from science as a public activity. That's fine, but one is risking injecting one's own biases in the guise of addressing algorithmic bias.

So far so good?

Practical approximations to the above limit are necessary. People will argue that their definition of "bias" is "evidence-based" whether they are advocating an opinion or not. That's fine. Let them provide their evidence and be sure to include it in the canonical set of data to be modeled. This canonical set of data will grow to a large size but is not "the set of _all_ data". It is a practical approximation thereof that leads to an unbiased definition of "bias" "as long as the model selection criterion is unbiased".

So now we are at the question of which model selection criterion is unbiased or least-biased of those proffered.

In the limit of Ray Solomonoff's proof that the Kolmogorov Complexity of a set of data drawn from a computable universe is optimal, we have just such a model selection criterion. Metaphysical arguments notwithstanding, we are again looking for practical approximations. A practical approximation of Kolmogorov Complexity is smallest existing program that outputs the canonical data set, as in The Hutter Prize for Lossless Compression of Human Knowledge. It's fine for people to correctly assert that this approximation of Solomonoff Induction doesn't yield the best of all possible models of the data, but they do have an obligation to seriously offer a superior model selection criterion.

[gianpierocea]

Amazing. I took the time to see the video as soon as it was posted and I was looking forward to see the conversation it would have sparked here in the comments. Thanks for making me more aware of these topics, which I think are bound to become more and more pressing.
As someone else has said already, it seems to me a bit naive to assume that the results should "bend" to our concept of what is good for the world (more: someone might say this is also very pretentious. Again looking at history it would not be the first time).
Then again I also realize this is just my opinion and I can very much understand the points brought in from the other side.
In any case the debate is open and it's very exciting (with even bits of scary thrown in there) given the enormous impact this can have on society as a whole.

[MachineLearningStreetTalk]

11:00 Don't you think it's inconvenient when... Reality drop truth bombs into your data. Damn you reality! Damn you!!! 😂😂

[joomyjoo]

great video! this is great teaching material!

[piernikowyloodek]

This discussion happens on a very abstract level. On a more down-to-earth / practical-engineering level - the question instead becomes -- Do the tools I build achieve their purpose?

If I train a face detector on a dataset where only 10% of people have darker skin tones, my system will underperform on people with darker skin tones. Even if 10% might be a correct representation of the population I operate in - I have built a poor face detector. We could even call it a biased one. If there are AT ALL any techniques to fix this, or for that matter fix poor performance on any identifiable edge cases, they should be used. Here debiasing is well-aligned with building a better product.

[jahcane]

Next Minecraft series > analogue hopefield network?

[FlyingOctopus]

Also isn't choosing how to represent reality with data a kind of "biased" choice. What attributes to include? Are they binary or not? What to do with exceptions? Reality is more complicated than we can formalize it.

When google shows results of doctors, should it show images according to distribution of doctors in images(including actors), or of actual doctors in your country, or of doctors in the whole world. There is always a decision what slice of reality your data should represent. So I think the problems with "reality" are more of a "I want to represent a different slice of reality" rather than "I want to change reality".