Hacker News Reader

Comment by panarky | Hacker News Reader

This is an interesting ethical challenge.

If the training corpus for a machine learning model contains stereotypes and biases, then the output of the model will reflect those prejudices.

When that model is used in large-scale applications, it will not just repeat those biases, it will amplify them. It propagates biased language, which feeds back into the model in a self-reinforcing vicious cycle of increasing bias.

Should we attempt to detect prejudices in machine learning models and actively counter them, creating a self-reinforcing virtuous cycle of balance and tolerance?

----

As an aside, for an article about language, this one has a lot of typos. Doesn't anyone proofread before posting?

View on HN

This is the same question as 'What values should we impose on society?' And that is an irreducible conflict. It's politics.

Programming a computer to perform a task is a powerful way to reveal hidden assumptions. What the problem of machine learning on human data reveals, is there is no objectivity. Our most 'objective' models will simply learn and then reinforce biases and prejudices and cause harm to people. We can't build objective, neutral, value-free models when it comes to human behavior, because humans change their behavior in response to the models. When we reject stereotypes we are imposing a set of values, just as much as when we embrace them. Machine learning forces us to confront the fact that this is exactly what we are doing.

There's a kind of philosophical crisis going on here. We need an entirely new way to think about the limits of objectivity in human sciences, and how to create ethical models in the presence of feedback between science and society. The language of objective physical science doesn't work.

Especially as it relates to this: "This means that blogs or websites that employ African-American language could actually be pushed down in search results because of Google’s language processing."

Isn't the ideal system to not assume anything about the content it's analyzing, but to base the model on behavioral patterns instead? In the case of Google ranking, either a site has traffic/low bounce/backlinks/social cred or it doesn't--why is Google attempting to 'read' and analyze the content itself? (Other than to serve ads of course)

> Isn't the ideal system to not assume anything about the content it's analyzing ...

Because Google no longer just searches for literal strings of text, ranked by links.

They now infer your meaning and find results that match your intent -- even if the exact search string doesn't exist in the search results.

For more on this, read about Hinton's work on "thought vectors" [1].

So if Google can't determine the meaning of African American language as well as it can standard English, then that content is likely to be less visible in search results.

[1] https://www.theguardian.com/science/2015/may/21/google-a-ste...

I can't say I would think that Google does a particularly good job at interpreting the intent of my search queries. I often need to reformulate them to get what I'm after. But I doubt that my searches are similar to those of the general public so that is understandable.

I get the same feeling - that Google is often missing the meaning and giving me unrelated keyword matches.

It's even worse than that - it explicitly ignores quoted keywords in many cases. The only intent they care about is what guides you towards their interests.

my favorite is from today when searching for "comparison prohibited benchmarks" google decide to give me results to "comparison ALLOWED benchmarks"

"They now infer your meaning and find results that match your intent -- even if the exact search string doesn't exist in the search results."

Bing seems to do the same thing. I really wish there were a grep for the web because both engines get it wrong even when I try to exclude results.

Try verbatim mode with google: http://paste.click/SWqasi

It's basically grep for the web, since it uses the actual query instead of the meaning of it.

Thanks, I did not know that was an option. It should make a few different queries quite a lot nicer. I was starting to get scared that I was wishing for the old Alta Vista.

But if I searched in the dialect of the content I would get the content. This is simply segregating the internet. Not "pushing down" rankings.

As an aside - this is a slippery slope conversation. Some people believe language requires rules. Other, more special people, think language is fluid and anarchist. The latter of those two do not make search engines.

For your aside, I don't think it's true. Search engine makers are well aware that language usage is what it is, and not what the thesaurus or grammar book says it is. Unfortunately, using machine learning to work out synonyms means you're going to miss out on minority usages, such as specialized academic jargon, or communities with millions of speakers which have slang and conversational grammar unlike "standard" English.

Try some sample "AAVE" texts in Google, it does a decent job of capturing intent.

My guess is that it's unlikely to provide sources that are in one of these dialects, as the link density will be low.

Enjoyed

> Some aspects of communication are likely to prove more challenging, Hinton predicted. “Irony is going to be hard to get,” he said. “You have to be master of the literal first. But then, Americans don’t get irony either. Computers are going to reach the level of Americans before Brits.”

Showing only (or prioritising) same-language results helps with relevancy. A German speaker who searches for “gift” is not looking for ideas for presents, and an Italian speaker who searches for “peperoni” is not looking for the American variant of salami.

A further problem, though, is that search engines try to understand the query and the content searched at a deeper level than mere keyword matches. For example, they may try to pick up synonyms (if I search for “movies in <local area>”, perhaps results for “films” might be helpful too) and associations.

> why is Google attempting to 'read' and analyze the content itself?

So that when you search for "cheesecake recipes", you don't get a well-sourced, high-traffic article from a respectable website describing in detail how the new business initiative by Cheesecake Factory is a recipe for disaster.

* Disclaimer: I pulled that example out of thin air. There's actually no such article.

> If the training corpus for a machine learning model contains stereotypes and biases, then the output of the model will reflect those prejudices.

Depends what you mean by "the corpus". Let's take the example of loan applications. Maybe the training data contains credit scores, biographical information, and personal essays about the applicants from loan officers.

Let's say loan officers tend to use more negative language when talking about black people.

Any decent machine learning system given racial data will not adopt those biases. In fact, it will counteract them. It will determine that e.g. the semantic content of the essays is an indicator of loan suitability, and it will also notice if it has to e.g. add something to the essay score for black people to make the best decisions.

As long as the output data is objective, competent ML systems will account for any bias in the input data.

> Should we attempt to detect prejudices in machine learning models

We don't have to. If the output examples we are training on are true, the ML algorithm won't adopt any incorrect biases.

This is literally the point of machine learning. You want to make the best possible decision. If your algorithm has incorrect human biases, you're losing money. ML naturally accounts for this stuff.

You are only talking about supervised learning.

And the bias people are probably talking about is bias in the data, not bias in the "input", as you put it.

See http://arxiv.org/abs/1607.06520 for a counterexample to your point.

In that paper it sounds like the algorithm is finding true relationships that the author doesn't like. What's the problem with the algorithm?

I don't understand what you mean by "the algorithm" or "true relationships".

Do you know what I mean? There's truth as in what the data says, and there's truth as in what's underlying before social processes " corrupt " the truth. This paper examines ways to get at the latent truth.

Do you think that's a useful endeavor? It's dismissive to say it has anything to do with what the "author likes".

> and there's truth as in what's underlying before social processes " corrupt " the truth.

Ah, so you're defining "truth" as "whatever I imagine the world would be like if society worked the way I wished it did". There's another word for that; "fantasy".

I asked you to define what you mean by "true relationship". And you did not.

> If the output examples we are training on are true, the ML algorithm won't adopt any incorrect biases.

This is rarely the case when working wtih real data, and thus inspecting whether our models are biased against protected classes is probably one of the most important things an ML practitioner should do.

We have a good understanding of the error profiles of most ML algorithms. We can put tight bounds on the difference between predictions and validation data. If ML algorithms make mistakes, it's usually due to noise or low-quality data, not the algorithm itself.

There is no reason an ML algorithm would be biased against a "protected class". It doesn't know what those are. It's possible that the algorithm will uncover a truth that you don't like, like that there as differences in risk profiles across race or gender, but that doesn't mean the algorithm is incorrectly biased. It just means that reality is at odds with how you might want it to be.

Not talking about the algorithm. It's the data that are biased. The choice of algorithm just determines how interpretable those biases actually are.

> It's the data that are biased.

Which data are you referring to? In most cases, the training data isn't human-generated, and if it is, we usually want to match human behavior as close as possible.

Virtually all data used to predict crimes or recidivism is fraught with human bias, for example. Not sure that we want to reproduce the bias of criminal justice system in any prediction problem involving this type of data.

Read anything written by Solon Barocas: http://solon.barocas.org/

How is recidivism data biased? I'm sure that the information gleaned from parole officers and cops might be biased, but as long as the ML system is trained on whether or not someone actually reverted to committing crimes, it should be able to detect bias on the part of P.O.s and other functionaries and give a more accurate determination as to someone's chances of recidivism.

Since most languages have dialects of similar "standard" form divergence, it's an NLP problem that would have to be addressed anyway. While slang might be hard to keep up with, the grammatical patterns of AAVE are well-documented and understood by linguists, and there's plenty of data to model.

The issue of prejudice stems from refusing to treat AAVE as a valid input to an AI, mirroring the racism in American culture that similarly devalues and excludes it. Output can conform to some approximation of "standard" form without being problematic, following the pattern of interaction between standard and dialect speakers. As a speaker of the standard, it's gauche and derisive to imitate or refuse or comprehend a dialect speaker, not to respond in your own form.

So humans do this too naturally - this is how new languages develop over hundreds of years and groups evolve and split. Aside from that, I'm a little nervous about seeing machines accelerate this process.

Yes! Here's an example: http://arxiv.org/abs/1607.06520

Was reading the article and, in true wiki fashion, click the links that seem relevant to get more context. As is often the case I ended up in some other corner of the internet watching interviews with Stephen Pinker. Something in all that input triggered a thought when I read your comment:

What is really the difference between spelling a word wrongly and using a non-dictionary word to express something. We except one as "slang" but will criticize the other because they don't comply with established rules. Isn't that what slang is? A break from established rules.

Anyways, just a thought, because I agree with you; if you publish - proofread. I just find it interesting to throw a Socrates argument at myself from time to time.

Btw and somewhat of topic.. this was the Pinker interview that was most interesting:

https://www.youtube.com/watch?v=egU0dxzFKAQ

> When that model is used in large-scale applications, it will not just repeat those biases, it will amplify them. It propagates biased language, which feeds back into the model in a self-reinforcing vicious cycle of increasing bias.

Which is precisely why social media (e.g. your newsfeed) is such an awful source for news, especially when you consider how low the standards are on digital publishing these days...

> Doesn't anyone proofread before posting?

No.

Reply on news.ycombinator.com