4 ms·

> More impressive still, all this happens successfully in the middle of a crowded, noisy station. Each kiosk has to figure out who is speaking to it; zero in on

> More impressive still, all this happens successfully in the middle of a crowded, noisy station. Each kiosk has to figure out who is speaking to it; zero in on that person’s voice within the crowd; transcribe the incoming speech; parse its meaning; and compare the person’s face against a massive database of photos—all within a few seconds. At the risk of sounding cynical, what is more impressive is that even if the kiosk either mis-identifies a voice or zeroes-in on the wrong person or mis-transcribes the incoming speech or parses incorrectly or cannot compare a person face or unable to connect to the database or identification takes more than a minute - no one will ever ever hear about it. The person complaining might get some points from his social score or go to jail. But, we will get an article from MIT making it sound like everything works flawlessly.

mark_l_watson 9y ago
It is probably a prototype system. That said, this is the way you make progress: you build something, and see if it performs well enough, meeting people’s needs. Andrew Ng had a great demo of detecting one voice in a party environment in one of his old classes. It was a matlab/octave one-liner. Here in the US, we have had some fine successes by organizations like DARPA ‘going long’ and taking funding risks. Now the Chinese are doing the same.
- tibbetts 9y ago
  DARPA and friends should go long. But it would be nice if journalists didn’t act like problems were completely solved the first time they see a prototype. That kind of hype is how you get AI winter.
  - jjoonathan 9y ago
    How do we disincentivise clickbait?
    - njarboe 9y ago
      Stop clicking? Teach others to do the same.
      - jjoonathan 9y ago
        That's what I'm doing, but I don't expect it to work.
taneq 9y ago
Man with clock knows what time it is. Man with two clocks never quite sure. If you don't test something then it's 100% reliable.
cscurmudgeon 9y ago
In Soviet China, AI measures your performance.
wenc 9y ago
> But, we will get an article from MIT making it sound like everything works flawlessly. Well, I don't want to speak to the draconian political implications of this -- that is a separate issue with troubling implications -- but I do want to address the general notion that imperfection should prevent a system (any system) from being practical. Just because it's not perfect doesn't mean it is not usable -- you just need to create rules that govern what actions can be taken. We do this with almost any real world system (vehicle routing, plant control, etc.). No predictive model is perfect, and we practitioners understand this and regulate our behavior accordingly, often through simple rules. We build in checks, overrides and recourses. Us techies often think of technological solutions as being all-encompassing, forget how much of a human element (and human judgment) there is in real-world feedback loops. Also, lest people think parsing out multiple sources from a signal is a new thing, Blind Source Separation [0] is the name that has been given to the study of this type of problem (sometimes called the "cocktail party problem"), and it has been studied for decades. Like most model inversion problems, it's known to be an underdetermined problem, and it is also known that you cannot get perfect separation all the time, but you can get a certain degree of separation if there is some signal (however weak) that helps discrimination. So it's not like practitioners have no idea about the potential issues -- most are acutely aware of them. This doesn't diminish the fact that the algorithms do actually work, for some definition of "work". Techniques like ICA and NMF preceded DL approaches. It sounds like DL (deep learning) approaches hold promise on these types of problems, and this is something of note. [0] https://en.wikipedia.org/wiki/Blind_signal_separation https://en.wikipedia.org/wiki/Blind_signal_separation

mark_l_watson 9y ago

It is probably a prototype system. That said, this is the way you make progress: you build something, and see if it performs well enough, meeting people’s needs. Andrew Ng had a great demo of detecting one voice in a party environment in one of his old classes. It was a matlab/octave one-liner. Here in the US, we have had some fine successes by organizations like DARPA ‘going long’ and taking funding risks. Now the Chinese are doing the same.

tibbetts 9y ago
DARPA and friends should go long. But it would be nice if journalists didn’t act like problems were completely solved the first time they see a prototype. That kind of hype is how you get AI winter.
- jjoonathan 9y ago
  How do we disincentivise clickbait?
  - njarboe 9y ago
    Stop clicking? Teach others to do the same.
    - jjoonathan 9y ago
      That's what I'm doing, but I don't expect it to work.

taneq 9y ago

Man with clock knows what time it is. Man with two clocks never quite sure. If you don't test something then it's 100% reliable.

cscurmudgeon 9y ago

In Soviet China, AI measures your performance.

wenc 9y ago

> But, we will get an article from MIT making it sound like everything works flawlessly. Well, I don't want to speak to the draconian political implications of this -- that is a separate issue with troubling implications -- but I do want to address the general notion that imperfection should prevent a system (any system) from being practical. Just because it's not perfect doesn't mean it is not usable -- you just need to create rules that govern what actions can be taken. We do this with almost any real world system (vehicle routing, plant control, etc.). No predictive model is perfect, and we practitioners understand this and regulate our behavior accordingly, often through simple rules. We build in checks, overrides and recourses. Us techies often think of technological solutions as being all-encompassing, forget how much of a human element (and human judgment) there is in real-world feedback loops. Also, lest people think parsing out multiple sources from a signal is a new thing, Blind Source Separation [0] is the name that has been given to the study of this type of problem (sometimes called the "cocktail party problem"), and it has been studied for decades. Like most model inversion problems, it's known to be an underdetermined problem, and it is also known that you cannot get perfect separation all the time, but you can get a certain degree of separation if there is some signal (however weak) that helps discrimination. So it's not like practitioners have no idea about the potential issues -- most are acutely aware of them. This doesn't diminish the fact that the algorithms do actually work, for some definition of "work". Techniques like ICA and NMF preceded DL approaches. It sounds like DL (deep learning) approaches hold promise on these types of problems, and this is something of note. [0] https://en.wikipedia.org/wiki/Blind_signal_separation https://en.wikipedia.org/wiki/Blind_signal_separation