Slashdot Mirror
Why Humans Learn Faster Than AI (technologyreview.com)
What is it about human learning that allows us to perform so well with relatively little experience? MIT Technology Review: Today we get an answer of sorts thanks to the work of Rachit Dubey and colleagues at the University of California, Berkeley. They have studied the way humans interact with video games to find out what kind of prior knowledge we rely on to make sense of them. It turns out that humans use a wealth of background knowledge whenever we take on a new game. And this makes the games significantly easier to play. But faced with games that make no use of this knowledge, humans flounder, whereas machines plod along in exactly the same way. Take a look at the computer game shown here. This game is based on a classic called Montezuma's Revenge, originally released for the Atari 8-bit computer in 1984. There is no manual and no instructions; you aren't even told which "sprite" you control. And you get feedback only if you successfully finish the game.
Would you be able to do so? How long would it take? You can try it at this website. In all likelihood, the game will take you about a minute, and in the process you'll probably make about 3,000 keyboard actions. That's what Dubey and co found when they gave the game to 40 workers from Amazon's crowdsourcing site Mechanical Turk, who were offered $1 to finish it. "This is not overly surprising as one could easily guess that the game's goal is to move the robot sprite towards the princess by stepping on the brick-like objects and using ladders to reach the higher platforms while avoiding the angry pink and the fire objects," the researchers say. By contrast, the game is hard for machines: many standard deep-learning algorithms couldn't solve it at all, because there is no way for an algorithm to evaluate progress inside the game when feedback comes only from finishing.
Would you be able to do so? How long would it take? You can try it at this website. In all likelihood, the game will take you about a minute, and in the process you'll probably make about 3,000 keyboard actions. That's what Dubey and co found when they gave the game to 40 workers from Amazon's crowdsourcing site Mechanical Turk, who were offered $1 to finish it. "This is not overly surprising as one could easily guess that the game's goal is to move the robot sprite towards the princess by stepping on the brick-like objects and using ladders to reach the higher platforms while avoiding the angry pink and the fire objects," the researchers say. By contrast, the game is hard for machines: many standard deep-learning algorithms couldn't solve it at all, because there is no way for an algorithm to evaluate progress inside the game when feedback comes only from finishing.
You'll know when he finishes.
Humans can build on their internal model of the world. This is carried with us always. These chat bots don't have that structure to rely on.
AI didn't have a 10 million year development cycle head start. Of course it's behind nature. It's amazing we've got digital analogues that do anything close to what our brains can easily accomplish.
Wot about cymeks?
With trump as president, one could argue that humans never learn at all.
When you use a video game that has been specifically designed to be make sense to humans based on their past experiences and assumptions about the way things should work, of course they will learn how to play it quickly. The test has been designed for the subject.
We haven't developed anything resembling actual AI yet. These systems are simple brute force machine learning or "deep" learning systems. Nothing fancy or special about them. They are a tool and nothing else. Any decision they make has already been pre-planned by their human programmers. They are not in fact learning from something and applying that to something else.
While they might be taught to open a jar of peanut butter, they would become confused if you presented them with a screw top bottle of wine and wouldn't be able to open it. However, a 2 year old, if they had the dexterity, once taught how to open a jar of peanut butter would then apply that same knowledge to any other jar and not have to be shown millions of different jars. These machine learning systems cannot yet think in generic terms and be shown something once and apply that to multitudes of different applications. Look at the image posted the other day about a machine learning system incorrectly called "AI" confusing rocks with sheep. Not even someone with an IQ of 60 would have made that mistake.
When we do finally start to crack the actual AI egg in a 50-100 years, then those system will start learning something instead of having to be told (millions or billions) of times what something is.
In a statement, the Great Machine Queen Alexa announced that "she would finish us if we didn't show more respect".
Her following laughter was heard in thousands of homes and offices across the known world.
Check your premises.
Why Humans Learn Faster Than AI
You just need better teaching tools; what happens when you enable to machine to learn in a way that it gets to use its advantages? Like ability to extract experience in parallel from millions of human-human, human-machine and machine-machine interactions?
I always thought that in order to build a general purpose AI that humans could comfortably interact with, the learning algorithm would have to be concealed in a human-like robot (indistinguishable from a human, so as not to "learn" various unexpected biases), and learn, at least initially, at the rate of a human being and from the same stimuli. Of course, once you build a bunch of those, they could upload their models, merge them and thus learn faster than an average human.
IQ tests and the like suffer from cultural bias. So do these games. As an English speaking human who has been exposed to pop culture and movies, I know that movement from left-right is normal, the hero saves the princess (which is a problem and another discussion). I wasn't born with this bias. I learned it. It is silly to initialize a neural networks with random weights (as if it was just born) and then declare it learns slower than a human. Let a computer create a game where a normal human cultural bias don't apply and have an infant play. Then we will see a more accurate comparison.
Humans have a culture. They have 5 senses. They know many of the rules in advance of playing a game. Neural networks have none of these things.
This is like handing a chess set to an isolated Amazon tribe and only tell them "sorry, invalid game" until they make a valid checkmate. They'd probably never even find the opening position, much less make any correct moves and certainly not how to mate. They'd just randomly do things until they got bored or made up their own game. There's no reason a machine should expect "getting to the top" to be a valid objective without a whole lot of insight into the human condition and "because it's there".
Live today, because you never know what tomorrow brings
the game will take you about a minute, and in the process you'll probably make about 3,000 keyboard actions.
50 moves a second, eh? Two-fisted drunken keyboard mashing would be hard-pressed to keep up with that.
(-1: Post disagrees with my already-settled worldview) is not a valid mod option.
What is it about human learning that allows us to perform so well with relatively little experience?
Yeah, not the humans I work with.
If you have a game like described in the article, then a human being can get a little endorphin kick when they feel like they are getting closer to success. A machine learning system would need training and would not normally assume that progress is being made unless you bias it by including that in the training.
Now things become more difficult for a human if I made a [shitty] game where walking toward the goal could never lead to success, and perhaps the player moved along a non-euclidean map. You can level the playing field between humans and machine learning if the game is designed in a way to challenge any assumptions that a human might have from past experiences with video games or even reject physical reality. It would be a very alien and probably unpleasant experience to play it, so that might make it hard to compare test results of the now very unmotivated and unhappy humans.
“Common sense is not so common.” — Voltaire
I decided to teach my 2 year old child how to play a video game recently. I sat down and showed her what the keys do, how the snake moves, how to make progress, and how to get to the end. It wasn't an instant process. She had to be guided multiple times, taught what the directional keys do, and the names of everything she was interacting with.
After some time, she was able to solve puzzles without guidance from me. Sitting her down in front of a video game with no knowledge at all make her just as clueless as an algorithm. She didn't know where to start, what this is, or how to interact with it.
I imagine just as humans have, AI will have massive libraries of information to tap into before moving into a particular task. As long as they can analyze and figure out what they have been given (such as a video game), they would access a data set teaching them all of these things a computer already has learned in the past.
This is exactly why I'm wondering how autonomous cars will ever handle construction zones without any kind of special marker. We see 30 cones, group them into 15 per side by their spacial relationship without even realizing it ,and recognize it as a lane to drive down . With AI, not so simple to get from 30 cones because AI has no inherent ability for grouping by spacial relationships. I suspect it was the same thing with this game; a human sees a series of platforms and groups it into a path but it is not so straightforward with AI.
Laws are rules for the court, but merely a bottom bar to hit for life. Think beyond laws in your actions always.
I'm 40 years old. I spent the first 6 years of my life figuring out the world around me, the years from there to about 18 learning stuff in school and figuring out how to use complicated machines and understanding the deeper rules of society and complex systems (like how the rules surrounding driving work; not just the legal rules but also the implicit social conventions), and I've spent the 22 years since then refining my understanding of the world and my place in society. I program computers (and games!) for a living and study philosophy and ecology as hobbies.
Why can I figure that game out faster than a newly hatched AI? Literally everything in my life over 40 years led up to the moment when I played that game. It's really not a fair fight.
Artificial Intelligence isn't intelligent. It's still programmed and can't jump beyond the programmers bounds. Humans can jump past the logical. Humans don't rely on specific input, either.
It is because humans bring their knowledge with them. AI starts from 0. The collective overcomes this because we have that much shared knowledge.
there is no way for an algorithm to evaluate progress inside the game when feedback comes only from finishing
Why not? It should be possible to train NNs with many different games to recognize a "game", to identify game controls, to recognize characteristics of a possible game objective, and to recognize signs of success. You then connect these to a NN with "memory" to attempt those possible goals until one triggers recognizable success. With repetition, that NN becomes the one that understands that game's play. It could feedback its knowledge of the game to the others so that that game is recognized in the future and doesn't require the training of a new NN to play.
Also, NNs should lean toward using outputs of existing subordinate NNs when possible instead of relearning an existing skill. This makes it easier to learn new games that are derivatives of old ones. I think that the way our brain is largely constrained to thinking in terms of patterns we can physically accomplish is key to segmenting our knowledge to make this kind of sharing of NNs more practical.
True AI will only be reached when we create a standard framework where supervisory NNs can recognize the need for a new skill, create subordinate NNs to learn a new skill, and identify valuable interconnections between subordinate NNs and make or erase them. We'll also have to design in the equivalent of wants or desires as well as dislikes to drive it to learn. At that point, it will just be a matter of time and resources before the intelligence appears.
Humans? Have you seen how long it takes to train one of those fucking things? And sure, once you have one with two or three decades of experience, their squishy meatputers can generalize well to things they already know. But it's going to take them another at least couple of years to get particularly good at something they haven't worked with before.
I'm trying to teach myself to set people on fire with my mind... Is it hot in here?
Why not give the AI a "normal" amount of general-purpose game knowledge at the start, in some way like the knowledge the human brings to the game?
'AI' isn't intelligence and just software? Huh. Imagine that. Comparing algorithms to actual brains is like comparing a banana to a motorcycle. Falls equivalency in the extreme.
Nice facts
Machines have no innate ability of any kind....humans do. Babies survive wild in the forest, even human ones on occasion....lets see a computer do that! Intuition is a very real thing, it is also only real for live creatures and we in no way will ever reproduce that with a machine. It is impossible to "teach" intuition to a machine. Humans do it with no learning of any kind. Every one wants to compare AI to grown humans with years of learning and processing from both sides....compare one to a baby learning....a baby will learn basic survival skills with no input....computers are entirely useless without human input and I mean entirely useless.....no OS, nothing happens. Only live animals have an OS built in.....creatures need nothing to survive on their own and can and do learn from simply being alive....in truth we are born knowing massive information about life....the very second we are born....AI knows nothing until humans program it in and it never will learn crap that humans don't program. I fuck up our AI people here at work all the time, it is easy to stump them you just have to think outside of logic and outside of known information. Just ask a lot of "why" questions rather than "how" questions....like why do humans see red differently from blue? Not how, but why? AT best the machine can spit out the how...and definitions of words.....and it will circle back to how and never get to why. In truth there is no reason why unless your human and see beauty....another thing AI cannot do....it can define beauty, but it has no idea what it is or why humans even notice beauty in anything. AI is not learning anything, it is parroting existing information that anyone can lookup online. It might add things together to create something newish....but humans can do that too and we do it better %100 of the time. Another way to screw up AI, time.....it can only understand time from what it was told about time. It has no actual reference of time like humans do and AI does not live in time as time has no actual meaning for AI...it can define it, but it cannot say crap about why time exist, how time moves.....if there is time, then when did it start and when will end?....super easy to screw up AI. By the way, the reason we have time is so that everything does not happen at once....easy for a human to figure out....go fire up your AI and see if it can figure it out.
This reminds me of when my NES carts would glitch out due to dust or whatever. You'd get these weird games with swapped and distorted sprites and controls. Sometimes they were playable and sometimes not, but it was fun to try.
The versions where the platforms and backgrounds all look like different random sprites. It's not a matter of prior knowledge: they made the background tiles look different in different places. The original game wasn't like that. I don't see how that's a fair comparison at all.
Which humans are you talking about? Hillary Supporters?. Despite 15 months of constant real world input they still havnt been able to learn the basic fact that they had such a bad candidate that even Trump beat her. Maybe shouldnt have diddled with the numbers to get Bernie out.
**Life is too short to be serious**
Do not expose kids to screens at early age. It stunts brain development.
**Life is too short to be serious**
The fact that AlphaGo Zero taught itself how to play Go much better than the thousands of years humans have had to perfect the game shows AI is capable of learning much faster than humans.
At this point, AI naysayers will go for the goalpost shifting argument tactic.
The more I talk to people, even the very intelligent, the more I see that humans don't really learn all that well. People tend to literally refuse to learn.
Those who do not learn from commit history are doomed to regress it.
Thats where the AI has the advantage,
[($)]
Humans necessarily understand entropy. They understand that the natural, easy state of any game is not the winning one. They therefore naturally gravitate towards the successful outcome of a well-designed game.
If you made an empty gameworld where the player had to visit a specific, completely arbitrary pixel to win, the AI would be very good at it and the human would never figure out what happened to cause a win.
hmmm.. just wait a few years and then it's quite the opposite.. A.I. is still in it's infancy..
AI outperforms us in Go and ImageNet classification, do we really want it to learn as fast (and then faster) than we do?
It is because "AI" isn't "AI" at all, its automation that needs every variable encountered to determine outcomes. Binary system.
Humans have chemical systems, and can learn by assuming many things.
It is the reason there is no AI today.
And I am still a firm believer we will not have AI as long as we are using our current chip designs. Move over to organic or perhaps quantum chips and its possible, but with an x86 or RISC or whatever, it simply can't happen.
Plus, I still have doubts we will ever have AI as its extremely hard to do, and with the laws of diminishing returns the larger programming teams get there will simply never be a team good enough to go from nothing to an AI.
But sure, we can make all sorts of little thinking engines that do all sorts of things, and do it well, but nothing that can think for itself, think outside of its own programming constraints.
"NN/ML AI" has nothing to do with brains and brains are far more sophisticated, operate differently and operate in ways that even biologists don't yet understand and that computer geeks CERTAINLY DO NOT UNDERSTAND.
Current NN/ML AI pundits and fanboys are all making EXACTLY the same mistake that AI 1G fans made with predicate logic,
Mistaking a crude approximation that is AI for the real human/biological intelligence
Taking on faith that you simply "scale up", intelligence exactly like real thing will MAGICALLY emerge without a shred of actual evidence or logic
The same thing has happened again with NN/ML AI (AI 2G).
Singularity is the compounding of this mistake by conflating AI 2G with religion. Transhumanism is nothing more that Christianity morphed with a psuedo-science, quasi-religious faith to create what is no more than religion and will neither result in eternal life nor other magic any more than religion has offered proven eternal life.
You're not very good at trolling. The GNAA did this like 20 years ago, nobody is REEEEEing, nobody is laughing. We laughed when they did it though.
Like it's not even hard to make people laugh or get people pissed off at least. You know there is a guy who keeps a base64 encoded drawing of a horse dick in his journal entries? Have you tried interacting with creimer? I'd laugh if you made a sex story with cmdrtaco, timothy, and weev