They're not supposed to learn things like that, it will affect their close rates

--dave
My local Chief of Police has fought for years to get his people to "keep the peace" instead of "show high case-closed numbers". He's started to succeed, and the crime rates are going down, but he's been rewarded by budget cuts and being phased out for being too expansive... Bummer!

You need as many 9's after the decimal point as you have digits in (N * N-1). As N is unbounded and accuracy is bounded, you get screwed. It's fine for a 10-person company (90 comparisons, negligable false positives) It's out of the question for airports (10,000 * 9,999 comparisons)

As the ARPAnauts would say "it doesn't scale"

It's TERRIBLE public policy for people to be pulled aside for mere physical resemblance to a third person. A person the cop's never seen, and only has a photo of, but they've been told by a computer that this is the person in the photograph.

And computers are never wrong

better technology doesn't help enough!

To oversimplify, if you have 1 error in a thousand, and you have 10,000 (crooks + innocent people), you do (10,000 * 9,999) comparisons and get 99,990,000 / 1,000 = 9,990 errors. In stats, it's a selection of every two persons out of 10,000.

It's really something like (select one of 100 crooks from 10,000 innocents), but it's still an insanely huge number of comparisons. Hoeever good your technology, adding more people will give you (N * N-1) more chances of getting an error.

Facial recognition vendors are very careful to NOT report their error rates in ways that expose this problem: it's the "elephant in the room" for that industry. And that includes Siemens, my former employer.

ABSTRACT
This article argues that our current laws on the topic are counterproductive, because they protect child molesters instead of bringing them to justice, they criminalize a generation of normally-behaving teenagers which diverts valuable police resources from the criminals we should be going after, and they lead to censorship and electronic book burning as well as unacceptable collateral damage to innocent families. Child abuse as such is not condoned by anybody, and this article argues that current laws are counterproductive in preventing and prosecuting it.

In http://falkvinge.net/2012/09/0... The abstract is there because the title of the article will enrage the folks doing the prosecution...

"Do one thing well" is how Unix kernel functions are written, and it's just plain a good idea. Systemd probably follows the first principle internally, many programs do.

Creating production systems[1] out of single-purpose commands connected by pipelines is a different principle, and only works if you keep them pretty simple. It's not a prionciple, but it is how a lot of Unix scripts are written, NOT including the shell that glues the parts together, and not including all the more complex programs, like ed or mail. Systemd doesn't follow the second, because it's more like ed than a text transformation like spell.

A more useful question is whether systemd as a whole does one thing, and does it well. About that, one might usefully discuss whether the Unix principle applies.

--dave
[Pipelines were patterned after a subset of "production systems" in early AI, which applied transforms to "produce" new things. They're not the kind of production systems you put on a raised floor]

You can have 99.999 accuracy, and if the number of comparisons is (N choose 2), then the probability is (N chose 2) * 0.00001, which will be (really huge number * 0.00001) which is (merely huge number).

I don't care how good or bad the implementation is, it has to have more 9's to the right of the decimal than I have zeroes to the left in the number of people, N. That's a known hard problem in computer science (;-))

--dave
[And yes, Siemens was getting crappy even then, but that isn't the problem that the FBI has to solve]