AI can now tell if you're a criminal or not Through machine learning, researchers have repeated the historic criminology experiment of telling criminals apart from law-abiding people using facial recognition. Physiognomy, the ability to judge a person’s character from appearance alone, has been around since ancient Greece and was widely accepted by philosophers. …
No, just the return of "the data says what we want it to say, after we sculpt it into shape". Oh dear, let's put these two guys into a trial set of 'scientists' ranking them by 'stupidity' and see what happens when the result is "Chinese are more likely to be stupid scientists". I think a few of their colleagues might have a bone to pick or two. Skulls indeed.
But “Unlike a human examiner/judge, a computer vision algorithm or classifier has absolutely no subjective baggages, having no emotions, no biases whatsoever due to past experience, race, religion, political doctrine, gender, age, etc, no mental fatigue, no preconditioning of a bad sleep or meal,”.
Just curious who decides inputs, weights and bias nodes.
Phrenology was just skull measurement. You are looking at Lombroso's "Criminal atavism" in its worst form here.
Anyway nothing can surprise me in a world where someone who was thrown out of Congress hearing for a judge for being too racist for the 1980-es standards is appointed to lead the USA judiciary.
No; the return of those eternal researchers who don't understand Bayes's Theorem.
(On another forum I've just managed to upset a Jewish poster who is determined that Jews are a genetically distinct group by pointing out that if you used his supposed genetic markers to identify Jews, most of the positives would be false. Even citing to him an actual, peer reviewed article posted on nih.gov didn't work. People really do cling to weak statistical associations.)
supposed genetic markers to identify Jews
You could not have chosen a more difficult case to argue. "Jewishness" is traditionally conveyed solely by the mother (if your father is a Jew and your mother is not you are not a jew as per their religious canon - you become a gentile).
As a result the "main" DNA in Jews is pretty much equivalent to the rest of the population. There is little or no difference in genetic markers between _ALL_ jews and the rest of the Caucasian population. That is no longer valid for specific Jewish groups by the way - there are some with significantly higher frequency of some recessives and different genetic makeup than the overall population. The overall "Jewish" however is not distinguishable from overall Caucasian as per that NIH paper (and many others).
This was the situation until someone finally analysed mitochondrial DNA (which is exclusively from the mother - father never provides any of it). The end results were startling - the mitochondrial DNA diversity in Jews is practically NIL - they all trace down to 4 distinct women somewhere in the Middle East ~ 2000 bc (estimates based on genetic distance). If memory serves me right, that paper is in Cell by the way. A few years back, forgot exactly which issue.
In any case - trying to argue this with a numpty in a political forum is described in the Old Testament: you are throwing pearls to pigs.
"The end results were startling - the mitochondrial DNA diversity in Jews is practically NIL - they all trace down to 4 distinct women somewhere in the Middle East ~ 2000 bc (estimates based on genetic distance). If memory serves me right,"
Your memory doesn't serve you right - research showed that about 40% of Ashkenazi Jews had that very narrow band of mitochondrial DNA. Sephardi and Mizrahi Jews had no such association.
And this is my point. Let's assume that you decided to use that mitochondrial evidence to identify Jews. For Ashkenazi alone, you would get 60% false negatives. For the Jewish population as a whole, you would get a very high percentage of false negatives. And it was false positives (or false negatives) that my post was about, since that is the importance of Bayes's Theorem.
This post has been deleted by its author
"On another forum I've just managed to upset a Jewish poster who is determined that Jews are a genetically distinct group by pointing out that if you used his supposed genetic markers to identify Jews"
I went to school in north Norfolk with a girl (born locally) who converted to Judaism. Fairly sure her DNA remained unchanged.
>I went to school in north Norfolk with a girl (born locally) who converted to Judaism. Fairly sure her DNA remained unchanged.
There was a human-interest story on Radio 4 earlier in the year about a British woman who wanted to convert to Judaism. Her conversion was recognised by the appropriate bodies in Israel, but not by those in the U.K.
I'm not sure that says anything about Judaism other than a group of people spread across dozens of countries for hundreds of years entertain a variety of views about things, whodafunkit.
Previous studies in the area had datasets in excess of 100000 - the whole "bio-measurement" (as it is incorrect to call it biometrics) database from criminal identification using the Lambroso method has been fed into statistical analysis a gazillion times.
Each and every time the idea that "this persons shape equates to increased probability of criminality" has failed to pass more detailed statistical analysis.
Perhaps we might want to more carefully rephrase 'half of whom were convicted criminals' into 'half of whom had been convicted of a crime' - since the facial features in question might simply be attracting unwanted attention from biases within the police, whom then attempt to (and succeed) in convicting that facial type more often, regardless of the degree of "proof".
... And it's actually been tested somewhere. There was a small excercise with people who were facing court, where they gave unfortunate-looking people plastic surgery before their court case, which, as expected, gave them a lower conviction rate.
But, (and this was the interesting bit), did not reduce the rate at which previously-charged people were re-arrested. Being good-looking doesn't make criminals less criminal: it just makes them less likely to be convicted.
>Being good-looking doesn't make criminals less criminal: it just makes them less likely to be convicted.
Yes and no. I take your point, but all things being equal, good-looking people have less motivation to commit crime. My reasoning is based on all the studies that suggest that good-looking people are more likely to be promoted at work, or attract more sexual partners. Therefore they can fulfil their needs without resorting to criminal behaviour. *
It's a bit like psychopaths - most aren't convicted criminals, because they can get all they want by manipulating people within the letter of law (if not the spirit), so they have no need to risk breaking any laws. As a result, most psychopaths are to be found in upper-middle management and not behind bars.
* There's a great episode of 30 Rock in which John Hamm's character is made to realise that people only think that he is competent at things (tennis, being a medical doctor, cooking, riding a motorcycle) because he is really, really good looking. He's 'in the bubble', which causes him to think that people are all just really nice and accommodating.
There is actually quite a lot of scope of sample bias here. The characteristics the article describes sound quite a lot like the facial type you see with Foetal Alcohol Spectrum Disorder, i.e. children whose mothers boozed heavily during pregnancy.
People with FASD are basically damaged in a lot of ways. Facial features are altered, and brain function is compromised. These people are more likely than the general population to be criminals, and there's a fairly good chance that police consciously or subconsciously recognise this facial type as a likely sort to check for criminal activity, hence these people are going to feature disproportionately in the database.
I somehow can't blame the dataset.. neither the size nor the content. However it's more likely the programming behind the AI. Computers only do what we tell them to do and if the programmers on this particular bit let their prejudices or pre-conceived notions take sway, then the results are already pre-determined.
Nope, it is not a single dataset. It's been constructed from two separate sources, one entirely composed of criminals and another composed of people of unknown criminal stature. The criminal photos were taken from wanted pictures and the non-criminal photos taken from the internet. The two sets are completely incomparable. There's also the issue that the authors can't actually state whether the people in the non-criminal group are or are not criminals. The classifier only learned to distinguish wanted photos from internet selfies.
Yeah, we've been building these classifiers for 20 years now and we keep having to teach the idiots how to use them properly (even the ones that have PhDs). Obvious problems:
Their dataset has a prior probability of criminality around 50%. That's way higher than normal and leads the system to think that criminality is common. Same problem with a lot of diagnostic medicine ANNs. They try to detect rare diseases with an equal handful of normal and diseased cases. They look great in the literature, but never get adopted, because they keep flagging up healthy people--they've been heavily biased to think that the problem exists.
Second problem is the data. Are the pictures random? I doubt it. They've started by just looking at Han Chinese. Then they picked pictures of non-criminals by browsing the web and picked pictures of criminals by scouring for wanted posters. Looking at their conclusion faces, I can easily classify criminals vs. non-criminals simply by noticing whether the person is smiling.
Third problem is feature selection. I'm sure the algorithm didn't automatically choose to look at facial features. A raw ANN will pick out tons of useless data like the color of the pixel at location (3,42). It won't automatically go looking for complex features like facial feature ratios. So, the authors picked out a bunch of features they thought might be relevant (neo-phrenology as previously noted) and discovered that some of them were more relevant than others
From this paper, I would conclude that Chinese people tend to post pictures of smiling people online and criminals tend to look unhappy in mugshots. Thus, it's easy to distinguish between a selfie and a mugshot.
Oh Sigh, Sigh!
You have a point, but I think you take it far too far.
"... teach the idiots how to use them properly (even the ones that have PhDs). Obvious problems:"
We see things here, between us, of markedly differing severity.
"Their dataset has a prior probability of criminality around 50%. That's way higher than normal and leads the system to think that criminality is common." And "Same problem with a lot of diagnostic medicine ANNs. They try to detect rare diseases with an equal handful of normal and diseased cases. They look great in the literature, but never get adopted, because they keep flagging up healthy people--they've been heavily biased to think that the problem exists."
I'm not sure at all that this is relevant, especially the first bit. Training on the examples is best done with near equal numbers of samples for each class: otherwise there is likely to be criticism on that very issue. Evaluation is, likewise, best done on datasets of near equal class size; and it's easier with equal-size evaluation sets.
For operational use: Bayesian statistics does indeed require weighting with the real-life class occurrence rates - this can be dealt with totally outside of class-specific modelling. This by use of the a priori knowledge of class occurrence statistics.
"Second problem is the data. Are the pictures random? I doubt it."
Read the paper, as linked. It is much better than you (think and) write, though it does have its deficiencies.
They've started by just looking at Han Chinese.
Looking within one racial characteristic (especially on such a small dataset) it actually sound science.
"Then they picked pictures of non-criminals by browsing the web and picked pictures of criminals by scouring for wanted posters."
No! Read the paper. All the photos are from non-criminal identification sources. I suspect this is from existing photos on ID cards or driving licences, or similar. Whilst this is not ideal, there is no bias in data-capture mechanism or in the likely 'happiness' of the subjects.
"Looking at their conclusion faces, I can easily classify criminals vs. non-criminals simply by noticing whether the person is smiling."
No you cannot: see above!
However, there is a problem with the demographics, particularly of the non-criminal dataset. There is a high preponderance of university-educated people. I suspect (only suspect) that this is derived from using current students/staff and their spouses or near-spouses. Note in the paper, the collared shirts of the non-criminals and the non-collared shirts of the criminals. Some clear demographic selection would have been useful here: most likely on employment status and earnings for the non-criminals; also on the type of crime for the criminals: violence against the person, violence against property, white-collar crimes - and so on.
"Third problem is feature selection. I'm sure the algorithm didn't automatically choose to look at facial features."
True, but so what?
"So, the authors picked out a bunch of features they thought might be relevant (neo-phrenology as previously noted) and discovered that some of them were more relevant than others."
Again, so what? Whether the individual or composite features are designed manually or automatically matters nothing, providing their training and the evaluation is unbiased (including lack of bias by repeated manual feedback).
"From this paper, I would conclude that Chinese people tend to post pictures of smiling people online and criminals tend to look unhappy in mugshots. Thus, it's easy to distinguish between a selfie and a mugshot."
See the paper and above: neither 'mugshots' (definitely) nor 'selfies' (it seems) are used. Thus neither data capture quality nor associated (mood/stylistic) effects are relevant deficiencies.
Best regards
What they've trained their AI to recognize is the physical symptoms shown in children of drug and alcohol abusers; symptoms of DNA damage. Doctors can recognize these features too. They can throw all the computing power in the world at this problem but the prediction accuracy isn't going to get any better. The correlation is limited and there's nothing more in the data to process. Does the Shanghai Jiao Tong University not have medical books that could have prevented this wasted research?
They have all the equipment to instantly identity potential criminals at the border and deny them entry.
Plus a database of all the ones they let in by mistake before this technology was available.
I would have found the study more convincing if it had been based on passport photos. All of mine make me look like a retarded criminal (ummmm.......) and a small sample from others are equally unflattering.
Compare two relatively small sample groups on a chosen criterion and you'll find statistical differences anyway, even without the more sophisticated criticisms above. You might find, say, that known criminals are more likely to prefer noodles to rice - so noodle eaters are potential crooks? Or that one group have 19% more chance of liking Man United than Man City or whatever.
I remember reading a comment on another story about AI. Short version: In the eighties the US fed high quality photos of their kit and grainy photos of Russian kit into a system. Pretty soon the error rate was near zero percent. Of course what it learnt was the difference between high quality and low quality photos, not the characteristics of different machine.
the beginning of a new era in plastic surgery. Late night tv - "Got a face like a criminal? Come see us." "Poor Johnny couldn't leave his house without fear of getting arrested what with his beady little eyes, and mouth like a hydra. After our in-home procedure, he now bears a striking resemblance to the late Marty Feldman, but he hasn't been arrested once. "
The law is arbitrary, changeable and often capricious. In country A(merica) I may not be a criminal because there is no law against (say) free speech but in country C(hina) I would be a criminal. Pretty sure my head shape doesn't change between the two countries.
Perhaps the head shape of the lawmakers is a better guide to the criminality of the people?
What exactly is the definition of "criminality"?
Does it exclude, say, rich individuals who manage to find ways round tax laws, but include drunks who damage a bit of property? Does it exclude wealthy politicians who paint slogans on the side of a bus about giving millions to the NHS, but include people who claim that they are going to share millions of GBP that are in a Nigerian bank account? Does it exclude Bingo hall and betting shop owners/managers who install machines that swallow addicted punters' money but include sellers of soft drugs and practitioners of the three card trick? And considering the provenance of this, does it include people who criticise the Chinese Govt?
If there have ever been any real studies into what makes people criminals, i.e. breaking the law
From my naivety I would presume any study would have to distinguish between those that commit crime because they do so or starve from those with an excess of resources who commit crime for reasons other than necessity.
Given that in the Western world the laws are very strict upon ownership and yet do not guaranty a minimum quality life for their citizens so as to prevent them falling into the "crime by necessity" bracket one wonders who actually have the most negative effect upon society.
The UK for example used to have pride in it's social safetynet but now we have vulnerable,homeless and starving citizens across the country in real need being ignored whilst we send charity abroad. The same charity money ends up back in the pockets of the UK people promoting the destruction of that safetynet along with the idea that the destitute were somehow to blame for their situation.
I do not need a machine to tell me who the real criminals are, their actions make them stand out clearly enough but I have no doubt that they will never be punished for their deeds until the law applies to ever citizen rather than just those the police are allowed to prosecute
Machines do however break down, sometimes in byzantine ways that causes error in judgement.
A machine that may have been struck by lightning in the past may not fail immediately, but will become more and more error prone before failing.
So how can you trust the output of the machine if you can't tell if the machine is failing or not?
The Register 
Biting the hand that feeds IT
Copyright. All rights reserved © 1998–2024
