Nature: Mastering the game of Go with deep neural networks and tree search
The game of Go has long been viewed as the most challenging of classic games for artificial intelligence owing to its enormous search space and the difficulty of evaluating board positions and moves. Here we introduce a new approach to computer Go that uses ‘value networks’ to evaluate board positions and ‘policy networks’ to select moves. These deep neural networks are trained by a novel combination of supervised learning from human expert games, and reinforcement learning from games of self-play. Without any lookahead search, the neural networks play Go at the level of state-of-the-art Monte Carlo tree search programs that simulate thousands of random games of self-play. We also introduce a new search algorithm that combines Monte Carlo simulation with value and policy networks. Using this search algorithm, our program AlphaGo achieved a 99.8% winning rate against other Go programs, and defeated the human European Go champion by 5 games to 0. This is the first time that a computer program has defeated a human professional player in the full-sized game of Go, a feat previously thought to be at least a decade away.
h/t hoooos
― Karl Malone, Wednesday, 27 January 2016 21:10 (eight years ago) link
i guess this should be the AI thread. post your comments about how AI is impossible because you saw a clip of a robot falling over here.
― Karl Malone, Wednesday, 27 January 2016 21:11 (eight years ago) link
important work they're doing over there *eyeroll*
― Οὖτις, Wednesday, 27 January 2016 21:11 (eight years ago) link
the go thing, you mean?
― Karl Malone, Wednesday, 27 January 2016 21:12 (eight years ago) link
yeah
re: AI in general, I wouldn't say it's impossible but it is very very very far away
― Οὖτις, Wednesday, 27 January 2016 21:12 (eight years ago) link
speaking of neural networks, there's this link caek accidentally posted: http://www.wired.com/2016/01/apple-buys-ai-startup-that-reads-emotions-in-faces
― Karl Malone, Wednesday, 27 January 2016 21:13 (eight years ago) link
it's close enough to figure out how you react to advertisements
I am not impressed
― Οὖτις, Wednesday, 27 January 2016 21:14 (eight years ago) link
I mean congratulations you've spent billions of dollars and tons of other resources on doing something a baby can do, good job
― Οὖτις, Wednesday, 27 January 2016 21:15 (eight years ago) link
(sorry I don't mean "you" you, not trying to make this personal)
― Οὖτις, Wednesday, 27 January 2016 21:16 (eight years ago) link
This is pretty exciting:http://www.eurekalert.org/pub_releases/2016-01/miop-sba012716.php
― schwantz, Wednesday, 27 January 2016 21:17 (eight years ago) link
haha, it's ok
buuuuuuut, when i was a baby, i wasn't capable of reading human emotions from millions of people at any given moment and then feeding that information to advertising corporations. of course, as i grew older i developed this ability but by that time other babies had already submitted job applications so mine was at the bottom of the pile
― Karl Malone, Wednesday, 27 January 2016 21:18 (eight years ago) link
lol
― Οὖτις, Wednesday, 27 January 2016 21:21 (eight years ago) link
but yeah the "reading human emotions" aspect does not impress me as a technological feat in and of itself. Biology still obviously way superior in that department. otoh the "helping corporations make even more effective advertisements!" aspect is just gross and sad.
― Οὖτις, Wednesday, 27 January 2016 21:23 (eight years ago) link
also, i think admitting that certain AI capabilities are similar to what a baby can do suggests enormous potential in the near term. the difference in capabilities of babies and adults seems enormous to us, but when you consider it on a logarithmic scale, they're very close. the difference between einstein and the livestreaming tech guy idiot in oregon is not very large in the grand scheme of things. if an AI's learning curve has already increased from an earthworm to baby level, einstein really isn't that far away.
obviously i'm referring to the scientific names of these universally agreed upon scales here
― Karl Malone, Wednesday, 27 January 2016 21:25 (eight years ago) link
thought the thread bump might be for Minsky
RIP big man
― Brad C., Wednesday, 27 January 2016 21:28 (eight years ago) link
yeah as i understand it the hope/fear is that at some unknown point of basic sophistication the gap all of a sudden closes itself
something i don't get about the superintelligence fear is why these new gods intelligent in ways we can't even imagine are just assumed to also be terminally discompassionate and sociopathically fixated on widget-making or nuclear supremacy
i do sometimes worry about the politics and very notions of intelligence of a lot of the people who do the actual work on this stuff, let alone of course the people who pay for it
rip minsky, yeah.
― denies the existence of dark matter (difficult listening hour), Wednesday, 27 January 2016 21:30 (eight years ago) link
enormous potential in the near term
there's always been enormous potential lol, it's the "near term" part that seems to be constantly pushed out
― Οὖτις, Wednesday, 27 January 2016 21:32 (eight years ago) link
I mean this fear of robots becoming *actually intelligent* and destroying humanity has been around basically since the concept of "robot" was first formalized, well before the first computers even existed.
― Οὖτις, Wednesday, 27 January 2016 21:34 (eight years ago) link
nick bostrom's book is basically about this (a lot of people seem to assume it's a kurzweil style book, but it's really all about risk management). he talks a lot about the end goals of an AI and their unintended consequences. one thing that comes up often is that for just about any goal, having more resources would be beneficial. or eliminating obstacles to the goal (such as humans).
The risks in developing superintelligence include the risk of failure to give it the supergoal of philanthropy. One way in which this could happen is that the creators of the superintelligence decide to build it so that it serves only this select group of humans, rather than humanity in general. Another way for it to happen is that a well-meaning team of programmers make a big mistake in designing its goal system. This could result, to return to the earlier example, in a superintelligence whose top goal is the manufacturing of paperclips, with the consequence that it starts transforming first all of earth and then increasing portions of space into paperclip manufacturing facilities. More subtly, it could result in a superintelligence realizing a state of affairs that we might now judge as desirable but which in fact turns out to be a false utopia, in which things essential to human flourishing have been irreversibly lost. We need to be careful about what we wish for from a superintelligence, because we might get it.
http://www.nickbostrom.com/ethics/ai.html
― Karl Malone, Wednesday, 27 January 2016 21:36 (eight years ago) link
xp ninety years not rly a v long time... in pre-singularity years
― denies the existence of dark matter (difficult listening hour), Wednesday, 27 January 2016 21:37 (eight years ago) link
i guess i am unconvinced that something can simultaneously be "superintelligent" and have an extremely rigid and unadaptable "goal system". people would get bored caring about paperclips, let alone one of these things.
really the problem w the whole line of speculation right is a lack of understanding of what we mean by intelligence let alone superintelligence
― denies the existence of dark matter (difficult listening hour), Wednesday, 27 January 2016 21:40 (eight years ago) link
agree that it could well be an alien brain w v incompatible values, also i suppose agree w the unmade point that the only really altruistic and compassionate thing to do is exterminate us
― denies the existence of dark matter (difficult listening hour), Wednesday, 27 January 2016 21:42 (eight years ago) link
yeah, i think part of the misunderstanding is that everyone tends to anthropomorphize AI. i mean, you're right: people would get extremely bored caring about paperclips. but computers aren't people. they'll switch off between 0 and 1 until they're gone.
agreed about lack of understanding of what the terms mean, though. i read through one of Edge's little collections of smart people talking about stuff, the AI issue, and it was incredibly frustrating because just about every writer seemed to be defining the same terms in different ways.
― Karl Malone, Wednesday, 27 January 2016 21:45 (eight years ago) link
^^otfm
― Οὖτις, Wednesday, 27 January 2016 21:48 (eight years ago) link
humanity's never really gotten around to a good working definition of what "consciousness" is but here we are thinking we have some fancy new way to create it (apart from the old fashioned way of biological reproduction + social engineering), even when we don't know or can't agree on what *it* really is
― Οὖτις, Wednesday, 27 January 2016 21:50 (eight years ago) link
people have been overestimating the proximity of AI for decades, in the sense of an AI as some kind of autonomous problem-solving agent, but maybe to an equal extent underestimating the kinds of intelligence programmers have built to work in specific problem spaces
if you had shown me Google search autocomplete 25 years ago, I don't think my reaction would have been, "Oh that's just an algorithm, where's some real AI?"
― Brad C., Wednesday, 27 January 2016 21:56 (eight years ago) link
obviously there's no denying technological advances. But yeah I don't think I consider what a Google search engine does "intelligence" in any meaningful way, and yeah maybe that is related to it being in the service of a specific, non-autonomous function.
― Οὖτις, Wednesday, 27 January 2016 22:02 (eight years ago) link
it's more like a representation/prediction of group intelligence
though yeah "intelligence" maybe not the word for millions of Google searchers
― Brad C., Wednesday, 27 January 2016 22:03 (eight years ago) link
xposts
i don't know, i guess i don't think that emulating "consciousness" is necessarily essential to a superintelligence. again, the anthropomorphizing thing is a problem. but maybe i'm going too far down the Turing road, thinking that the most important things to measure are outcomes (if an AI can detect human emotions via facial muscle movements more accurately than a human being can and react accordingly, then who cares if it's "conscious" or not?)
― Karl Malone, Wednesday, 27 January 2016 22:13 (eight years ago) link
if an AI can detect human emotions via facial muscle movements more accurately than a human being can and react accordingly
begs the question of what constitutes "more accurately" and "react[ing] accordingly"
― Οὖτις, Wednesday, 27 January 2016 22:17 (eight years ago) link
of the top of my head, for "more accurately", assume that 100 people are asked to film themselves while thinking of either a tragic or sexy memory. a test group of humans then views the videos and guesses the emotions on display, while an AI also completes the same task. the AI is more accurate if they guess the emotion more frequently than the humans do.
"react accordingly": i don't know, i guess it would just be recognizing when someone is sad or angry and backing off for a while (in contemporary Siri terms, maybe holding off on the automated reminder that your toddler's doctor's appointment is tomorrow afternoon). or laughing at a joke. or doing the fake laughter thing you have to do when someone that is respected tells a mediocre joke in a public setting.
it might seem that "consciousness" is necessary to register human emotions and react as humans do, but i'm not sure that's true.
― Karl Malone, Wednesday, 27 January 2016 22:26 (eight years ago) link
dreaming seems to be a signal of consciousness. but does an AI have to dream in order to complete tasks at superhuman levels?
― Karl Malone, Wednesday, 27 January 2016 22:28 (eight years ago) link
do you really want me to deconstruct the problems with your examples cuz they seem obvious to me
― Οὖτις, Wednesday, 27 January 2016 22:28 (eight years ago) link
what if the superintelligent ai says stuff like that all the time
― I expel a minor traveler's flatulence (Sufjan Grafton), Wednesday, 27 January 2016 22:31 (eight years ago) link
oh great thx for outing me
― Οὖτις, Wednesday, 27 January 2016 22:37 (eight years ago) link
better test would record how 100 people react to jute gyte dinner music and compare with ai's reaction
― I expel a minor traveler's flatulence (Sufjan Grafton), Wednesday, 27 January 2016 22:39 (eight years ago) link
xpost i don't know, do what you want i guess
i guess it's likely that i'm just totally misunderstanding what you're saying. i mean, you're saying that emulating "consciousness" is a pre-requisite for superintelligence, right? (earlier you wrote "humanity's never really gotten around to a good working definition of what "consciousness" is but here we are thinking we have some fancy new way to create it"). i'm trying (and failing) to argue that it's not necessary to obtain superintelligent results.
i guess an unstated premise i'm using that others might not agree with is that our brains and computers are already very similar. when i feel "sad" i think that's the result of many stimuli working in concert, leading to my neurons doing what they do.
loool sufjan
― Karl Malone, Wednesday, 27 January 2016 22:42 (eight years ago) link
also, i guess an obvious point, but there are AI research paths like machine learning that aren't trying to emulate the brain's behavior, and certainly aren't trying to create "consciousness"
― Karl Malone, Wednesday, 27 January 2016 22:44 (eight years ago) link
anyway lolz aside I'm gonna take a crack at this cuz it's a slow day at work
assume that 100 people are asked to film themselves while thinking of either a tragic or sexy memory. a test group of humans then views the videos and guesses the emotions on display
this scenario is subject to a lot of problems that plague sociology/psychology experiments, not all of which can be controlled for. Are "tragic" and "sexy" memories actually typically accompanied by facial expressions? (OK tears stereotypically accompany "tragic", but "sexy"? I dunno what facial expression correlates to "sexy"). Are the test subjects intentionally emoting for the camera or otherwise not presenting an objective sample set? Do the people doing the filming write down or otherwise indicate what they're thinking of while their being filmed? How reliable is that? What if their faces are not expressive? Are they all filmed the same way (lighting and framing do a lot of work with film...)? etc. etc.
this is something that actual humans have problems doing. People misread other people's emotional cues *all the time*. It is socialized, learned behavior, and it varies really widely among people, situations, social strata, culture. This is hardly a simple operation for an AI to complete.
xxp
― Οὖτις, Wednesday, 27 January 2016 22:49 (eight years ago) link
please keep in mind that i came up with the scenario in less than 60 seconds
― Karl Malone, Wednesday, 27 January 2016 22:50 (eight years ago) link
i guess an unstated premise i'm using that others might not agree with is that our brains and computers are already very similar.
yeah I don't agree with this at all. When I was referring to consciousness upthread you can just swap that out for "human brain" or whatever. We have a very very limited understanding of how the brain works. By contrast we have a very detailed understanding of how computers work. There is a massive gap between the two, not the least of which is how our brains manage to process, sift and recall such a vast amount of information while using so little energy.
xp
― Οὖτις, Wednesday, 27 January 2016 22:52 (eight years ago) link
AI as "a machine that thinks like a human" is a pretty dated definition, the idea that machine intelligence can augment human cognition in ways that are nearly instant or imperceptible is the goal of most projects, or creating software that can adapt to new situations using past recorded data
the article about the hacker who is trying to out-tesla tesla on the augmented driving front, building a self-driving system that reacts based on recorded human responses to traffic conditions seems to be on the right track, whether or not his work is viable
general emulation of things we consider "consciousness" is a route that's well-trodden in the chatbot "can I tell whether this is a human" way and isn't really that important outside of customer support or w/e
― μpright mammal (mh), Wednesday, 27 January 2016 22:55 (eight years ago) link
imo we're going to find out more about the human brain by creating systems that learn than we are going to create systems that learn by determining how the human brain works
― μpright mammal (mh), Wednesday, 27 January 2016 22:56 (eight years ago) link
the idea that machine intelligence can augment human cognition in ways that are nearly instant or imperceptible is the goal of most projects
sure, this is something we're already living with.
but when people talk about AI superintelligences taking over, I don't think this is what they're referring to - they're referring to something that not only does what a human brain can do, but does it exponentially better. And we're nowhere near the former, much less the latter.
― Οὖτις, Wednesday, 27 January 2016 22:58 (eight years ago) link
I think it's more a matter of creating systems that have a gestalt decision-making process or evolutionary algorithm that comes up with things that humans would not, or would possibly not even conceive of
making machines think like humans is silly, imo, we should determine the better parts of abstract reasoning and develop that
― μpright mammal (mh), Wednesday, 27 January 2016 23:00 (eight years ago) link
machines that not only _do not_ do what human brains do, but do things in a way so differently that it seems foreign to our ideas of cognition
― μpright mammal (mh), Wednesday, 27 January 2016 23:01 (eight years ago) link
that makes more sense to me than trying to build the nine millionth robot that can't walk through a door
― Οὖτις, Wednesday, 27 January 2016 23:04 (eight years ago) link
(just to bring it all back full circle)
yes
i always warn against conceptually anthropomorphizing AI in these kind of discussions, and then end up up in a wormhole of rebutting anthropomorphic arguments anyway. and inevitably i mention sexy memories and things fall apart
― Karl Malone, Wednesday, 27 January 2016 23:06 (eight years ago) link