Current Events > Racist artificial intelligence algorithm

(+) Yeah! (+)

https://metro.co.uk/2019/09/17/racist-ai-algorithm-classes-black-people-negroids-blackamoors-10757032/

is this allowed to post on GameFAQs?

---

It shouldn't be. Just for posting it, even though I'm not clicking it, you're racist.
---

How did some of those words even get into the system? Apart from the racial stuff there's a picture with a woman tagged as prophetess.

The researchers probably should do a better job reviewing the annotations from racist assholes on MTurk

Good luck cleaning that up now
---

I did a Machine Learning course at Udacity and one of our tasks was to program a Dog breed classifier. In the project, we also had it able to recognize human faces as well. I'm darker skinned than the guy in that article and I tested my project on my own face. No dice. I tried a friend of mine who is "clear skinned" (or several shades lighter than caramel for a reference point) for a black guy, no dice. Grabbed a photo of a Russian dude who is the CEO of a company I'm invested in, boom, works.

We need more diverse people in the AI/ML field to bring in more diverse datasets (this is probably the only time I'd push for diversity). Datasets are going to reflect the priorities of those who make them. On the flipside, it's a hard sell because this is a deep field and you have to be genuinely interested to commit to learning this stuff.
---

^I've heard this before but why doesn't AI recognize darker people? People say "we need to fix this" but no one seems to say what the actual problem is. Is this something to do with the actual input of the developers or is it just a visual thing the machines have a problem with?

specialkid8 posted...

but no one seems to say what the actual problem is

Labels don't pop out of nowhere. They're created by human annotators whether it's the developer or people hired to do it

That's all there is to that. And for computer vision problems, you can't make algorithms that are robust to skin color if the developers dont bother to reach out and gather a more diverse data set. There are black and other minority researchers making a push to stop these dumbass labels from popping up in systems by collecting their own data or modifying existing data sets but there are not that many of them out there in the first place

https://www.businessinsider.com/biases-ethics-facial-recognition-ai-mit-joy-buolamwini-2019-1
---

As opposed to mongolod or caucasoid, which the man clearly isn't. It's just a human taxonomy term. You should all be far more offended by the word Mulatto

---

specialkid8 posted...

I've heard this before but why doesn't AI recognize darker people?

Hanlon's Razor suggests the problem may be sub-optimal lighting, rather than racism.
https://create.adobe.com/2017/11/28/_10_tips_for_photographing_darker_skin_tones.html

Questionmarktarius posted...

specialkid8 posted...

I've heard this before but why doesn't AI recognize darker people?

Hanlon's Razor suggests the problem may be sub-optimal lighting, rather than racism.
https://create.adobe.com/2017/11/28/_10_tips_for_photographing_darker_skin_tones.html

There are an assload of sensors out there that can handle suboptimal lighting. All rgb-d sensors have infrared and they are cheap

What's the razor for laziness or being satisfied with the status quo instead of stupidity
---

@_@
---

specialkid8 posted...

^I've heard this before but why doesn't AI recognize darker people? People say "we need to fix this" but no one seems to say what the actual problem is. Is this something to do with the actual input of the developers or is it just a visual thing the machines have a problem with?

Not enough data. That's solely it. The developers have to collect images to create the dataset to be used to train the models and you need A LOT, gigs worth. You can imagine that the demographics of the developers are primarily Asian, Indian, or some Whites. So they're going to prioritize collecting data that fits their environment or what they want to classify and they're going to make sure the model works well enough for them before anything else. They can worry about having equal representation later.
---

MacadamianNut3 posted...

What's the razor for laziness or being satisfied with the status quo instead of stupidity

Uh... being stupidly lazy.

Sad_Face posted...

Not enough data. That's solely it. The developers have to collect images to create the dataset to be used to train the models and you need A LOT, gigs worth. You can imagine that the demographics of the developers are primarily Asian, Indian, or some Whites. So they're going to prioritize collecting data that fits their environment or what they want to classify and they're going to make sure the model works well enough for them before anything else. They can worry about having equal representation later.

This just seems like a big assumption. Is "oops, we forgot to put black people in the computer" really that big of a problem?

specialkid8 posted...

Is "oops, we forgot to put black people in the computer" really that big of a problem?

Uh yeah it would be a huge problem if this technology is used for anything serious and not memey snapchat filters
---

https://www.businessinsider.com/viral-ai-selfie-classifier-imagenet-roulette-part-of-bias-project-2019-9

Apparently this was intentional though to raise awareness of the problem
---

MacadamianNut3 posted...

https://www.businessinsider.com/viral-ai-selfie-classifier-imagenet-roulette-part-of-bias-project-2019-9

Apparently this was intentional though to raise awareness of the problem

Oh, so it really is malice instead of stupidity, huh?

specialkid8 posted...

This just seems like a big assumption. Is "oops, we forgot to put black people in the computer" really that big of a problem?

I can assure you this is no assumption. Study Machine Learning yourself and you'll see how insanely well developed and these tools are, freely available to boot, and how you can train your own datasets. It's not "oops we forgot to put black people in the computer", it's "alright, we just compiled a million+ images and spent the last couple months finally getting this thing to bloody work to recognize our faces. Oh, what's that? We should do black folks too? We'll get around to them... later"
---

how many big brains in here have taken a CV course? Graduate level?
---

IllegalAlien posted...

how many big brains in here have taken a CV course? Graduate level?

I have a PhD in computer science, spent the first two years doing biometrics which is literally just computer vision and machine learning, and I'm dealing with those two and robotics right now for my job

What's the point
---

IllegalAlien posted...

how many big brains in here have taken a CV course? Graduate level?

I did my master's thesis on CV and published it after, but now I'm in another topic. Might move back to some CV applications tho,
---

Nice CV is great, Viola-Jones also had this bias problem IIRC
---

did the AI study at 4chan?
---

For the record though, usually CV algorithms run on black and white images and there is certainly a legitimate argument to be made that darker faces present more of a challenge.

...but certainly not so much as to result in the issues talked about here. There's definitely a bias in training data, and no one serious is really denying that tbh
---

MacadamianNut3 posted...

What's the razor for laziness

C'mon, guys, let's get Ivynn's Razor trending.
---

scar the 1 posted...

For the record though, usually CV algorithms run on black and white images and there is certainly a legitimate argument to be made that darker faces present more of a challenge.

...but certainly not so much as to result in the issues talked about here. There's definitely a bias in training data, and no one serious is really denying that tbh

Yeah I'm wondering what kind of equipment they used to collect the data. Even we had success with iris biometrics which were also black and white after being converted, and that worked for people with light eyes where anyone could clearly see the texture from a couple of feet, and people with dark eyes where you would have to shine a light in their eyes to see their texture close up

I haven't seen that many midnight black people wandering around so I'm wondering how this is still a big problem in face recognition in 2019 other than intentional laziness
---

I'm assuming they're using top-k classification model w/ CNN. So, obviously, the labels are in the dataset and are biased towards classifications based on race for black people. Whereas other non-racial labels, such as doctor for example, would be labeled more often for other races.
---

is "blackamoor" even a word?
---

MacadamianNut3 posted...

I haven't seen that many midnight black people wandering around so I'm wondering how this is still a big problem in face recognition in 2019 other than intentional laziness

Idk it makes sense for descriptors for dark complexion to be different since the gradients will likely be weaker, so if the dataset is shit then yeah. And if there's anything I know about CV, it's that datasets are shit
---

ROBANN_88 posted...

is "blackamoor" even a word?

maybe 800 years ago.