From the "What could possibly go wrong" files ...
https://www.technologyreview.com/s/602955/neural-network-learns-to-identify-criminals-by-their-faces/?utm_campaign=socialflow&utm_source=facebook&utm_medium=post
QuoteNeural Network Learns to Identify Criminals by Their Faces
The effort aimed at identifying criminals from their mugshots raises serious ethical issues about how we should use artificial intelligence.
Soon after the invention of photography, a few criminologists began to notice patterns in mugshots they took of criminals. Offenders, they said, had particular facial features that allowed them to be identified as law breakers.
One of the most influential voices in this debate was Cesare Lombroso, an Italian criminologist, who believed that criminals were "throwbacks" more closely related to apes than law-abiding citizens. He was convinced he could identify them by ape-like features such as a sloping forehead, unusually sized ears and various asymmetries of the face and long arms. Indeed, he measured many subjects in an effort to prove his view although he did not analyze his data statistically.
This shortcoming eventually led to his downfall. Lombroso's views were discredited by the English criminologist Charles Goring, who statistically analyzed the data relating to physical abnormalities in criminals versus noncriminals. He concluded that there was no statistical difference.
And there the debate rested until 2011, when a group of psychologists from Cornell University showed that people were actually quite good at distinguishing criminals from noncriminals just by looking at photos of them. How could that be if there are no statistically different features?
Today, we get an answer of sorts, thanks to the work of Xiaolin Wu and Xi Zhang from Shanghai Jiao Tong University in China. These guys have used a variety of machine-vision algorithms to study faces of criminals and noncriminals and then tested it to find out whether it could tell the difference.
Their method is straightforward. They take ID photos of 1856 Chinese men between the ages of 18 and 55 with no facial hair. Half of these men were criminals.
They then used 90 percent of these images to train a convolutional neural network to recognize the difference and then tested the neural net on the remaining 10 percent of the images.
The results are unsettling. Xiaolin and Xi found that the neural network could correctly identify criminals and noncriminals with an accuracy of 89.5 percent. "These highly consistent results are evidences for the validity of automated face-induced inference on criminality, despite the historical controversy surrounding the topic," they say.
Xiaolin and Xi say there are three facial features that the neural network uses to make its classification. These are: the curvature of upper lip which is on average 23 percent larger for criminals than for noncriminals; the distance between two inner corners of the eyes, which is 6 percent shorter; and the angle between two lines drawn from the tip of the nose to the corners of the mouth, which is 20 percent smaller.
They go on to plot the variance in the data from criminal and noncriminal faces in a simplified parameter space called a manifold. And this process reveals why the difference has been hard to pin down.
Xiaolin and Xi show that these datasets are concentric but that the data for criminal faces has much greater variance. "In other words, the faces of general law-biding public have a greater degree of resemblance compared with the faces of criminals, or criminals have a higher degree of dissimilarity in facial appearance than normal people," say Xiaolin and Xi.
This may also explain why certain kinds of statistical test cannot distinguish between these data sets. Indeed, Xiaolin and Xi show that when they combine criminal and noncriminal faces to create "average" faces, they look almost identical.
Although controversial, that result is not entirely unexpected. If humans can spot criminals by looking at their faces, as psychologists found in 2011, it should come as no surprise that machines can do it, too.
The worry, of course, is how humans might use these machines. It's not hard to imagine how this process could be applied to data sets of, say, passport or driving license photos for an entire country. It would then be possible to pick out those people identified as law-breakers, whether or not they had committed a crime.
That's a kind of Minority Report scenario in which law-breakers could be identified before they had committed a crime.
Of course, this work needs to be set on a much stronger footing. It needs to be reproduced with different ages, sexes, ethnicities, and so on. And on much larger data sets. That should help to tease apart the complexities of the findings. For example, Xiaolin and Xi find that criminal faces can be sub-divided into four subgroups, but noncriminal faces into only three. How come? And how does this vary in other groups?
And the work raises important questions. If the result does hold water, how is it to be explained? Why would the faces of criminals have much greater variance than those of noncriminals? And how are we able to spot these faces—is it learned behavior or hard-wired behavior that has evolved?
All this heralds a new era of anthropometry, criminal or otherwise. Last week, researchers revealed how they had trained a deep-learning machine to judge in the same way as humans whether somebody was trustworthy by looking at a snapshot of their face. This work is another take on the same topic. And there is room for much more research as machines become more capable. Examining what our clothes or hair say about us is one obvious angle. And machines will soon be able to study movement, too. That raises the possibility of studying how we move, how we interact, and so on.
Ref: arxiv.org/abs/1611.04135: Automated Inference on Criminality Using Face Images
It couldn't be used to catch Hitler.
Next up: craniometry apps.
Quote from: Tamas on November 24, 2016, 11:40:01 AM
He is dead, duh.
That's what they wanted you to believe. If it is on History Channel, it must be History. Would they lie?
You know, reading the title very quickly, I read totally something else. :sleep:
Sounds like Phrenology 2.0.
What is phrenology? Is it saying that your face is the cause effect behind your behavior, or just saying that you can tell something about a person by just studying their face?
I'd like to see which pictures they used
Quote from: LaCroix on November 24, 2016, 01:14:09 PM
I'd like to see which pictures they used
Some are here: https://arxiv.org/pdf/1611.04135v2.pdf
Quote from: Syt on November 24, 2016, 01:15:50 PM
Quote from: LaCroix on November 24, 2016, 01:14:09 PM
I'd like to see which pictures they used
(https://languish.org/forums/proxy.php?request=http%3A%2F%2Fwww.sighedeffects.com%2Fwp-content%2Fuploads%2F2014%2F10%2Fphreno01.jpg&hash=2e3de9e62ecb7f63e0300be9d9aa3f7112718fed)
FYP
These features sound similar to the ones caused by higher testosterone.
Quote from: MadImmortalMan on November 24, 2016, 04:18:48 PM
These features sound similar to the ones caused by higher testosterone.
Sure. In cats.
Quote from: DGuller on November 24, 2016, 01:07:40 PM
What is phrenology? Is it saying that your face is the cause effect behind your behavior, or just saying that you can tell something about a person by just studying their face?
The latter, IIRC, about the shape of people's heads.
In China murdering millions isn't a crime, while trying to introduce freedom is, so it's unclear how helpful this is in a real country.
Quote from: The Larch on November 24, 2016, 04:23:13 PM
Quote from: DGuller on November 24, 2016, 01:07:40 PM
What is phrenology? Is it saying that your face is the cause effect behind your behavior, or just saying that you can tell something about a person by just studying their face?
The latter, IIRC, about the shape of people's heads.
Then it's a bit of a stretch to imply that any attempt to accomplish such a thing is pseudoscience. Other factors may affect both the appearance of your face and your criminal proclivities, or the appearance of your face may cause you to be suspected and caught in the first place.
Quote from: DGuller on November 24, 2016, 04:43:15 PM
Then it's a bit of a stretch to imply that any attempt to accomplish such a thing is pseudoscience. Other factors may affect both the appearance of your face and your criminal proclivities, or the appearance of your face may cause you to be suspected and caught in the first place.
I wonder how they sourced the pictures...
Quote from: Jacob on November 24, 2016, 05:00:09 PM
Quote from: DGuller on November 24, 2016, 04:43:15 PM
Then it's a bit of a stretch to imply that any attempt to accomplish such a thing is pseudoscience. Other factors may affect both the appearance of your face and your criminal proclivities, or the appearance of your face may cause you to be suspected and caught in the first place.
I wonder how they sourced the pictures...
I read the paper quickly, they described it. There are a few places in that paper that made me a little skeptical, but by machine learning academic paper standards, this appears to me to be a pretty good one.
They got the pictures of criminals from police, and the pictures of non-criminals from trawling the Internet. They made sure that nothing about the actual picture-taking methodology remained in the data, and that the pictures did not contain any traces that could've tipped off the algorithm.
Quote from: DGuller on November 24, 2016, 05:17:16 PM
I read the paper quickly, they described it. There are a few places in that paper that made me a little skeptical, but by machine learning academic paper standards, this appears to me to be a pretty good one.
They got the pictures of criminals from police, and the pictures of non-criminals from trawling the Internet. They made sure that nothing about the actual picture-taking methodology remained in the data, and that the pictures did not contain any traces that could've tipped off the algorithm.
EDIT: never mind.
Quote from: CountDeMoney on November 24, 2016, 01:40:56 PM
Quote from: Syt on November 24, 2016, 01:15:50 PM
Quote from: LaCroix on November 24, 2016, 01:14:09 PM
I'd like to see which pictures they used
(https://languish.org/forums/proxy.php?request=http%3A%2F%2Fwww.sighedeffects.com%2Fwp-content%2Fuploads%2F2014%2F10%2Fphreno01.jpg&hash=2e3de9e62ecb7f63e0300be9d9aa3f7112718fed)
FYP
I thought this was what they actually used until I read the text :lol:
re: pictures, so the program can tell the difference between a sad face and a happy face :hmm:
Ugh. Eugenics rears its ugly head, yet again. The problem with this is that it's a false correlation. Criminal status is also correlated with social status, which is also correlated with physical appearance. This isn't scanning for criminals, it's scanning for people likely to be penalized through the criminal justice system- useful in its own way, but knowing how social status decreases likelihood or severity of penalty, its usefulness for crime prevention is dubious, to say the least.
The correlation seems real enough. :hmm:
Quote from: DontSayBanana on November 24, 2016, 11:38:36 PM
Ugh. Eugenics rears its ugly head, yet again. The problem with this is that it's a false correlation. Criminal status is also correlated with social status, which is also correlated with physical appearance. This isn't scanning for criminals, it's scanning for people likely to be penalized through the criminal justice system- useful in its own way, but knowing how social status decreases likelihood or severity of penalty, its usefulness for crime prevention is dubious, to say the least.
Social status correlates with the curve of your upper lip? :huh:
There are plenty of meaningless but real correlations, a shortage of bananas in the UK correlates with us being at war with Germany for example.
Quote from: MadImmortalMan on November 24, 2016, 04:18:48 PM
These features sound similar to the ones caused by higher testosterone.
There's one very easy indicator for high testosterone exposure in the womb, and that's finger length. If the middle finger is noticeably longer than the index, it means that mand had a high exposure to testosterone, if the fingers are roughly the same length it means he had a lower exposure. A higher exposure to testosterone has been linked to things ranging from athletic prowess (the higher your testosterone, the more likely you are to be competent in sports and physical skills) to predisposition to violent crime (the more testosterone, the likelier you are to commit it). By itself it doesn't mean anything, it's what happens in life what matters, in this case nurture beats nature. That's why things like the OP have the potential to be Orwellian nightmares, your features don't make you a criminal, your actions do.
Quote from: The Larch on November 25, 2016, 05:17:59 AM
Quote from: MadImmortalMan on November 24, 2016, 04:18:48 PM
These features sound similar to the ones caused by higher testosterone.
There's one very easy indicator for high testosterone exposure in the womb, and that's finger length. If the middle finger is noticeably longer than the index, it means that mand had a high exposure to testosterone, if the fingers are roughly the same length it means he had a lower exposure. A higher exposure to testosterone has been linked to things ranging from athletic prowess (the higher your testosterone, the more likely you are to be competent in sports and physical skills) to predisposition to violent crime (the more testosterone, the likelier you are to commit it). By itself it doesn't mean anything, it's what happens in life what matters, in this case nurture beats nature. That's why things like the OP have the potential to be Orwellian nightmares, your features don't make you a criminal, your actions do.
Damn, I want to know what happened then. Why aren't I a b-ball phenom?
Wonder if this neural network can recognize my snail drawing :angry:
Quote from: DGuller on November 24, 2016, 11:44:51 PM
The correlation seems real enough. :hmm:
But the correlation seems to be between pictures obtained from the police vice pictures selected online. That pictures obtained from the police are notably more likely to be pictures of criminals isn't exactly a startling breakthrough.
Quote from: grumbler on November 26, 2016, 09:03:14 AM
Quote from: DGuller on November 24, 2016, 11:44:51 PM
The correlation seems real enough. :hmm:
But the correlation seems to be between pictures obtained from the police vice pictures selected online. That pictures obtained from the police are notably more likely to be pictures of criminals isn't exactly a startling breakthrough.
Does it? The paper seems to go to great length describing the process by which they normalized the pictures so that backgrounds, lighting, equipment signatures, etc. are no longer a factor. Also, while the pictures are obtained from police, those weren't the pictures taken by police, they were just normal pictures of people known to be ciminals to police. Do you have reason to believe that their methods to normalize the pictures were insufficient?
Quote from: DGuller on November 26, 2016, 09:52:45 AM
Does it? The paper seems to go to great length describing the process by which they normalized the pictures so that backgrounds, lighting, equipment signatures, etc. are no longer a factor. Also, while the pictures are obtained from police, those weren't the pictures taken by police, they were just normal pictures of people known to be ciminals to police. Do you have reason to believe that their methods to normalize the pictures were insufficient?
If the photos of the criminals were taken under the same conditions as those taken of the non-criminals, then they are comparable. I don't know why you would get any pictures from the police in a study like this; get the names of criminals, and then find their pictures online, just as you find the non-criminals. Then you have a valid sample for comparison.
If you get some pictures from the police, and others not from the police, and you discover that there is a statistically significant variance between the two sets of pictures, you still cannot attribute the variance to criminality vice origin of the pictures, no matter how much you normalize for background or whatever.
Quote from: Admiral Yi on November 25, 2016, 12:12:06 AM
Social status correlates with the curve of your upper lip? :huh:
Aristos keep theirs stiff.
Quote from: DGuller on November 24, 2016, 11:44:51 PM
The correlation seems real enough. :hmm:
Sure. Between convicts and social status. It only holds if you're naive enough to assume that everyone that commits a crime is convicted, and that social status plays no part in conviction rates.
Quote from: DontSayBanana on November 27, 2016, 10:14:06 AM
Quote from: DGuller on November 24, 2016, 11:44:51 PM
The correlation seems real enough. :hmm:
Sure. Between convicts and social status. It only holds if you're naive enough to assume that everyone that commits a crime is convicted, and that social status plays no part in conviction rates.
I was just pointing out that, as usual, you were saying things authoritatively that were completely false. Correlation either exists or doesn't, interpretation doesn't enter into it.
Quote from: DGuller on November 27, 2016, 10:29:20 AM
I was just pointing out that, as usual, you were saying things authoritatively that were completely false. Correlation either exists or doesn't, interpretation doesn't enter into it.
What things did I say "authoritatively that were completely false?" Content validity is that it measures what it's supposed to measure- since not all criminals are convicts (additionally, not all convicts are criminals, but that's a completely different rant), saying that this measures criminality because it identifies convicts is demonstrably false.
Quote from: DontSayBanana on November 27, 2016, 11:19:06 AM
What things did I say "authoritatively that were completely false?"
QuoteThe problem with this is that it's a false correlation.
Unless the researchers doctored the study or were grossly negligent in the technical details of their study, the correlation they found is real. Just like the correlation of the number of pot holes and flu cases is real.
Quote from: DGuller on November 27, 2016, 11:30:19 AM
Unless the researchers doctored the study or were grossly negligent in the technical details of their study, the correlation they found is real. Just like the correlation of the number of pot holes and flu cases is real.
No, the correlation between these photos and felon status is real; the correlation between these photos and criminality is not- it's a claim made with mountains of evidence to the contrary already sitting there.
Quote from: DontSayBanana on November 27, 2016, 11:34:05 AM
Quote from: DGuller on November 27, 2016, 11:30:19 AM
Unless the researchers doctored the study or were grossly negligent in the technical details of their study, the correlation they found is real. Just like the correlation of the number of pot holes and flu cases is real.
No, the correlation between these photos and felon status is real; the correlation between these photos and criminality is not- it's a claim made with mountains of evidence to the contrary already sitting there.
That's a pretty extraordinary claim to make. The legal system has to be perfectly random in its actions for this claim to actually be true. I think even in Saddam Hussein's Iraq there was some correlation between conviction status and criminality.
Quote from: derspiess on November 25, 2016, 12:09:25 PM
Wonder if this neural network can recognize my snail drawing :angry:
lol, stupid internet tests.
And in a completely unrelated development, DGuller is being more of a Russian pseudoscience knucklehead than usual. Bet you guys didn't see that correlation coming.
Quote from: DGuller on November 27, 2016, 11:39:46 AM
That's a pretty extraordinary claim to make. The legal system has to be perfectly random in its actions for this claim to actually be true. I think even in Saddam Hussein's Iraq there was some correlation between conviction status and criminality.
Not at all. The legal system just has to be shown not to have a direct correlation, so here are some counter-examples that law is applied unevenly among different social statuses:
- Pre-trial intervention, expungement, and "second chance" programs that heavily weigh the candidate's social status to determine their likely future social contribution.
- Income (which is correlated to social status) and quality of legal representation. Rule of law means accepting the loopholes as well as accepting the law- law firms with offices on three continents and hundreds of paralegals are just more likely to find them than an underfunded public defender's office.
- Juror bias - jurors will feel more responsibility and pressure to dilute punishment towards a socially similar defendant than one that's dissimilar- they won't hold back on outcasts.
All three of these effects are empirically provable, and all three chip away at the correlation between criminality and felon status.
Quote from: DontSayBanana on November 27, 2016, 12:23:34 PM
Quote from: DGuller on November 27, 2016, 11:39:46 AM
That's a pretty extraordinary claim to make. The legal system has to be perfectly random in its actions for this claim to actually be true. I think even in Saddam Hussein's Iraq there was some correlation between conviction status and criminality.
Not at all. The legal system just has to be shown not to have a direct correlation, so here are some counter-examples that law is applied unevenly among different social statuses:
- Pre-trial intervention, expungement, and "second chance" programs that heavily weigh the candidate's social status to determine their likely future social contribution.
- Income (which is correlated to social status) and quality of legal representation. Rule of law means accepting the loopholes as well as accepting the law- law firms with offices on three continents and hundreds of paralegals are just more likely to find them than an underfunded public defender's office.
- Juror bias - jurors will feel more responsibility and pressure to dilute punishment towards a socially similar defendant than one that's dissimilar- they won't hold back on outcasts.
All three of these effects are empirically provable, and all three chip away at the correlation between criminality and felon status.
:unsure:
Quote from: DontSayBanana on November 27, 2016, 12:23:34 PM
All three of these effects are empirically provable, and all three chip away at the correlation between criminality and felon status.
Now you're saying something substantially different. There is a big difference between something chipping away at the correlation and something completely removing it. The argument you made in the previous post is only valid if the correlation is completely removed. Correlation is not algebra, counter-examples don't invalidate it, only weaken it.
Quote from: DontSayBanana on November 27, 2016, 12:23:34 PM
All three of these effects are empirically provable, and all three chip away at the correlation between criminality and felon status.
You know what else chips away at the correlation between criminality and felon status? Pleading down to a misdemeanor.
Quote from: DGuller on November 27, 2016, 12:35:12 PM
Quote from: DontSayBanana on November 27, 2016, 12:23:34 PM
All three of these effects are empirically provable, and all three chip away at the correlation between criminality and felon status.
Now you're saying something substantially different. There is a big difference between something chipping away at the correlation and something completely removing it. The argument you made in the previous post is only valid if the correlation is completely removed. Correlation is not algebra, counter-examples don't invalidate it, only weaken it.
Are you joking? Between uncaught criminals and criminals treated differently by the system, I can't imagine there being an R value of more than 0.55 or so. Maybe, just maybe if the system was rock solid before, 0.6, and I wouldn't be putting any crime prevention programs into place with under a 0.75. Hell, I wouldn't want less than a 0.8.
Quote from: DontSayBanana on November 27, 2016, 02:22:26 PM
Quote from: DGuller on November 27, 2016, 12:35:12 PM
Quote from: DontSayBanana on November 27, 2016, 12:23:34 PM
All three of these effects are empirically provable, and all three chip away at the correlation between criminality and felon status.
Now you're saying something substantially different. There is a big difference between something chipping away at the correlation and something completely removing it. The argument you made in the previous post is only valid if the correlation is completely removed. Correlation is not algebra, counter-examples don't invalidate it, only weaken it.
Are you joking? Between uncaught criminals and criminals treated differently by the system, I can't imagine there being an R value of more than 0.55 or so. Maybe, just maybe if the system was rock solid before, 0.6, and I wouldn't be putting any crime prevention programs into place with under a 0.75. Hell, I wouldn't want less than a 0.8.
Let's assume for the moment that your numbers are correct. Have you noticed how all of these numbers appear to be significantly greater than zero?
Quote from: DontSayBanana on November 27, 2016, 02:22:26 PM
Are you joking? Between uncaught criminals and criminals treated differently by the system, I can't imagine there being an R value of more than 0.55 or so. Maybe, just maybe if the system was rock solid before, 0.6, and I wouldn't be putting any crime prevention programs into place with under a 0.75. Hell, I wouldn't want less than a 0.8.
Completely removing it would mean a correlation of zero.
Way to keep moving the goalposts. I said this is not a useful predictor, and the threshold for something to be predictive is way higher than 0.1, 0.2, 0.5, none of those are strong enough correlations to be predictive.
Quote from: DontSayBanana on November 27, 2016, 02:44:36 PM
Way to keep moving the goalposts.
You're the one moving the goalposts. When you say that something isn't correlated, you mean it has an R of less than 0.8. Okay...
Quote
I said this is not a useful predictor, and the threshold for something to be predictive is way higher than 0.1, 0.2, 0.5, none of those are strong enough correlations to be predictive.
The threshold for a correlation to be predictive is >0.
Quote from: DGuller on November 27, 2016, 02:48:16 PM
The threshold for a correlation to be predictive is >0.
Not for any application where I've actually had to use R-squared. And sure as hell not for anything with possible civil rights implications.
For everything I've had to deal with involving peer review, 0.6 has been the absolute bare minimum to suggest a causal link. 0.5 is "it could be related, but it could also not"- high enough to warrant further study, but not high enough to reliably suggest a causal link.
Quote from: DontSayBanana on November 27, 2016, 08:36:35 PM
Quote from: DGuller on November 27, 2016, 02:48:16 PM
The threshold for a correlation to be predictive is >0.
Not for any application where I've actually had to use R-squared. And sure as hell not for anything with possible civil rights implications.
For everything I've had to deal with involving peer review, 0.6 has been the absolute bare minimum to suggest a causal link. 0.5 is "it could be related, but it could also not"- high enough to warrant further study, but not high enough to reliably suggest a causal link.
The threshold is the same for any predictive application. If it's significantly different from zero, it's predictive, causal or not makes no difference. What you may want to do with that predictive relationship is a different issue.
(https://greatcaesersghost.files.wordpress.com/2012/03/nerd.jpg)