4 ms·

It's not a useful assist if it's misinformation. There's a tendency to trust what comes out of a machine as infallible. I certainly don't want anyone armed ca

It's not a useful assist if it's misinformation. There's a tendency to trust what comes out of a machine as infallible. I certainly don't want anyone armed carrying a badge getting notified in real-time of false-positives for potentially violent, armed, most-wanted criminals. It just becomes a green light for use of excessive force. "The officer feared for his safety, the automated identification system reported the driver with the busted tail light as being the violent drug lord known as El Chapo. Our condolences go out to the substitute teacher's friends and family."

m0zg 7y ago
>> It's not a useful assist if it's misinformation. You're making a typical mistake by thinking these systems will not improve over time as they get more training data. They will, and given enough time, data, and elbow grease they will eventually reach superhuman levels of accuracy. But for them to get there they need to be exposed to the world under realistic circumstances.
- d1zzy 7y ago
  So what we need is a restriction of the officer being able to see the (potentially faulty) results, not a restriction of having facial recognition enabled?
  - m0zg 7y ago
    Serious answer: from the perspective of a deep learning researcher (me), the right answer would be to surface some metric of probability, and have the officer confirm, upon completing secondary validation, whether the prediction was right or wrong. With a few thousand officers doing this over the course of a year (collecting potentially a million training samples or more), the accuracy figures will go up tremendously, and the probability estimates will become far more accurate. That's assuming that the body cam manufacturer has a capable deep learning research team, though, which might not be the case.
- traderjane 7y ago
  Then let another community pay the R&D costs of misidentification of criminals while the tech community tries to grease the joints.

m0zg 7y ago

>> It's not a useful assist if it's misinformation. You're making a typical mistake by thinking these systems will not improve over time as they get more training data. They will, and given enough time, data, and elbow grease they will eventually reach superhuman levels of accuracy. But for them to get there they need to be exposed to the world under realistic circumstances.

d1zzy 7y ago
So what we need is a restriction of the officer being able to see the (potentially faulty) results, not a restriction of having facial recognition enabled?
- m0zg 7y ago
  Serious answer: from the perspective of a deep learning researcher (me), the right answer would be to surface some metric of probability, and have the officer confirm, upon completing secondary validation, whether the prediction was right or wrong. With a few thousand officers doing this over the course of a year (collecting potentially a million training samples or more), the accuracy figures will go up tremendously, and the probability estimates will become far more accurate. That's assuming that the body cam manufacturer has a capable deep learning research team, though, which might not be the case.
traderjane 7y ago
Then let another community pay the R&D costs of misidentification of criminals while the tech community tries to grease the joints.