Sampling bias: how a machine-learning beauty contest awarded nearly all prizes to whites


#1

Originally published at: http://boingboing.net/2016/09/06/sampling-bias-how-a-machine-l.html


#2

Neat phenomenon, but I don’t understand the outrage.

In other words, they used non-representative samples to make statistical inferences.

Only if you choose to misinterpret those inferences. The samples were 100% representative of the self-selected contributor community, as were the inferences. Nothing else was claimed.


#3

With this and Automated Sentencing, when the Machine Revolution comes, first to the wall will be blonde women and black men.


#4

Did I miss something? It sounds like the ‘winners’ (per racial category) were actually roughly in proportion to their representation among the entrants. If 75% of the contenders were white Europeans (no word on how many whites of other nations) then 33/44 would be an average representation among the winners and it sounds like there were probably other white (non-European) entrants; the breakdown in the article being incomplete.


#5

These types of articles all trip over the same thing. They pretend that there isn’t a more or less global standard of beauty. There is one of course. The best representation of that standard is the actress Kate Hudson ten years ago. Show anybody on the planet a picture of Kate Hudson ten years ago and any other female that looks markedly different from Kate Hudson ten years ago and ask them to choose the most beautiful. Kate Hudson ten years ago will win the majority of the time.


#6

I don’t know who Kate Hudson is and I don’t want to add the phrase “Kate Hudson ten years ago” to my search history, but this made me laugh.


#7

To the people “so what”-ing – this is a variation on garbage in, garbage out. Not the contestants’ actual faces, but the evaluation inputs. It’s going to become more and more of a problem as “big data” is used to determine who’s allowed, for instance, to get a mortgage. Syllogisms based on the wrong metrics are still wrong if a machine makes them rather than a human.

The article’s right to point out that there are a lot of people who will believe something is true because math and science were involved. 19th century racists used math and stats to “prove” all sorts of things. They were still wrong.


#8

Yes! If these exercises reveal anything at all, it’s the biases in their inputs. Deep Dreaming sees dog faces everywhere because it was fed dog faces, not because dog faces are somehow special.


#9

The combination of this new oracular authority and the dismal, familiar sampling bias is a truly dangerous thing.

Not so new. I can remember reading a paper from the 1950s about the “cult of data processing.” It’s kind of the Revenge of the Nerds.


#10

Crap in, crap out. One of the fundamentals of software development. Feed your machine learning AI with bad data, you’ll get a bad black box.


#11

If you had said Alyson Hannigan any time in the last ten years, I’d have listened.


#12

wait…so 75% of entrants were white europeans, presumably some where white non-europeans (im going to guess 5% just from pulling it out of my ass), which makes 80% of entrants were white. and 84% of winners were white, which is roughly in line with the number of entrants. So…what the hell? It seems like this is a story about absolutely nothing; how did this become a story about a racist AI and bad data? The result is absolutely what one would expect.


#13

not because dog faces are somehow special.

Okay, hold on, let’s not be hasty.


#14

Regardless of actual results, it seems like using a machine to assign an “objective”, one-dimensional numeric score for a measure that is multivariate, entirely human and subjective is Doin’ It Rong.

I worry that the problem with AI is that we don’t really know what intelligence is, so we won’t know it when we see it.


#15

Here is the leaderboard.

It looks like a lot of the white winners don’t fit the beauty ideal that would get them chosen in a beauty contest with human judges. Not that they’re unattractive, they just aren’t the sort of faces I’d expect to see as finalists. I also notice that three of the men broke the “no beards” rule, while another hasn’t shaved recently. Quite a few of the women have ignored the “no makeup” rule too.

In any case, this represents an improvement in algorithmic beauty contests, since 100 years ago there was a sample of one as the ideal of female beauty (and actual women didn’t match up to it):

In October 1922, a particularly contentious competition took place at the Physical Culture Show and Beauty Contest held at Madison Square Garden. At this event, according to the Pittsburgh Press, five male judges, all sculptors, “led the young women contestants into a private room in the Garden and minutely inspected the competitors one by one.”

The women were naked during these inspections, “wearing not even so much as a pair of slippers on their feet.” One of the contestants, Ann Hyatt, told the Press that she was instructed to remove her bathrobe and then bathing suit. When she murmured that the situation was “very embarrassing,” Herman Moens, the head judge, remarked, “But the body is far more beautiful nude.” Then, according to Hyatt, “[t]he other four repeated in a kind of chorus, ‘Yes, indeed. By all means. It certainly is.’”

After every contestant had been subjected to a thorough visual inspection, 18-year-old Dorothy Knapp (five-foot-three; 35.5-25-35) was declared the winner. This devastated Hyatt (five-foot-four; 34.75-28.5-36.5), who had been determined to claim the Miss American Venus title. Refusing to accept the judges’ decision, Hyatt sought the services of a lawyer, who announced his client’s plan to take the case to the Supreme Court of the State of New York. “Miss Hyatt is really, in fact, the most perfectly formed woman in America, the modern Venus of the United States,” her lawyer, David Steinhardt, said. “[I]t is a matter of serious importance to herself, to her husband, and to her children that she should not be defrauded of this conspicuous distinction.”

Three months after the “new Venus” proportions were announced, Dr. R. Tait McKenzie, sculptor and former director of physical education of McGill University, told a crowd of female college students that he had given up on looking for ideal figures among America’s young women. Having studied hundreds of measurement cards from colleges in the country’s east, he declared that the women were all “slab-chested and knock-kneed”—the “antitheses of Venus,” according to the Oakland Tribune.


#16

Pretty fuckin rude for a Canadian!

Maybe he meant the Venus of Willendorf. “These chicks aren’t morbidly obese, and none of them have beehives for heads!”


#17


#18


#19

This topic was automatically closed after 5 days. New replies are no longer allowed.