SoylentNews is people
SoylentNews
Another Scientific Incarnation of Selective Correlation
When the 19th century was young, a Viennese physician Franz Joseph Gall got the ball rolling for the "science" of phrenology. (Not to be confused with phenology.) Phrenology believed that the shape and contour of a person's skull revealed their character, and thus could be used by employers and the criminal justice system to identify the lazy and the miscreants with simply a few quick measurements.
It also came in handy to justify slavery in the U.S., as depicted in Tarantino's Django Unchained.
Phrenology never went away, but went on to lurk in spin-offs such as eugenics. And if there were to be an updated incarnation of using a few quick body measurements to find the evil among us, it would have to employ sci/tech terms as "researchers", "algorithms" and "AI".
And so it does: Convict-spotting algorithm criticised
Researchers trained an algorithm using more than 1,500 photos of Chinese citizens, hundreds of them convicts.
They said the program was then able to correctly identify criminals in further photos 89% of the time. But the research, which has not been peer reviewed, has been criticised by criminology experts who say the AI may reflect bias in the justice system. "This article is not looking at people's behaviour, it is looking at criminal conviction..."
So, will AI ever get this god-like?
[Continues...]
AI Can Predict the Future Criminals Based on Facial Features
The bankrupt attempt to infer moral qualities from physiology was a popular pursuit for millennia, particularly among those who wanted to justify the supremacy of one racial group over another. But phrenology, which involved studying the cranium to determine someone's character and intelligence, was debunked around the time of the Industrial Revolution, and few outside of the pseudo-scientific fringe would still claim that the shape of your mouth or size of your eyelids might predict whether you'll become a rapist or thief.
Not so in the modern age of Artificial Intelligence, apparently: In a paper titled "Automated Inference on Criminality using Face Images," two Shanghai Jiao Tong University researchers say they fed "facial images of 1,856 real persons" into computers and found "some discriminating structural features for predicting criminality, such as lip curvature, eye inner corner distance, and the so-called nose-mouth angle." They conclude that "all four classifiers perform consistently well and produce evidence for the validity of automated face-induced inference on criminality, despite the historical controversy surrounding the topic."
[...] The study contains virtually no discussion of why there is a "historical controversy" over this kind of analysis — namely, that it was debunked hundreds of years ago. Rather, the authors trot out another discredited argument to support their main claims:, that computers can't be racist, because they're computers:
[...] Absent, too, is any discussion of the incredible potential for abuse of this software by law enforcement. Kate Crawford, an AI researcher with Microsoft Research New York, MIT, and NYU, told The Intercept, "I'd call this paper literal phrenology, it's just using modern tools of supervised machine learning instead of calipers. It's dangerous pseudoscience."
(Score: 2) by jmorris on Sunday November 27 2016, @06:09AM
Let us accept there might be problems with the science in this study, it isn't peer reviewed yet, they haven't released source code or anything that could be examined, etc. Accept all that as givens. Now lemme ask a question.
For the purposes of debate in replies to this post, let us ignore those issues. Assume the full commented source is available, a paper has been fully and properly peer reviewed and that the results have been duplicated. Duplicated by Google, the poster child for big data and huge throbbing brains. No doubt remains, this is Science and it is settled; not by consensus (i.e. politics) but in the customary way of Science, by results so clear any remaining doubters just can't say anything meaningful. So assume it is reality. So here is my question: Now what? We probably should at least run the thought experiment while we have time to carefully consider the options.
Damned if I like any of the things that would come from such a thing.
(Score: 0) by Anonymous Coward on Sunday November 27 2016, @07:00AM
I see it catching easy to catch criminals if it relies on facial data from those that are caught.
(Score: 3, Insightful) by AthanasiusKircher on Sunday November 27 2016, @07:52AM
No doubt remains, this is Science and it is settled; not by consensus (i.e. politics) but in the customary way of Science, by results so clear any remaining doubters just can't say anything meaningful. So assume it is reality. So here is my question: Now what? We probably should at least run the thought experiment while we have time to carefully consider the options.
Except what does it mean that it is "a reality"? What exactly are the statistics going to show? That if you have X facial feature, you might be 10% more likely to end up in jail? Or maybe, in some extreme case, double or triple the chances? I can give you an algorithm for that right now in the U.S. -- "Is person A black? Then person A has a higher chance of ending up in prison... err... 'being a criminal'."
That's a statistical fact, and if you ran random photos of the U.S. prison population through this thing, it would likely come up with that "anatomical feature."
So what? What conclusions do we draw from such "settled Science"? Clearly, if you look at the way such statistics are received by Americans, the fact that it is "settled Science" would be meaningless. Some would choose to see it as proof that Blacks are all likely criminals. Others would see it as an indictment of larger cultural or demographic issues, like the fact that Blacks are more likely to live in poverty, crime tends to be greater in poor neighborhoods, police sensitivity to crime is likely higher and thus leads to more arrests... but poverty is harder to measure in a facial photo.
And even if you could prove somehow that this wasn't biased by arrest or conviction or whatever stats -- and that the sample was free from ALL of that, so what? It's not like you're going to get settled science saying that if a person has X, Y, and Z, they are going to be a murderer 99% of the time. It's NEVER going to be that precise. At best, I imagine such studies might say that people with X, Y, and Z are a few times more likely to commit crimes. But, I'd also bet that the MAJORITY of people with X, Y, and Z still will NOT commit crimes.
So, are you going to start taking actions with those people because some small percentage MIGHT commit more crimes? Again, it's a politicized question, not one for "science" debate. Some will look at those few and say "Ban all persons with X, Y, and Z from entering the country." Others will see that as ridiculous since the majority of those people aren't guilty of anything, nor are they likely to be.
These aren't questions that have a clear answer -- "settled Science" or not.
(Score: -1, Redundant) by Anonymous Coward on Sunday November 27 2016, @11:37AM
Have you ever heard of a hypothetical question? Because you don't act like you have the faintest idea what one is.
jmorris posed the very reasonable hypothetical "if this was proven to work [100%], then what?".
Attacking the premise of an hypothetical is moronic.
So, IF it worked, then what?"
(Score: 0) by Anonymous Coward on Sunday November 27 2016, @12:34PM
Have you ever heard of a hypothetical question?
Is that a rhetorical question?
(Score: 2) by Entropy on Sunday November 27 2016, @06:56PM
Certainly subgroups of people, say males 15-35 from ___ group are about 1000% more likely to commit violent crime. That's a little more than the 10-30% you're suggesting. That's a huge, crude, and almost completely unrefined group. What if this narrows it to a subset of 3% that are 100,000% more likely to commit violent crime? What if it narrows that further by measuring the folks actually in the area a particular violent crime such as rape was committed?
Obviously this isn't justification to toss someone in jail, but it could be useful(though "unpopular") to look at them more closely.
(Score: 2) by HiThere on Sunday November 27 2016, @07:34PM
Well, what your figures actually say is that members of your specified population are x% more likely to be arrested or convicted of committing a crime of some particular variety. They don't say whether the frequency of them committing the crime is higher. One can reasonably suspect that many crimes are committed by those who are never caught, and even that lots of crimes never come to official notice. And if this is the case, then a greater tendency to arrest members of some particular group may just reflect a preference in arresting. More likely it's a complex of interacting factors, and that group actually *IS* more likely to commit the kind of crime that they are more likely to be arrested for, but to what extent this is true is far from clear, and there's no obvious way to determine which factors are more significant. All the accessible data is known to be biased, but the degree to which it is biased is not only unknown, but probably varies in an undeterminable pattern. (I'd say random, but I don't mean actually random, but merely that there is no accessible way of determining the variation.)
Your proposition also assumes that all arrests/convictions are based on valid evidence rather than, say, threats and plea bargains made with a public defender who doesn't want the district attorney's office mad at him. And there are other sources of noise in the data which are non-random in the way they affect it.
Javascript is what you use to allow unknown third parties to run software you have no idea about on your computer.
(Score: 3, Interesting) by AthanasiusKircher on Sunday November 27 2016, @09:58PM
Certainly subgroups of people, say males 15-35 from ___ group are about 1000% more likely to commit violent crime. That's a little more than the 10-30% you're suggesting.
Actually, I'm pretty sure I also said "perhaps double or triple," i.e., well over 100% more likely. Yes, I didn't go up to 1000%, but I wasn't thinking of age restrictions. I agree that you MIGHT get that high based on pretty crude measures.
What if this narrows it to a subset of 3% that are 100,000% more likely to commit violent crime?
First, I really doubt we could get such precision. But even if we could, what are we measuring, though? There are HUGE numbers of confounding factors here. You have selection bias (as was already pointed out in another post) -- what crimes are reported, which are investigated, which are investigated "more thoroughly," which are charged, which are actively prosecuted vs. pled down until these people don't end up in prison at all. And that's just possible bias created by police and the district attorney. Now introduce various other demographic factors (which might be used -- either implicitly or explicitly -- by the attorneys to argue for conviction), whether the defendant can afford competent defense, what biases juries might have for or against certain "types" of defendants (which, yes, includes how they look), etc., etc.
And that's only selection bias for who gets convicted. Then you have the actual statistical confounding factors, like what I brought up about poverty. If you control for poverty and educational level, the majority of the supposed white/black difference in crime rate disappears. There's still some difference, but it's much smaller (and perhaps could be explained by some other confounding factors). So, if we only look at facial features or other anatomical features, are we really measuring the RELEVANT differences? Or are we merely isolating some groups that have apparent higher crime rates based on appearance (pun intended), while overlooking other significant causes of crime that might not show so readily on a face?
Obviously this isn't justification to toss someone in jail, but it could be useful(though "unpopular") to look at them more closely.
Greater than 99% of people -- whether White, Black, or otherwise are NOT murderers. Greater than 99% of Muslims are NOT terrorists. We clearly live in a culture of fear, rather than judging people rationally.
Oh and for one last example: Greater than 95% of Catholic priests are NOT child molesters. The reason I mention that last stat is because a number of statistical studies have shown that the rate of child molestation doesn't appear to be higher among Catholic priests than it is among Protestant ministers, or Boy Scout leaders, or teachers, or coaches, or whatever. The difference is the Catholic Church kept central records, not that they had a higher incidence than the general population. Lots of things can make incidence appear to be worse among certain populations. (And just in case you try to go this way, note that I am NOT Catholic and have no tie to the church. I just am wary of statistical biases created by perception.)
(Score: 2) by JoeMerchant on Sunday November 27 2016, @02:31PM
In today's social climate, this would be an Ouroboros: science eating itself. Science is based on a clear set of ideals/principles, but it lives in a political world. Not saying that the topic should be dropped, but if pursued too aggressively today, it will be bad for the pursuers on personal and professional levels.
The question has been, and continues to be, interesting. Can you quantify tendencies to criminal behavior? If so, should we intervene socially to give extra attention to these people to help them integrate with society? If we treat them the same (as we claim to do today), then you don't need much in the way of predictive science, you can look in the jails and get your data.
There are centuries of socioeconomic history that confound the topic, and those confounds will never be removed without some kind of bias... which is yet another confound.
End of the day, this data might predict increased probabilities of certain behaviors based on bone structure. You can do the same based on (gasp) skin color, blood chemistry, genetics, socio-economic background, prior offenses, and current level of wealth. What you cannot do is convict a pre-crime. If you're going to classify an organism as "human" and entitled to "human rights" then interventions at the level of arrest and imprisonment must be based on actions, not probabilities of future actions. Maybe offer them assistance, voluntary programs, targeted information, but if you're going to arrest based on looks, you might as well admit that you're treating them like an animal, not a person who is presumed to be capable of controlling their actions, at least to some degree.
🌻🌻 [google.com]
(Score: 2) by jmorris on Sunday November 27 2016, @04:03PM
The "punish the heretic" phase of politics generally stops if incontrovertible evidence supporting a theory appears. Which is why the forces of Progress are to vehement in their policing of science, so they can ensure 'bad ideas' are never funded, anyone pursuing 'politically incorrect' lines of research are quickly declared 'not scientists' and driven out, etc. But I assumed the final argument ending paper already exists.
There are centuries of socioeconomic history that confound the topic...
And there ya go rejecting the premise of the theoretical. Assume it works. Assume you can feed in 10,000 faces and get a list of 250 that will, with the classic wee P value representing 95% certainty that gets you declared "SCIENCE!", become criminals in the next decade. Yes, if this were running in America, it would be almost certain the results wouldn't be race neutral. That is just one of the consequences that would upset most people. But unless you are asserting that ALL blacks would be flagged as criminals that doesn't really matter. What does is you have a list of people almost certain to be, or become, criminals. So then what?
Progress would push to criminalize even possessing the software to generate such a list, but I hope there would be a consensus here that would be both unwise and immoral. So like most political questions, we know the Progressive position is both wicked and ineffective, but it doesn't leave a clear alternative path. What happens as employers gain access to the algorithm and training datasets? Potential dates? This is like public key crypto, it is math, once out of the bottle you don't just stuff the genie back in. So what sort of consensus could we arrive at to deal with this knowledge in a rational way?
I suspect that if it could be proven that the thing detects early enough, the education system would have to use it, to attempt to provide additional educational resources and training to the children at risk. But beyond that, no frikkin' idea other than a prayer we could in fact come up with interventions effective enough that the thing couldn't predict adult behavior good enough anymore for people to depend on the hits.
And this is just the start. What happens when a big neural net gets trained on DNA + extensive life history data and starts making accurate predictions?
(Score: 2) by JoeMerchant on Sunday November 27 2016, @05:13PM
Does it matter if it's skin color, bumps on the skull, or ratio of pinky to bird finger length that is a p0.95 predictor of future criminal intent? Whatever the determining factor, at what p-value do you pre-convict and address the problem with retro-active abortion?
Basically, at what point do you declare the child of two human beings non-human?
🌻🌻 [google.com]
(Score: 2) by jmorris on Sunday November 27 2016, @06:02PM
Yea, that is the root of this problem. That if we can ever accurately, for whatever value of 'accurate' we pick, predict criminality, is it acceptable tp pre-punish or 'pre-treat' the problem. Really hard to see how the solution doesn't quickly become worse than the problem, but equally hard to see the reasoning for why we would simply ignore the knowledge or even the ethics for doing do. It is just a real mess. And probably in our future. If this research flops there will be more later, eventually the odds are somebody does it.
(Score: 2) by JoeMerchant on Sunday November 27 2016, @06:13PM
To me, the "humane" thing to do is offer assistance, voluntary intervention programs, etc. and if/when the predicted does happen, focus on rehab instead of punishment. But, then, that's what I think of all "at risk" populations whether identified by science, social workers, or the justice system.
Science supports rehab as not only more effective for the criminal, but overall much less costly to the community... but, politics rules the day in that arena, so rehab gives way to punishment more often than the reverse - at least in the jurisdictions I know anything about.
🌻🌻 [google.com]
(Score: 2) by jmorris on Sunday November 27 2016, @06:45PM
Not sure what science you are looking at. Treating criminals as a health issue works for some criminals and fails utterly for others. Pedos for example are notorious for being all but untreatable, either as a mental health case or with punishment in the criminal justice system. Put them back into society and they almost always end up re-offending. Petty criminals are at the other end of the spectrum, remove them from the bad environment, get them trained in a marketable skill and gainfully employed and the re-offense rate is generally considered to be an acceptable risk.
(Score: 0) by Anonymous Coward on Sunday November 27 2016, @07:11PM
To me, the "humane" thing to do is offer assistance, voluntary intervention programs, etc. and if/when the predicted does happen, focus on rehab instead of punishment.
And how are these people going to be identified? Through mass surveillance? Well, mass surveillance is unconstitutional (in the US) and unethical in and of itself, so this whole scheme isn't really voluntary to begin with. If you have to destroy innocent people's rights in order to identify these people, then you should never be allowed to implement this system.
(Score: 2) by dry on Sunday November 27 2016, @09:25PM
The progressive idea is that people are innocent until proven guilty along with the idea that it is better to err on letting the guilty to go free rather then Bismark's thing about it being better to convict 9 innocent people rather then let 1 guilty person to go free.
Even with your unrealistic 95%, that means that 5% will not be criminals. The regressive would rather throw all 100% in jail for pre-crime whereas the progressive idea would be to try to stop those 95% from being criminals and if that fails, wait for them to become criminals.
This also raises the question of whether those people who are damaged in some way that leads them to being criminals should be punished or just removed from society.
In reality the numbers are closer to 60% rather then 95% according to the science that the progressives have already done. Expose a fetus to alcohol (actually just study people that were exposed to alcohol as fetuses) and their faces will have small eye openings, a smooth philtrum and a thin upper lip. 60% of those will drop out or be kicked out of school and 60% will end up in jail (aged 12 and older for both). As the study included different degrees of FAS, it follows that those with extreme FAS are very likely to end up as criminals due to being damaged.
The progressives try to prevent them becoming criminals and if that fails, humanely remove them from society. The regressive would go on about how they made bad choices and should be punished and gloat about putting them in a situation that leads to them being raped.
https://en.wikipedia.org/wiki/Fetal_alcohol_spectrum_disorder [wikipedia.org]
(Score: 2) by AthanasiusKircher on Sunday November 27 2016, @10:36PM
There are centuries of socioeconomic history that confound the topic...
And there ya go rejecting the premise of the theoretical. Assume it works.
I don't read this as "rejecting the premise of the theoretical." That issue is actually at the heart of whether and how "it works." A predictive statistical model need not deal directly with causality. In fact, one major point of the Scientific Revolution was essentially a rejection of the premise that science must always have a causal explanation for a mathematical model. Looking for causality is Aristotelian thinking. Newton spearheaded modern science by effectively saying, "Yeah, there are these unseen forces in my model, but I don't know if they really exist or what the true cause is... but the math works." (He specifically added such apologetics to later editions of the Principia, since many criticized his "unseen forces" as potential claptrap, which belonged more to mystical traditions like alchemy and Hermeticism than "science.")
To me, it sounds like you're attempting to skirt the same issue here, by "assuming it works," without addressing the mechanism. In some areas of science, that's reasonable if you're just trying to create a mathematical model. But we're trying to do more here -- we're presumably trying to find the CAUSE of crime and perhaps PREVENT it. We're trying to change nature. That's different, and it's less like Newton's physics as abstract math and more like the way people research diseases or sociologists look for causes of poverty or whatever.
So, in the present case, it's not just the "socioeconomic history," but the ongoing effects of that history. "Confounding factors" are essential to determine causality and to aid in effective prevention. As I posted separately here, when you take poverty into account, a lot of the apparent differences in white vs. black crime rate disappear. Not all, but a lot. If we want to address crime for the future, should just target those with "African" facial features (as a naive math model might tell us)? Or should we look at broader societal factors which may be behind the apparent differences in our facial feature study?
This isn't rejecting the premise that a mathematical model might "work" -- it's asking WHY the model appears to work, and how we might act on it, depending on whether the physical features are causally related vs. just a symptom of other issues that seem more causal upon examination.
Or are you actually proposing that "assume it works" means that we've determined causality -- that somehow we can prove that facial ridges cause crime? Or that the same developmental process creates facial ridges as creates brain patterns that lead people toward crime?? IF we could actually prove that beyond a reasonable doubt using rigorous science, I have no doubt it would lead to a new eugenics movement, selective abortion, etc. But I sincerely doubt studies like this are ever going to get close to proving causality like that. (And that wasn't even really suggested in your initial post, where you referenced "big data" replication, not neurophysical investigations of causal links or whatever.)
(Score: 2) by HiThere on Sunday November 27 2016, @07:17PM
Actually, there's no evidence that there's anything wrong with the program, what's clearly wrong is the method of training. This is a common problem with AI. That if they had properly trained it it wouldn't have been able to make any predictions is only true if the theory is wrong...which I believe it to be. Criticizing the AI program because of the training it was given, however, is inappropriate. A biased dataset will bias any program. It's just that AI tend to have more edge cases, and the edge cases tend to be harder to detect. (Well, that happens as the task gets more complex.)
N.B.: For a typical AI problem even an unbiased dataset won't work, because of edge cases that get ignored during training, but which show up in use. People solve this by never turning off training, thought they tend to turn it down as they get experience in an area. That's probably the right solution, but it requires continued feedback, which adds to the cost.
Javascript is what you use to allow unknown third parties to run software you have no idea about on your computer.