3 ms·

> I thought that actually was the pairwise number for some algorithms in use. They're designed to produce a list of suspects to investigate, and then if you get

by mlyle 3y ago

> I thought that actually was the pairwise number for some algorithms in use. They're designed to produce a list of suspects to investigate, and then if you get a dozen hits against your surveillance camera photo from your database of a few thousand local mugshots, that's what you're after. You can certainly accept a whole lot more false positives in exchange for a lower false negative rate depending upon application; that's what figure 1 shows. > Doing the numbers this way implies that the error rate would be ~2800% higher if the database contained the entire population, even under these conditions. Disagree. Error rates with 1 million people aren't anywhere near 1/12th; they're perhaps half. It's nowhere near a linear relationship. Also, if you had everyone in the database, you wouldn't have lots of people presenting who aren't in the database, which is where most of the false matches presented. > But even supposing that this is a problem, wouldn't it still be better to ban facial recognition databases than news reporting? Don't take my disagreement on the magnitude of the threat of facial recognition as supporting something else.

AnthonyMouse 3y ago
> Disagree. Error rates with 1 million people aren't anywhere near 1/12th; they're perhaps half. It's nowhere near a linear relationship. Is this only because of the trade off between false positives and false negatives? To avoid raising the false positive rate excessively you could accept a higher false negative rate, but how much does one rate change if you hold the other constant? > Also, if you had everyone in the database, you wouldn't have lots of people presenting who aren't in the database, which is where most of the false matches presented. This is just a facet of measuring false positives like that: If you're not in the database, the algorithm may be confident that someone who isn't you, is you. If you get added, you may look slightly more like yourself than the other person does and you may end up at rank 1, and then it isn't calling this other person a false positive anymore even though it does consider them as looking enough like you to exceed the threshold for returning a match. But as long as the database isn't fully comprehensive, that means the false positive rate for someone who isn't in the database is going to get higher the more people who are in it, and then anyone would just claim that they aren't in the database. As would be the case for someone who just changed their name. Unless they add themselves to the database under their new name, in which case the opposite happens: The database has Bernie Madoff with your mugshot from 2009, you change your name and go to the Department of Privacy to have your picture taken as Altria Academi, and your recent picture looks more like you do now than your old mugshot so Altria Academi comes up as rank 1 and Bernie Madoff comes up as rank 2 and you've got yourself a false negative.

by mlyle 3y ago

AnthonyMouse 3y ago

> Disagree. Error rates with 1 million people aren't anywhere near 1/12th; they're perhaps half. It's nowhere near a linear relationship. Is this only because of the trade off between false positives and false negatives? To avoid raising the false positive rate excessively you could accept a higher false negative rate, but how much does one rate change if you hold the other constant? > Also, if you had everyone in the database, you wouldn't have lots of people presenting who aren't in the database, which is where most of the false matches presented. This is just a facet of measuring false positives like that: If you're not in the database, the algorithm may be confident that someone who isn't you, is you. If you get added, you may look slightly more like yourself than the other person does and you may end up at rank 1, and then it isn't calling this other person a false positive anymore even though it does consider them as looking enough like you to exceed the threshold for returning a match. But as long as the database isn't fully comprehensive, that means the false positive rate for someone who isn't in the database is going to get higher the more people who are in it, and then anyone would just claim that they aren't in the database. As would be the case for someone who just changed their name. Unless they add themselves to the database under their new name, in which case the opposite happens: The database has Bernie Madoff with your mugshot from 2009, you change your name and go to the Department of Privacy to have your picture taken as Altria Academi, and your recent picture looks more like you do now than your old mugshot so Altria Academi comes up as rank 1 and Bernie Madoff comes up as rank 2 and you've got yourself a false negative.