Cameras That Can Spot Shoplifters Even Before They Steal?
bloomberg.com
bloomberg.com
How do these AI systems account for causation?
It's the fundamental issue I have with the penetration of these sorta "AI" systems in daily life: from what I understand about the basic maths, they're ultimately all seeking correlation and real-world data used to improve models seems to achieve verification, not validation.
I'm guessing I'm not the only person who reads this article and experiences a visceral reaction.
This is what I would call "ugly" tech.
i think it’s quite great tech, definitely a huge need in the market for this sort of prevention mechanism.
How is it not about policing thought?
If you present physical attributes/symptoms of someone who is thinking about shoplifting (nervousness, anxiety, fidgetiness, etc., as mentioned in the article), the models can classify you are "about to shoplift".
That's pretty close to thought policing if you ask me.
The causation gap for these models presents a really terrible scenario.
Without causation, these models are capable of correlating things like race, gender, etc. with outcomes like shoplifting.
Nobody infringes on your rights, if a system pings you as suspicious and so a security guard is alerted and watches you more closely. There are no false positive consequences here: either they find that can of Red Bull on your person without a paid receipt, or they don't. You won't get banned from the store, because the system inadvertently deemed you suspicious.
A whitebox statistical model still can provide the factors that contributed to its prediction, without resorting to causal inference.
I do not think a certain race, gender, age, income causes shoplifting, but they sure are correlated, and effective at finding shoplifters or tax fraudsters. I deem the act of shoplifting more unethical than the act of watching you on a camera feed. This already happens anyway, these systems just manage attention better.
i think we should stop trotting out this tired nugget without understanding how sparse the alternative is. Even within the academic stats community, there’s a fair amount of handwringing that rho is honestly all we have. fpci is one of the most dispiriting theorems in statistics. we get by with rubin models & sem, but these are pretty hokey and outside of social sciences you get laughed at if you say ran an sem. you can regress student score on study time & get a positive coefficient, but then you are setting yourself up - why did you get that positive coefficient ? well it means if he studied longer he would have scored higher. yeah but then why did he not study longer ? the fuck i know! oh wait a minute let’s do sem! ok we can’t observe depression so that’s a latent. and we can’t observe teenage angst so another latent. soon your friendly neighborhood psychologist thinks up half a dozen latent causes. put the whole thing into the sem mixer grinder with score as response. yay sem software gave me significance! ok so depressed teenagers will score poorly on test because they won’t study for it...but wait a minute why are they depressed- oh lets do another sem throw it divorce rate & obesity & lack of religious attendance & regress again....this is when the statistician gets up & says fuck all you assholes you don’t have anything to establish any of these things, you are just grossly misusing linear regression. then the social scientists respond well we don’t need your math kind anyways, we have sem software, its latent turtles all the way down. so what if beta hat is xprime x inverse xprime y? can’t publish a psych paper with just that. gotta have my latent goodies. causal inference ftw!
god sometimes this whole field makes me hurl.
1) The "signs of shoplifting" are human-designed, and so are heavily biased 2) The "signs of shoplifting" are entirely machine learning-based, meaning that they are overwhelmingly biased towards the people who actually get caught and the people who are of communities more likely to shoplift, either because of economic necessity or because of being more likely to live in areas with higher crime rates for whatever reason.
Both lead to racist results without any explicit racism being required.
Also in the absolute best-case scenario that this system works, all you're achieving is locking up more people stealing out of desperation for the benefit of corporations. Even that terrible situation, though, is a pipe dream compared to the reality of systems like these.
If the discrimination (not racism!) causes a bias towards a certain race, neglecting other races who shoplift, and focusing on innocents, then the accuracy will suffer. These systems are not deemed problematic because they are ineffective, they are problematic, because they do a "good" job.
Can you elaborate on this? As stated, it sounds not just inaccurate, but impossible: it reads like youre saying that any policy that doesn't result in a uniform (or proportional) racial ratio is illegal, which obviously isn't the case.
I think you're talking about disparate impact (as a legal concept), but this is 1) IIRC likotes to certain domains, like housing or employment and 2) avoidable if the racially-skewed policy is demonstrably related to the job (in the employment case).
As a trivial example, requiring a college degree for a job dramatically shifts the racial distribution of the candidate pool, and obviously isn't illegal.
If the shoplifting system is found to be illegal, AIUI, it would probably be because the case is made that it's _directly_ racially biased (disparate treatment).
Same group that complains when a TSA agent frisks a five year old white kid. All in the name of fairness.
Police has a limited capacity. I want that capacity, paid for with taxes, to be optimized. If that means racial profiling, because some races are more prone to crime, then so be it. That is another form of fairness. Humans employing these techniques make use of common sense, a long sought after feat that AI barely mimicks.
You do not fix societal problems, such as racism or racial crime statistics with algorithms anyway. These communities should put the blame elsewhere, starting with themselves, before they point the finger at "racist" algorithms. It is not statistics fault when certain groups are more or less likely to commit crimes.
So the communities of people who are direct descendants of people were forcibly brought here, enslaved, and then systematically denied access to the mechanisms the rest of our ancestors used to build wealth, should blame themselves because they have higher crime rates?
White families have more than 10 times the wealth of black families on average. This is directly caused by the years of systemic oppression faced by their ancestors.
In addition, to this day, we arrest them at a higher rate rate than White people for the same crimes. We are more likely to convict them for the same crimes. And we give them longer sentences for the same crimes. Elementary school teachers even give harsher punishments to Black students for the same infraction.
Job seekers with black sounding names are significantly less likely to be called back for interviews. Black renters have a harder time renting than White renters with the same credit score. The list goes on and on. When you combine all of this, you end up with a community with a higher crime rate. The problem is caused by systemic injustice, it's not a problem that black communities can fix themselves. If you apply those same conditions to any demographic, you would see a higher crime rate--it's inevitable.
>If that means racial profiling, because some races are more prone to crime, then so be it.
And it's absurd to use that higher crime rate to justify racial profiling, which inevitably has a disparate impact on innocent Black people, which erodes trust in police and institutions, which directly leads to even more criminality. In turn that leads to even more racial profiling. You can't just ignore positive feedback loops by mentioning them and pretending they don't exist.
>That is another form of fairness
Trial by combat, enslaving defeated enemies, and separate but equal are all "other forms of fairness" that our society has rejected.
You can’t say if [some race] then that. But you can use other things (are they loitering, barefoot, poor hygiene, has sunglasses, truckers cap, pulls over hoodie on entering, have a decoy, etc)
If the proportion of white shoplifters and black shoplifters being caught is different, AI will just scale up that disparity. If reality is biased, we should work to fix that: not make it more efficient.
To specifically address your point:
"so long as their accuracy is within reasonable range of effectiveness"
This is a far too narrow look at what accuracy means - that's why we usually look at sensitivity, specificity, false negative rate, and false positive rates, and why we break those rates up over important subgroups. To the example numbers you give, let's posit (without evidence) that your numbers are correct - they don't tell us what proportion of people visiting the store belong to each group (is the model just learning to identify ethnicity?), and they don't tell us what proportion of each group is really shoplifting.
In many areas (drug crimes come to mind), there's good evidence that different ethnic groups commit crimes at similar rates, but arrest rates are wildly, grotesquely disproportionate. Imagine that AI is applied to such a system - it learns, correctly, that black people are more likely to be arrested for particular crimes. It then targets black people, and the accuracy numbers look good on paper. In fact, because of the trust we have in technology, such a system might help normalize or justify this crazy preexisting disparity.
If you were building a statistical model to describe predictors of shoplifting, this kind of bias wouldn't make it past a cursory level of peer review. When you're slapping "AI" on a product and implementing it at a huge scale, nobody bats an eye.
My point is about the data generating mechanism - making inferences conditional on a bad selection procedure is a bad idea. It would be bad statistics, and it's bad ML. In this case, my argument is that it's also bad for society.
For example I can legally profile short people, but it isn't moral to do so. And I would rightfully be judged if I did so.
Asking if it is right to judge people based on their mental health is an entirely different topic than asking if we can judge people based on a physical characteristic they were born with and have no control over. Apples and oranges.
Well you can hope, but it's made by humans, so I wouldn't count on it.
Edit: to people wondering why I do this, I can only say that it’s something that I never even thought of as weird. That’s what growing up rich can do you.
I have a friend that does the same thing and it just seems really weird to me.
My groceries aren't mine until I've paid for them. And also, as a person with not always a lot of money I sometimes (very rarely, but still) find that I don't have enough money for the things I wanted to buy. What if I'd started consuming the goods, got to the till and couldn't pay?
If I needed to drink first I'd buy the drink first, drink it, and then proceed with the remainder of the shopping.
Anyway, people can do what they want. I'm not going to berate anyone for this. But I still think it's really weird.
I think a restaurant is different though because that is explicitly how it works (eat first then pay).
To take your comparison further, what would happen if you sat down in a grocery store and started consuming hundreds of dollars worth of food and drink? I think it would raise a lot of eyebrows, and the employees would probably confront you. So there is a difference.
Furthermore, in a restaurant you place an order and they serve you, and from a legal perspective (IANAL) there might be something about some sort of implicit contract that you will pay them. Whereas with a grocery store you are not allowed to remove the items from the store until you've paid for them. And I feel like consuming the items inside of the store is sort of analogous to removing them from the store.
I would also add that if the store owner came up to you and challenged you to please pay for that right now, you're obligated to do so. Considered as a whole, it's not a good idea for someone to do that over a bottle of soda for all kinds of reasons, but I'd say the right still is essentially there.
I think a potential risk here is from small initial sample sizes. Here's a thought experiment. A system like this is set up and runs successfully for some time. A person with some distinctive characteristic enters the store and steals something and is caught. The system learns that so far 100% of people with that characteristic stole. Next time it detects that characteristic, it flags that person and they're ejected from the store. The system will never learn that this characteristic can be benign, because from now on it will be excluded from it's data set.
That's a bit of an artificial example, making assumptions about how the system might work, but something along those lines could be a potential failure mode, and we might never know what that characteristic is because the system can't express the reasons for it's decision in a meaningful way.
At least the won't try to arrest you before committing a crime like in the movie
Imagine being banned from your main grocery store option because of a false positive from this system.
Every time we make an advancement someone uses it in exactly the opposite way than intended. In this case people will get too confident in the technology, and start accusing random people with a tick or other false-positive indicator.
At the same time, we're reaching a point with data science and sensing technologies that we've not looked at from a moral and ethical perspective at large as a society. These technologies will certainly be used by the ruling class (wealthy, powerful) for their best interests to further entrench this societal structure. As with any technology, it can also be used to serve humanity at large so the technology itself isn't bad, it's always a balancing act of how it's utilized.
We, as a society, really need to begin considering moral and ethical implications of the uses of these technologies by the ruling class and decide what we want and don't want to allow. If it's left for them to decide, I can assure you, it won't be good for vast majority of us plebes/proles (citizens).
Maintaining a democracy is a never-ending fight and we meager citizens currently seem to be losing the battle as technology accelerates the ruling class's grip at every turn. Heck, people on this very site are indirectly contributing to these development efforts for what now seems like a great payoff. It's easy to get wrapped up in the beauty and complexity of technology without considering its implications, especially when you're getting paid quite well (for now). You only need to look back a few decades at retrospective assessments from some of the greatest minds in the century that developed modern nuclear weapons as a guide.
We need to organize and step-up our game to take back control of our government to serve us (the people), not a select few extremely wealthy and powerful folks. We need to consider moral and ethical implications of use of these technologies and enact laws with serious protections and penalties before mass abuse of technology grows more rampant.
Good security would see your "target" as the "subject".
The target is the guy taking advantage of the new camera prediction system by waiting for a subject of interest to enter the area before proceeding to rip you off while you're otherwise engaged.
My guess is that the shoplifting rings will have a field day in any mall relying on tech like this.
I can have sympathy for the teenager who steals a snickers bar, ok. But these roving bands of shoplifting thieves, I hope they get caught and put out of business.
Isn't this survivorship bias? You can correlate suspicious mannerisms with shoplifting only because the shoplifters without suspicious mannerisms don't get caught. Has there been research into what percentage of inventory shrinkage is attributable to people who exhibit such suspicious behavior? This also leaves unanswered the question of how many people exhibit these behaviors without being shoplifters, e.g. people with social anxiety (exhibiting general nervousness), ADHD (insatiable restlessness), Tourette's (uncontrollable physical tics), or on the autism spectrum (which can compel one to conspicuously touch everything in sight).
Some say that the neural network will be racial profiling instead of working on body language tells. However, having worked in retail and knowing the attitudes of some of my colleagues, I think I would trust a computer over some of my former workmates when it comes to having fair suspicions.
The training for the neural network is simple, the known thieving incidents get traced back to the thief entering the door and doing what all thieves do when they enter the store - scope out the counter, the cameras and mirrors. It is all body language.
There can be off-duty thieves who don't steal always but will be showing some of the same tells. These will be false positives. But then there are the long tail thieving situations that can not be spotted by a mere human.
For example, a woman regularly stealing cat litter and pet food with every weekly shop for a family. She simply placed the items on the lower part of the shopping trolley, checked everything else out properly and conveniently forgot to pay for the 'minor items' on the lower rung.
No human would have thought a normal member of society would risk reputation for the small financial gain involved. This probably started as an accidental theft one week with no 'nerves' involved. By the following week the 'OMG silly me!' routine would have been prepared for. There would be no looking at the security arrangements walking in, so very different 'tells'. Humans could not spot this, but a neural network trained on every bit of shrinkage across a vast store network might stand half a chance.
If you do work in retail you actually want to deliver excellent customer service and to trust customers. So, from a point of view of wanting to do that and give customers undivided attention, I welcome our new neural network overlords.
Even if it doesn't, the opacity of ML means that it's going to difficult to defend against charges that it does.
Great, so sociopaths will go unchallenged because the computer deems them trustworthy, and socially-anxious people will endure YET MORE intrusive encounters during what should be a simple trip to the store.
"Can I help you? No, I won't go away, I don't understand this computer thing but it says I really need to stay near you while you're in the store."
That'll be lovely.