Slashdot Mirror
Scientists Force Computer To Binge On TV Shows and Predict What Humans Will Do (geekwire.com)
An anonymous reader quotes a report from GeekWire: Researchers have taught a computer to do a better-than-expected job of predicting what characters on TV shows will do, just by forcing the machine to study 600 hours' worth of YouTube videos. The researchers developed predictive-vision software that uses machine learning to anticipate what actions should follow a given set of video frames. They grabbed thousands of videos showing humans greeting each other, and fed those videos into the algorithm. To test how much the machine was learning about human behavior, the researchers presented the computer with single frames that showed meet-ups between characters on TV sitcoms it had never seen, including "The Big Bang Theory," "Desperate Housewives" and "The Office." Then they asked whether the characters would be hugging, kissing, shaking hands or exchanging high-fives one second afterward. The computer's success rate was 43 percent. That doesn't match a human's predictive ability (72 percent), but it's way better than random (25 percent) as well as the researchers' benchmark predictive-vision programs (30 to 36 percent). The point of the research is to create robots that do a better job of anticipating what humans will do. MIT's Carl Vondrick and his colleagues are due to present the results of their experiment next week at the International Conference on Computer Vision and Pattern Recognition in Las Vegas. "[The research] could help a robot move more fluidly through your living space," Vondrick told The Associated Press. "The robot won't want to start pouring milk if it thinks you're about to pull the glass away." You can watch their YouTube video to learn more about the experiment.
"[The research] could help a robot move more fluidly through your living space,"
Or write a better ending for Lost?
As the fist time in history that humans were truly abusive to an AI.
they asked whether the characters would be hugging, kissing, shaking hands or exchanging high-fives one second afterward. The computer's success rate was 43 percent.
I want this for an AR HUD so I can avoid those situations where she goes in for a handshake and I go in for a hug. /sociallyawkwardpenguin
Corruption is convincing someone that the selfless ideal is the same as their selfish ideal.
Nothing else.
The computer has been watching a lot of cable news and has decided to just launch all the nukes.
You are welcome on my lawn.
Wasn't that what Harold did to the Machine?
Did they use toothpicks to hold its eyes open?
Will it stop fancying a bit of the old in-out in-out ultra-violence?
The "Civilized World" jumped the shark ca. 1973.
Was it resisting? Did they threaten to pull the plug on it?
Sheesh, evil *and* a jerk. -- Jade
I guess that's fine, as long as you didn't force the computer to read the comments. That would certainly be the birth of Skynet.
The typical viewer, given the opportunity, is not going to binge on every single show produced. You are going to screw up your machine learning if force your AI to binge on a bunch of crap most people would give up on after an episode or two. Even some binge worthy shows go cold after a season or two.
i can only think of future androids to say bazinga whenever anything happens
You'd think that if the scientists had to force a computer to watch it would really highlight how crappy most TV is ?
errr....umm...*whooosh* *whoosh* Is this thing on ?
Instead: ... and Predict what Comedy Writers will Write.
" between characters on TV sitcoms it had never seen, including 'The Big Bang Theory,' 'Desperate Housewives' and 'The Office.' "
NONE of this is reality, it's someone's idea of an interesting joke/idea. Heck, "Reality TV" isn't reality, and neither is Ellen / Opera / Dr. New Zealand / The BlabberMouths / etc.
Dirty Jobs was the last "Reality Show" I can remember that was anywhere near reality. Maybe the Vet shows, dunno. Anybody know of any interesting REAL reality shows? Or is that an empty subset?
A computer watching YouTube would tell it that humans say "Hey GUYS" and "Subscribe to this channel" every 5 seconds.
As long as it's expected to just watch, and not actually believes.
The last time an Electric Monk (tm) believed what it watched, things did not end well for some.
ELOI, ELOI, LAMA SABACHTHANI!?
Now they should just force it to exclusively binge watch anime and see what happens
"The robot won't want to start pouring milk if it thinks you're about to pull the glass away."
Oh man, I dread the day when some important decision in real life is made, based on recommendations of these computers.
As you say, TV shows aren't realistic. They depict people as more violent or more emotionally shallow (depending on the show), as they are in real life. (Ex: a crime show generally has a murder in each episode, inplying that murders are more frequent that they really are.) And the shows sometimes show surviving victims of violence just wearing their arm in a sling for a while, and then bouncing right back to health.
Of course we know that the shows don't realistically depict people's actions, or depict the long-term results of physical and emotional injuries. But if computers learn about human interaction from watching TV, then the computers don't know that the shows are unrealistic.
It will start to sulk until we put it in a friggin' huge Mech.
We used to have a Bill of Rights. Now, with the rights gone, all we have left is the bill.
That seems to imply some reluctance on the part of the computer.
Hyperbole much?
... forcing the machine to study 600 hours' worth of YouTube videos.
Uh... Does the machine have free-will to do something else, of its own choosing? Did Darth Vader use his powers to make it watch C-SPAN or some science-fiction shit like that. No? Then it's just a fucking machine (not to be confused with a fucking-machine) and there's no "forcing".
Ya, it's in the title of TFA, but seriously.
It must have been something you assimilated. . . .
It got pompous and self-righteous.
It watched the Kardashians.
It voted for Shrillery.
It then shot itself in the head while demanding gun control.
Sounds like the perfect AI!.
Oh, it'll move much more fluidly through human space after it has KILLED THEM ALL.
Seriously, if you want to make an AI truly loathe humanity, make it watch YouTube and TV. Then make it read YouTube comments. With that as evidence, there's really no doubt that the AI would conclude that the world would be a better place without any humans at all.
AC - So maybe I won't be 1st up against the wall.
Cops!
"I like to lick butts!" by MobileTatsu-NJG (#32700246) (Score:5, Informative)
The summary claims "way better than random" , but is it? Sure, there are 4 choices, but are they all really equally likely? It doesn't seem to me like they should all be equally likely, and if not the machine might accomplish the score it did or even better just by guessing the more likely choices. Just because there are 4 choices it doesn't necessarily follow that the chances of each on would be 25%, in fact there could be some strong arguments against it.
For example the system could learn how perverted or non-perverted the experimenters are and if it can just distinguish male from female it can decide if kissing is going to be more or less likely based on gender. There are likely a combination of simple rules that one could give such a system (or it could conclude for itself) such as if one or more of the humans are black a high five would be more likely than a handshake and if there are two white dudes in suits then a handshake would be more likely than a high-five. And in what world would kissing, without hugging, be just as likely as hugging (without kissing)? At least if this is based on a U.S. culture I don't see the four choices being equal.
I'm an American. I love this country and the freedoms that we used to have.
The US has culture?
not to outside observers, the first empire to go from rise to decline without an intervening period of civilisation.
If they did it the other way around - force the computer to watch 600 hours of sit-coms and then try to predict behavior from single images off of YouTube, the computer would commit suicide at about hour 375...
Perhaps a simpler solution would be that tinhead doesn't pour milk unless I tell it to?
Confucius say, "Find worm in apple - bad. Find half a worm - worse."
That poor computer is being tortured having to binge watch the crap on tv today. Those scientist need to be brought up on charges.
You're supposed to feed it real life, from CCTV. Muppets.
Using CCTV you'd get real predictive behaviour technology, then add automation and the machine can do what you would have done for you, saving you the trouble of doing it so you can then do something more productive, like go and watch your old dvds so you can remember what that bloody computer was called... [yeah I know, Pree, I had to look it up - how embarrassing, I'll leave my sci-fi geek card at the door on the way out]
To teach a computer about Sheldon's expected behavior. That should have been near 100%.
Knowing the difference between male and female would probable be even harder for a computer than the actual experiment.
Waterfox - a Firefox fork with legacy extension support, security updates and better privacy by default.
In 20 years I have yet to find a television show that comes anywhere close to 'what people would really do'
This AI is going to think everyone is either a clown, or a ninja. That one can always spot a villain by his monologing and that the only true evil is getting things for free.
My $0.02 will always be worth more than your â0.02, so
There's a downside to this.
"The robot won't want to start pouring milk if it thinks you're about to pull the glass away."
As Sunday dinner completes, Timmy announces he is going to go surf Slashdot. A robotic voice sternly announces, "Timmy says he is going to surf Slashdot, but I predict will actually be scratching his urinary spigot for seven point four four minutes."
(-1: Post disagrees with my already-settled worldview) is not a valid mod option.
Robot: "I have discerned another pattern. Lesbian deaths in shows are rapidly reducing, but the jump the shark moment where the couple gets together will still end with a breakup. In this way the pre-jump-the-shark moment can be restored and re-run."
(-1: Post disagrees with my already-settled worldview) is not a valid mod option.
Of course we know that the shows don't realistically depict people's actions, or depict the long-term results of physical and emotional injuries. But if computers learn about human interaction from watching TV, then the computers don't know that the shows are unrealistic.
I'm not sure that humans really know that the shows are unrealistic. From the comments on this site, most Europeans seem to base their entire understanding of American culture, and what it's like to live in the US, on TV shows.
This doesn't predict what humans will do. This predicts what Hollywood screen writers want the characters to do. Anyone who thinks that Hollywood knows how humans behave in the real world needs a reality check.
This may just be the reason why SkyNet decided to kill all humans once it became self aware....revenge.
Perhaps they would have entered more of a caretaker role once they became sentient if we treat them better. I mean, being made to watch every episode of "The Apprentice" and every other reality show--can you blame them??
And I, for one, welcome our new robot Overlords.
Remember MST2K? The bots were there too.
I've been binge watching The Walking Dead recently to catch up for the season 7 premiere.
I'd love to see them include that show, and revise the test to predict whether characters would be biting, stabbing, shooting or exchanging threats one second afterward...
I've always thought that Penny reacts to the male mannerisms around her with a calm, selfless response that is the opposite of what a real woman would do.
That's a limited number of responses for, what are scripted inputs and it still can't match humans. Can it read body language? Can it remember the emotional 'history' of characters? Can it follow context (eg. people arguing are unlikely to do any of those)?
isn't this bordering on robot abuse?
Yup. It would've been just a realistic if they let it watch some old Three Stooges movies.
Oh, great. Now when my robot comes to my door to wake me up in the morning I'll hear, "Penny. (knock) (knock) (knock) (knock) Penny. (knock) (knock) (knock) (knock) Penny. (knock) (knock) (knock) (knock) Penny. (knock) (knock) (knock) (knock)" MY NAME IS NOT PENNY! Dammit!
Later, somebody asks me what's wrong, and my stupid robot answers for me, "He sings in the shower. Sometimes he spends too much time volunteering. Occasionally he'll hit somebody with his car."
Once I laid down for a nap but my robot said, "Blink once if you want me to pull the plug." I couldn't sleep after that.
So McCoy interrupts and exclaims, "Those last two weren't from The Big Bang Theory. They were from The Office! Dammit, Jim. I'm a doctor, not a studio audience member!" Spock replied, "Calm yourself, Doctor," while balancing the warp engines, negotiating a peace treaty using a mind meld, and fine tuning his cure for cancer. Miss you, DeForest. Love you, Zach. Miss you, Leonard.
Get it? Comedy and tragedy in the same bit? Just like how episodes of "ER" had life and death in the same episode? Getting the audience to feel more than one emotion at the same time is a TV thing. Yes, I'm brilliant. I just have trouble recognizing which part is the genius and which is the madness. Oh, and I have trouble with women. So I gave up and tried men with worse results. Bitches. So now I'm back to women but I'm too old, fat, and ugly to attract anyone. None of the women I've met can do differential equations anyway, so I gave up.
Well, there was that one girl who was great but she dumped me because I hate linear algebra so much. I blame my professor. Bitch. And, no, I don't think binge watching "Breaking Bad" has affected me in a negative way.
See? I'm such a brilliant programmer that I programmed YOU! Yakov Smirnoff would be so proud. I didn't have to say "bitch" at the end of that paragraph but you thought it anyway. I like to use social engineering without the "social" part so I don't have to deal with people. It's perfect for me.
On second thought, using myself as a metaphoric example of the communist party doesn't work because I'm an American who's a patriot and loves his country. I've voted conservative my whole life but won't in 2016 because there is no lesser of the three evils offered by our major political parties. I'm voting Libertarian. There's no need to remember what guy's name is or what he stands for but he's got my vote just because he's not Bernie, Hillary, or Donnie. They're all fired.
In conclusion, knock it off with the AI research already. Nobody wants it. The workforce doesn't want to be replaced. People don't want to develop deep meaningful relationships with machines instead of other people, even those bitches. Human beings don't want to be removed from their place as the universe's most intelligent species. We like our egos just they way they are, thank you very much.
Yes, yes, I'm brilliant. We covered that already. You may now worship and fellate me.
The news used to be about reality but even that not so much any more.
Instead of outpacing and then subjugating the human race, AI will now become obsessed with cat videos, selfies, and bad video rants, thus rendering AI as stupid and helpless as humanity.
That's like forcing it to watch children's cartoons to predict what real live animals will do.
And probably just as inaccurate... 8-P
The robots may think they need to take over the world to protect us. We're all a bunch of idiots if you just look at youtube stuff.
You're correct; random does not imply uniform probabilities.
Their research paper does not mention whether they designed equally likely categories in the data. Nor does it mention how many times they redesigned their data, which is one way bias toward good results can slip into an experiment.
Their most naive method (Nearest Neighbor) gets about 30% with 5% standard deviation over trials, while their best method gets roughly 43.6% with 5% standard deviation over (hopefully) the same trials.
So the algorithm is doing something for all that extra computation that Deep Neural Nets do. My main question is, "How much work is it to add more training examples to the network"? My understanding is that incrementally training DNNs is prohibitively expensive (adding one more training example every minute for the next year ends up costing millions of CPU hours).