> 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.
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.
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.
> 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.
Does the shanghai subway now require IDs to use? That doesn’t sound right, they do spot checks for sure but...foreigners use it also.
And you don’t need to buy a ticket if your RF transit card is filled up, so the point of the kiosk checking ID via facial recognition is even weirder. Maybe they setup that condition for expository reasons? There is no way kiosks are going to screen everyone during rush hour, even if they work perfectly.
I'm pretty sure this is only a prototype in a single location. When I took the subway yesterday, I didn't see any face recognition anywhere. Then again I haven't recharged my transportation card in half a year, so maybe they now require ID for that. If they do, foreigners are probably identified by their passport.
They might be, but face recognition isn’t going to work for that. Is this another case of foreigners waiting in line while Chinese get to use the kiosks (as when taking the train)?
I find the subway so annoying in china. It worked well until they started those damn security checkpoints, after that I just couldn’t justify taking it rather than a taxi.
The points where bags are scanned? It is annoying. It is a security farce. It does only take a few tens of seconds, less than a minute. Unless in peak hours, then perhaps a minute more, where taxis would take 4x to 5x to take where the subway puts you.
It depends on the station, but liangmaqiao Beijing can have waits up to 30 minutes, sometimes even an hour. The lines can be crazy crazy long in ways that can only happen in china.
Of course, traffic will make taxis slower. But it always seemed to work out faster for me unless I had to leave at peak peak, and Beijing traffic makes shanghai’s look reasonable.
No, the Shanghai subway does not need IDs to use, nor do any subways in China. Buying a ticket needs no need for ID, nor does buying a pre-paid RFID-based card require an ID. The bag-check xray is bottle-neck enough.
Taking the train does require an ID card (or passport), as does taking a domestic flight.
Yes, that’s how it was for me in 2016. They do spot check IDs in Beijing at least, but tend to leave foreigners alone. I never got what they would do if a chinese wasn’t carrying their ID, I imagine it isn’t very common to take that on the subway!
It's funny in a disturbing way. NSA and Google use technology in an oppressive way and everyone is up in arms. China develops oppressive technology and journalists cannot stop singing their praise.
I think it's pretty clear people have been up in arms about population monitoring in both the US and China. In fact, if anything, we've been MORE critical of China.
Here is the fundamental problem in a nutshell:
The governments of China and the US simply don't care what we think.
In America, we can do something about it through many avenues. In China, you get 'disappeared' for trying to do something about it. A bit of a difference.
So you're saying it's impossible to abuse USA terrorist safeguards to make people 'disappear' ( in terms of infinite prison time or destroying their social status)? I think the term freedom in debating democratic countries vs. totalitatian countries is a joke, there is always someone with more power than you and will use it against you if he/she can profit from it.
Perhaps not, but if abuse like that happens there are avenues to fight it. For example, see all the BLM protests regarding police brutality. In China, there is not even a need of pretense. If you criticize the government, you are gone. The BLM equivalent in China, the student uprising, was literally squashed with tanks. So, to conflate our two systems and say the difference in freedom is non-existent is the joke.
House of cards is a pretty accurate depiction of U.S. politics,there are many avenues for one to be “disappeared” in America too. It’s not so much different
Chinese AI "rewiring" the world is a terrifying notion, as I'm sure by mandatory design, China's version of Asimov's laws would basically put obeisance to the party and big daddy XI as their primary tenets.
If China won't respect human rights, I would support moving away from them as a trade partner. There are other producers with more liberty for their people.
Instead, China is viewed as an opportunity (to be exploited).
Apple, Dell, IBM, everyone, used them as cheap labor to get manufacturing costs down. So that they can sell that sweet iPhone for $1000, and make $800 in profits. Apple stock goes up. Your 401k is more valuable.
Then, there are more iPhone users in China, than there are in the United States. Apple can make more money selling services (music, apps, subscriptions) to them, than they do in the United States. Do you want to take away this sweet money tree from Apple?
Walmart views China as a cheap manufacturer. You can buy more plastic trinkets for less. So you spend more money, to buy more things. Win for you. Win for Walmart. Walmart stock goes up. Your 401k is more valuable.
If you want to kill the lucrative trade partnership with China, then you had better give a good explanation to all the people with their cushy 401k, why their retirement accounts are tanking.
The trade with China, may suck for the blue collar class, but it is very lucrative for the investor class. And for you, your 401k gets a nice bounce, so that you too, can get a share of the profits.
It's funny how many times that article mentions something that a Chinese company is doing that is basically a copy of what a US company has already done. And the only reasons the Chinese company is able to even be close is that China essentially bans US tech companies from participating in their economy, and probably because the government conducts espionage on their behalf.
This is a huge deal, and even though steel tariffs aren't a great solution, I at least appreciate that our government is treating China's behavior as a real threat.
40 comments
[ 3.6 ms ] story [ 80.9 ms ] threadAt 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.
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.
If you don't test something then it's 100% reliable.
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
And you don’t need to buy a ticket if your RF transit card is filled up, so the point of the kiosk checking ID via facial recognition is even weirder. Maybe they setup that condition for expository reasons? There is no way kiosks are going to screen everyone during rush hour, even if they work perfectly.
I find the subway so annoying in china. It worked well until they started those damn security checkpoints, after that I just couldn’t justify taking it rather than a taxi.
The points where bags are scanned? It is annoying. It is a security farce. It does only take a few tens of seconds, less than a minute. Unless in peak hours, then perhaps a minute more, where taxis would take 4x to 5x to take where the subway puts you.
Of course, traffic will make taxis slower. But it always seemed to work out faster for me unless I had to leave at peak peak, and Beijing traffic makes shanghai’s look reasonable.
Taking the train does require an ID card (or passport), as does taking a domestic flight.
Here is the fundamental problem in a nutshell:
The governments of China and the US simply don't care what we think.
That's the problem we need to solve.
Made it past the censors. I guess there are _some_ differences.
Your initial premise is incorrect, I'm ignoring the rest of your comment
Instead, China is viewed as an opportunity (to be exploited).
Apple, Dell, IBM, everyone, used them as cheap labor to get manufacturing costs down. So that they can sell that sweet iPhone for $1000, and make $800 in profits. Apple stock goes up. Your 401k is more valuable.
Then, there are more iPhone users in China, than there are in the United States. Apple can make more money selling services (music, apps, subscriptions) to them, than they do in the United States. Do you want to take away this sweet money tree from Apple?
Walmart views China as a cheap manufacturer. You can buy more plastic trinkets for less. So you spend more money, to buy more things. Win for you. Win for Walmart. Walmart stock goes up. Your 401k is more valuable.
If you want to kill the lucrative trade partnership with China, then you had better give a good explanation to all the people with their cushy 401k, why their retirement accounts are tanking.
The trade with China, may suck for the blue collar class, but it is very lucrative for the investor class. And for you, your 401k gets a nice bounce, so that you too, can get a share of the profits.
This is a huge deal, and even though steel tariffs aren't a great solution, I at least appreciate that our government is treating China's behavior as a real threat.
"steel tariffs aren't a great solution"