Facial recognition is still pretty shit

Amazon’s facial recognition technology falsely identified 28 members of Congress as people who have been arrested for crimes, according to the American Civil Liberties Union (ACLU).

The ACLU of Northern California’s test of Amazon’s controversial Rekognition software also found that people of color were disproportionately misidentified in a mugshot database, raising new concerns about racial bias and the potential for abuse by law enforcement.

There’s no AI – as yet – as good as a human at recognising a face. It’s a difficult and complex problem.

Of course, 100% of members of Congress should be identified as arrestees at least in a just world but that’s another matter.

19 thoughts on “Facial recognition is still pretty shit”

  1. people of color were disproportionately misidentified in a mugshot database

    Never mind racist computers. How do they find missing persons in China?

    “He’s about five feet tall, with straight black hair…”

  2. It is remarkably difficult to photograph ‘people of colour’, i.e. dark-skinned, because automatic camera settings tend to respond better to the hues in lighter skin and therefore give better feature definition. It seems to me that the misidentification of dark-skinned people is largely due to technology, rather than racial bias.

    Using flash is some help, but – as my experience, such as it is, is photographing people in hi-visibility clothing, – the nett result is to obscure the features entirely!

  3. South Africa’s identity documents and passports operate on a computerised biometric system. One of the big banks is now offering to open bank accounts on the basis of a selfie sent by smartphone. This will be a lovely experiment to watch (as long as one is not a shareholder).

  4. After almost everybody in the UK has had acid thrown into his or her face–two years?- the technology will be no use anyway.

  5. I’ve often argued computers are crap and processing visual information and humans much, much better but this stands in the way of people’s dreams of self-driving cars so I get laughed at.

  6. In other new, the ACLU reports that the Wright flyer is incapable of crossing the Atlantic, or of carrying 100 passengers, so the whole Aircraft thing is useless and all flying trials should be stopped.
    Sigh.
    Face recognition is an interesting AI problem and will get better with time, provided people can monetise the benefit.
    No technology is perfect on day 1.

    A more relevant concern would be: suppose it started working perfectly?

  7. Pingback: Computers and Visual Data | White Sun of the Desert

  8. Computers are stupid. Remarkably so.
    Teach it to catch a tennis ball and it will, better than a person. Lob a cricket ball, and it’ll just sit there.

    Obviously the programmes are getting better, but computers will never have the intuition, imagination and problem solving capabilities of a human.

    Cos all they ever do is add 0 and 0, 0 and 1, or 1 and 1.

    Pattern recognition is difficult, and it’s something we do without thinking.

  9. 100% of members of Congress should be arrested at least in a just world

    But the damn computers will still refuse to recognise 73.62%* of them …

    * Or some other large fraction.

  10. Bloke in Costa Rica

    The “Congressional Black Caucus” is just another way of saying “gang of unindicted felons”. And the rest of ’em are little better. It’s long been known that the absolutely best investment strategy known to mankind is to be a member of Congress.

  11. Tim Newman,

    “I’ve often argued computers are crap and processing visual information and humans much, much better but this stands in the way of people’s dreams of self-driving cars so I get laughed at.”

    Pretty much.

    Image recognition is mostly just about patterns. It doesn’t know what a dog is, it just knows that an image has certain patterns and is called “dog”. It doesn’t know what a cat is. It knows another image has certain patterns and is called “cat”. You get tens of thousands of those images, it weights certain patterns as more “cat” than “dog”. So when you give it an image, it checks the patterns and then tells you it’s more likely a dog or a cat.

    For some needs, that’s good enough. I did some fiddling with ANPR for a company for license plate recognition. The idea was that if an unknown motorist parked in a car park, it would detect and tell an operator to go check. It was only about 90% accurate. But that was good enough. Solving 90% of the problem is still a huge improvement.

    The problem with self-driving cars is that you can’t get 10% of children running into roads wrong. Or 10% of stop signs.

  12. “Computers are stupid. Remarkably so.”

    Nope. My computers were very smart. I made them so for 31 years.

    But there is an old adage that applies:

    “Computers do what you tell them to do, not what you want them to do.”

  13. I don’t really get the connection between pattern & recognition self driving cars. Humans are very good at pattern recognition. It’s when they try & use that ability to drive cars that they run into problems. They see things that aren’t there & miss seeing things that are.
    Human eyesight isn’t much good for determining distance out beyond a few feet. The eyes are too close together. So distance is judged by relative perceived size. ( a learned process & easily fooled) That’s combined by changing vectors to estimate relative speed. Add a high def field of view that’s not much better than the palm of your hand at arms length & it’s a wonder it’s possible to drive a car at all.
    You get the classic mistakes. An object approaching where the relative bearing doesn’t change will not appear in the peripheral vision until it’s too late to do anything about it. Misjudging distances & speeds because one’s trying to track numerous objects at continually changing angles.
    We’re lousy at perceiving changing accelerations, as well
    You get taught about all this stuff when you learn to fly. But for some reason, not when you learn to drive.

    Any practical self driving car system will have a much better suite of information inputs than a human. Wide angle & accurate distance measurement. Probably radar. The doppler giving relative velocity. Hand-shake capability with other vehicles – probably tied into a central traffic management network – so the car knows the position, vector & intentions of other cars. Thanks to GPS an accurate perception of position mapped to the location. Reaction times a fraction of human reflexes.

    No self driving system can ever be 100% perfect. But it would beat any human driver by a mile. Unless, for some crazy reason, you’re limiting it to the sort of information quality a human tries to cope with. Why would you do that?

  14. Bloke in Costa Rica

    Pattern recognition using AI can often do a lot better than humans. A discriminator tuned using a Generalised Adversarial Network is often capable of doing classification with a better false positive and false negative rate than a person. It’s not even that hard to get basic AI recognition up and running if you know Python and have a bit of spare time. I’ve played with TensorFlow and it’s pretty impressive.

  15. I would note that there is a significant difference between using facial recognition as identification (which has been claimed to fail in the Congress test but see previous posts) and as authentication.

    The difference is that although you need to use Bayseian maths to calculate the success and failure probabilities for both, but in the latter case, it collapses to straight probability calculations. It’s the same with all biometrics (although the difficulties of generating a biometric face map, especially of poc, from a photograph at a distance have been previously described.)

  16. Bloke in North Dorset

    Calling it artificial intelligence sets expectations way too high, its nothing more than comparing one set on binary numbers against another set and giving result a score based on some training against known patterns. OK there’s some clever maths and statistics as SE points but that really is all that’s going on and it certainly isn’t intelligent.

    This set of excellent videos explains how it works for the uninitiated: https://www.youtube.com/watch?v=aircAruvnKk

Leave a Reply

Your email address will not be published. Required fields are marked *