August 18, 2016

"Neville... feels that he has unlocked a mystery of the human mind which could signal a revolution in policing."

"He is frustrated when he encounters skepticism, as if he were claiming to have discovered officers with telepathy or E.S.P., and he feels a nagging suspicion that had he written an algorithm instead the brass would be rushing to embrace it. 'People don’t want to believe that humans could be better than a machine,' he told me. 'And the sad truth in this wicked world we live in is that people don’t want to pay a human. They want to buy a machine.'"

The end of an excellent New Yorker article by Patrick Radden Keefe, "THE DETECTIVES WHO NEVER FORGET A FACE/London’s new squad of 'super-recognizers' could inspire a revolution in policing."


traditionalguy said...

Human Face recognition is the basis of all Celebrity Status. The familiar face triggers emotions and will be paid for like a drug. The Stars are the familiar faces.

Nice to see this generation figure it out too.

Squints said...

The documentary "Married To The Mob" proves this. 8^)

FullMoon said...

Facial recognition test here

Paul Snively said...

"'People don’t want to believe that humans could be better than a machine,' he told me."

Because it isn't true, or rather, when it's true, it's only true temporarily, and that "tempo" is getting faster all the time.

It used to be that, if you wanted a machine to do something "only human beings can do," you had to sit down with a bevy of experts in the domain, interview them, codify their expertise somehow, and program the computer to do what they did. This famously led to the era of expert systems, which did OK for a while, but rapidly exhibited brittleness and, more baffling still to those who developed them, a tendency to simply regurgitate what any 12-year-old could immediately infer from the facts and inference-steps encoded in the system. This was where matters stood in the 1970s and into the 1980s.

Fast forward to today. We're well past the Bayesian revolution, the adoption of Maximum Entropy and Minimum Description Length machine learning systems, and Deep Learning approaches. It's no longer necessary to "try to capture what experts do." It's only necessary to define "success" and "failure," and provide as much training data as possible (in "unsupervised" machine learning, you don't even have to do that).

To sum up, we're at the point where a geek annoyed by his neighbor's cat hanging out in his yard can use a digital camera, a tiny one-board computer, and some open-source software to turn on the sprinkler when there's a cat in the yard. Not people, not other animals. Just cats. More serious hardware and more serious software can do a lot more than this.

JCC said...

That's fascinating. The implications are quite numerous, esepcially for eyewitness testimony in the courts.