4 ms·

> That is definitively false. We're really, really good at it. machines [are] cheaper and scale better. I am taking the "cheaper and scale better" part into ac

> That is definitively false. We're really, really good at it. machines [are] cheaper and scale better. I am taking the "cheaper and scale better" part into account when I say that machines will be more accurate. If you want to pick a face out of thousands in a crowd or match a single face to a list of thousands of targets facial recognition is going to outperform humans at the same cost. > the false positive rate is a few percent The false positive rate is whatever you configure it to be. The confidence level for flagging someone is a parameter and you can adjust it to whatever false positive rate is acceptable. > the vast majority of people the machine flags will be innocent Every single flag will be reviewed by a human who will decide whether or not to take action based on it. This is no different than a human looking at a list of suspects and then watching for matches, except that the computer can use a much larger suspect list, and give the human hints. > Worse, the people the machine thinks look like someone on the bad list, it always thinks look like them Every time a person is flagged that must be reviewed by a human. If the human marks the flag as a false positive then the algorithm learns from its mistake. The idea that people will be repeatedly flagged is FUD unless they look a lot like the target, which is a problem with humans too: http://www.chicagotribune.com/news/local/breaking/ct-identical-twins-murder-confession-met-20160922-story.html http://www.chicagotribune.com/news/local/breaking/ct-identic...

AnthonyMouse 9y ago
> I am taking the "cheaper and scale better" part into account when I say that machines will be more accurate. Then you used the wrong word, because cost and throughput are not accuracy. A machine that always returns "true" can be very cheap and very fast. > The false positive rate is whatever you configure it to be. The confidence level for flagging someone is a parameter and you can adjust it to whatever false positive rate is acceptable. When that makes the false negative rate ~100%, you might as well do yourself a favor and throw the entire system in the trash. > Every single flag will be reviewed by a human who will decide whether or not to take action based on it. This is no different than a human looking at a list of suspects and then watching for matches, except that the computer can use a much larger suspect list, and give the human hints. That's the problem though. Getting arrested and subjected to twelve hours of questioning every time you walk in front of a camera will ruin your life even if you ultimately get released every time. The much larger suspect list only exacerbates things. Even if the false positive rate is no worse, it gets multiplied by a much higher volume. When the problem is that people make mistakes, making the same mistakes faster only makes the problem bigger. And of course they'll look a lot like the target. That's what causes the false positive.

AnthonyMouse 9y ago

> I am taking the "cheaper and scale better" part into account when I say that machines will be more accurate. Then you used the wrong word, because cost and throughput are not accuracy. A machine that always returns "true" can be very cheap and very fast. > The false positive rate is whatever you configure it to be. The confidence level for flagging someone is a parameter and you can adjust it to whatever false positive rate is acceptable. When that makes the false negative rate ~100%, you might as well do yourself a favor and throw the entire system in the trash. > Every single flag will be reviewed by a human who will decide whether or not to take action based on it. This is no different than a human looking at a list of suspects and then watching for matches, except that the computer can use a much larger suspect list, and give the human hints. That's the problem though. Getting arrested and subjected to twelve hours of questioning every time you walk in front of a camera will ruin your life even if you ultimately get released every time. The much larger suspect list only exacerbates things. Even if the false positive rate is no worse, it gets multiplied by a much higher volume. When the problem is that people make mistakes, making the same mistakes faster only makes the problem bigger. And of course they'll look a lot like the target. That's what causes the false positive.