Sci-fi policing: L.A. police predicting crime before it occurs
sfgate.com
sfgate.com
I could see an algorithm come up with a lot of interesting correlations that make sense in hindsight but might not have been thought of ahead of time. For example, maybe a particular stretch of bars the weekend after payday are more likely to have confrontations just after closing time. Or maybe certain neighborhoods are more vulnerable to robberies during a specific holiday. Or maybe specific set of sports teams pitted against each other leads to conflict outside the game.
Using police to passively deter crime rather than punish it seems like a win-win.
Perhaps, not in the case of this example. But think about GPS tracking. Before if you wanted to follow someone around you usually had to dedicate at least 4 people 24/7 to watching them. Now, just use GPS and your good to go.
It's not so much about watching the good guys, but in my view it's about not having to make the choice to allocate resources. Is this person really worth taking 8 guys (4 per 12 hour shift) off of another case/patrol, etc to watch one person. That probably prevented a lot of abuse. Now, you slap a tracker on their car and come back in two weeks and pick it up and all the data is there.
Suppose you have an upper class neighborhood that tends to collect a lot of noise complaints on the weekend. In theory sending a cop there might make an algorithm that's trying to minimize response times happy, but the reason cops tend to be stationed else-ware is to better respond to serious crimes because the station is 2 minutes away in case of a serious issues so there are better places to send cops to cover a full county. Sort of how a lot of algorithms made bad risk assessments because they had a poor understanding of what happens during a significant downturn in the housing market.
In theory this works out a lot like the old halting problem proof, but in practice we may not see such an arms race.
Yes. It may not be a dominating factor, but yes, it is a factor.
As they said, there probably will be questions about profiling people and neighborhoods, but if the desired result is to reduce crime and it does reduce crime, does that not in fact subvert the notion of the stereotype? That's to say, if people think this might stereotype a neighborhood as one with more than normal criminal activity and this is implemented and this brings down crime, then in time, this neighborhood would be better than normal, in crimes statistically speaking and would rehabilitate the neighborhood, rather than vilify it.
Besides, in reality, it doesn't matter what the police or the mayor's office say, it matters what people on the street say. People on the street are unconcerned with being PC. They will tell you if a neighborhood is good or bad. One of the most honest assessment I have gotten is from RE agents. They don't offer any filters --they just tell you how it is, from my experience.
Don't deploy this and let crime continue to be fought traditionally --this results in more neighborhood crime victims and continued view of this neighborhood as crime ridden.
Deploy this technique with a long term result of decrease in crime such that it's less than that of places which did not implement this prevention technique. Once that is achieved, there would be no inclination to treat people in those places differently.
So, given those scenarios, would a neighborhood accept different treatment knowing that long term crime would decrease and thus obviate the need to be treated differently, or would they prefer to continue the status quo knowing this would allow crime to continue as it had but getting to keep their current level of civil liberties?
I know those not the only options, but given those two, I wonder what a group might choose?
Also, we do not know how this technique would affect people's civil liberties. We only guess that it would.
My only concern is this: if the long term effects of this project will indeed reduce crime, do we have a society that's evolved to be less criminal or so we simply lack the data (because we've actively changed it) to continue to be effective at preventing crime?
Some people have found stop-and-frisk unacceptable, yet others find it acceptable.
I don't think measuring decrease in crime is that difficult. One can find out statistically or by way of anthropological study at the other end of the spectrum. One of the easiest ways is probably to ask the residents. do they feel that crime has risen or fallen? Ask Real Estate agents --they have a good pulse on a neighborhood.
Now, something unrelated and this is just a hunch, but I guess that neighborhoods with high home ownership would have better reception for this than neighborhoods with low ownership --not because the population prefer crime, but because a reduction in crime would have the side effect of having a positive (higher) impact on rents and that's a negative for most renters.
Imagine people people being able to report, not actual crime, but what could be suspicious behavior, to the police. Such "information" could be used to move resources to an area. You certainly can't harass someone based on such suspicions, but certainly can be more ready to respond if a crime does occur.
And property owners might like to know this sort of information. If someone is hanging around your parked car for example. Imagine being able to reply with "Oh, that's my teenager waiting for me to come back." or "That's my psycho-ex stalker girlfriend! And she has a crowbar!"
And then we'll accidentally create artificial intelligence! (I hope)
I think the Minority Report allusion is purposeful in order to make it sound more sci-fi-ish and thus thus more questionable.
If you're talking about UCR that's not completely true. The FBI does have reporting guidelines, but individual departments have been known to "interpret" those guidelines differently for political and funding purposes. That is one of the reasons why the FBI discourages "ranking" of metropolitan areas for crime based on UCR stats.
Most of the data they use is probably available to the public already under various open data schemes, so it's only a case of wiring it up.
It'll turn into a game of cat and mouse like SEO is for example.
However, I think that your average smash and grab robbery, at least in the UK, is usually handled by the insurance companies rather than the police. The police provide a crime ref so you can claim that it is a crime and that's it. There's very little prevention or follow up done.
I doubt it'd change that much if they deployed it here.
The sad part: someone will eventually do it.
The second part, I agree, is completely criminal. I doubt it will become available on any mainstream app store.
Except as part of a sting operation.
Speaking from personal experience there is validity in that assertion. However, I'm question any algorithm that claims to predict human behavior, particularly when it involves the police.
They were pretty awesome and there is a huge amount of space for them to grow.
I'm assuming they mean where poor people live.
A two layer neural network with input (features) as the profile of a crime in a certain city and all of the related circumstances reduced to numeric columns. That should be able to produce reliable predictions on where/when crimes happen.
Andrew Ng talks about how to build this two layer neural network here: http://www.stanford.edu/class/cs221/
Glue it together, feed it the data, refine the inputs, test the outputs and revise the inputs/features, and there you go.
And yes for the liberals, this is going to profile people on all of the politically incorrect data points, race, color, sex, orientation, drug use, etc.
How's that going to be detected?
Thas one column would not be enough to decide with any certainty at all, however, if you pass in age, sex, weight, orientation, race, skin color, religious affiliation, address, marital status, prior convictions, languages spoken, income, has-vehicle, make/model of car, profession, where they work, smokes, drinks, and hundreds of other spot data points about them, a two layer neural network would be able to pull it all in, and use historical data to train against later data, and use recent data to predict where the crimes are going to occur, and it will be eerily right, gathering information gain from what seems out of thin air.
This program isn't profiling people, it's profiling crimes.
> "The new program used by LAPD and police in the Northern California city of Santa Cruz is more timely and precise, proponents said. Built on the same model for predicting aftershocks following an earthquake, the software promises to show officers what might be coming based on simple, constantly calibrated data — location, time and type of crime."
Location, time, and type of crime. Multivariate regression, not neural networks.
Why? Because we live in a dynamic, non-fate based world. Crime is a choice, and cannot actually be predicted (because destiny is not set). I can walk up to the door of a convenience store with a gun, stop, think, and turn back. Thus random patrols in known bad neighborhoods is guaranteed over a large dataset to be as effective as 'predictive' crime technology.
You can predict a crime, just like you can't predict where a drop of rain will fall. But when the weatherman says "80% chance of rain in the north", I'll be sure to take an umbrella.