Crime-Predicting Algorithm Outperforms Humans
technologyreview.com
Crime-Predicting Algorithm Outperforms Humans
1–10 of 18 posts
Re: Crime-Predicting Algorithm Outperforms Humans
#2> Where police used to sit in daily meetings to plan where to patrol, they can now spend more time actually out on patrol, since the computer's doing the planning.
While this is a nice application of machine learning, I can't help but wonder how much of their increase is due to the above fact.
Police being part bigger parts of their neighborhoods must be a non-negligible variable.
Re: Crime-Predicting Algorithm Outperforms Humans
#3> Where police used to sit in daily meetings to plan where to patrol, they can now spend more time actually out on patrol, since the computer's doing the planning. While this is a nice application of machine learning, I can't help but wonder how much of their increase is due to the above fact. Police being part bigger parts of their neighborhoods must be a non-negligible variable.
Even if you assume that the algorithm's quality is exactly on par with the seasoned police veterans, the fact that it's fast is a sufficient value-add by itself.
Re: Crime-Predicting Algorithm Outperforms Humans
#4What will be said to those who claim racial profiling by these algorithms?
Re: Crime-Predicting Algorithm Outperforms Humans
#5> Where police used to sit in daily meetings to plan where to patrol, they can now spend more time actually out on patrol, since the computer's doing the planning. While this is a nice application of machine learning, I can't help but wonder how much of their increase is due to the above fact. Police being part bigger parts of their neighborhoods must be a non-negligible variable.
Even if you assume that the algorithm's quality is exactly on par with the seasoned police veterans, the fact that it's fast is a sufficient value-add by itself.
In that case, I've got an even cheaper and faster solution: Pick some points on the map randomly.
Now that I think about it, this could be used as a control.
Re: Crime-Predicting Algorithm Outperforms Humans
#6What will be said to those who claim racial profiling by these algorithms?
Would people complain that the police are or aren't visiting their neighborhood?
Re: Crime-Predicting Algorithm Outperforms Humans
#7Re: Crime-Predicting Algorithm Outperforms Humans
#8Never quite understood the draw of headlines like this: "Human With Tool Outperforms Human Without Tool"
Yep. That's what humans do. Now, if you really want to catch my attention, give me a headline like this: "Humans Without Tool Completely Fail To Freak Out When New Tool Appears".
Re: Crime-Predicting Algorithm Outperforms Humans
#9What will be said to those who claim racial profiling by these algorithms?
There was previous discussion on this the other day... http://news.ycombinator.com/item?id=4185684
I think it depends on the inputs. If the algorithm looks at statistics blindly, ie. does not take into account race, then it would be more accurately called "crime profiling". So, unless people were to feel that crime should be more acceptable in some circumstances than others (i.e. a crime prone area should be allowed some low level crime because that's what people know), it should be a non-issue. I think the question is more about how the police on the ground respond, do they inject race into the the practical aspect?
It would seem that that should not change just because police are being dispatched by an automated system --what I mean, is I don't think this system should or will make beat cops more or less biased against a given ethnicity, per se. but in effect a neighborhood will get more attention and long-run, should end up in a better place, given that results show a significant improvement.
In the end, time will tell.
Re: Crime-Predicting Algorithm Outperforms Humans
#10Earlier quoted context omitted.
Even if you assume that the algorithm's quality is exactly on par with the seasoned police veterans, the fact that it's fast is a sufficient value-add by itself.
In that case, I've got an even cheaper and faster solution: Pick some points on the map randomly. Now that I think about it, this could be used as a control.
That would actually have the added benefit of avoiding profiling.