*Prof. Jürgen Schmidhuber: True Artificial Intelligence Will Change Everything*
Jürgen Schmidhuber is a German computer scientist and artist known for his work on machine learning, Artificial Intelligence, artificial neural networks, and more.
#ArtificialIntelligence #NeuralNetworks #MachineLearning #DeepLearning
Posted by Artificial Intelligence Blog on Friday, July 28, 2017
My speech will be about the most important about the grand theme of the 1st century which is the rise of artificial intelligence which is going o transform every aspect of our civilization and before we will look at the content rillettes have a brief look at the previous century what was the most important thing in the previous century the journal nature in 1999 made a list of the most influential inventions after twenty century and number one of class once the invention from 1908 which made the 20th century stand out among our centuries ever in the history of mankind because it was the one that drove the population explosion from 1.6 billion people in the year nineteen hundred too soon 10 billion it’s a chemical thing and a high pressure and high temperature nitrogen is extracted from thin air to make still 500 million tons of artificial fertilizer for a year now without that stuff half of humankind would not even exist this planet could sustain at most four billion people without that one invention billions and billions and billions would never have lived without it and soon two out of three people on this planet will depend on this one single mention nothing else was remotely as influential as an however the way I explosion of the present century is going to be much more impactful and grander than that because that we are not talking about smaller numbers such as for or 10 but we are talking about trillions of trillions and this has a lot to do with the fact that computers are getting faster by a factor of 10 per euro per five years and this trend has held at least since nineteen forty one man cannot souza built the first working program controlled computer and nineteen forty one seventy five years gone every five years since then computers became roughly 10 times cheaper which means that now we have a factor of a million billions and this trend has been running for a long time but only recently we have approached the computational power of a small animal brain and in the near future for the first time for a thousand euros.
We will have small computers which can compute as much as a human brain and then if the trend doesn’t break and there’s no reason why it should break it will take another 25 years and we will have it would take another 50 years and we will have a small device for the same price which can compute as much as our 10 billion brains combined as all brains of humankind together and that will not be only one of these devices but many many many we want programs these computers like we do with the con computers we are going to train them we are going to educate them like we educate babies and kids and one of the techniques that recently has become rebranded under the name deep learning which is just a rebranding of an attitude of an old hat is this technique card artificial neuronal networks has anybody ever heard about artificial neural networks
Has anybody not ever heard of artificial neural networks and came we have a third group in this room probably already asleep
planning goes back at least 50 years because and and much of that and there’s a good place this is in the middle of Europe and it’s a good place to point out that most of this research many of the basic insights come from Europe started 50 years ago in the Ukraine where a mathematician called even ankle but the first really deep neuron the tracks and efficient and networks that could learn from experience and we have made a lot of progress since then this Renaissance that you’d see down there is not a typo it’s really an RN s office because that stands for recon your networks which are the deepest of them are the recurrent neural networks are the deepest and most powerful artificial networks and they inspired a little bit by the human brain.
The human brain your brain has in your cortex about 10 billion little processors which are called neurons and each of them is connected to on average 10,000 other neurons which means you have got a hundred thousand billion connections in your brain. In the beginning these most of these connections are apparently randomly pre-wired but through learning through experience through trial and error they change and each connection strength indicates how much does this neuron influence this neuron over there and here we see our simple we can Network an artificial recurrent network which also has input units like for example your retina or a camera where pixels are streaming into the system or acoustic sounds or whatever or pain signals or pleasure signals are streaming to those little artificial brain and their output nodes which produce actions action sequences control robot muscles or control your muscles in your case and in between thinking takes place and these hidden neurons and the art of my profession is to come up with a learning algorithm that changes these connections such that he initially dump that right.
Over time learns to solve problems that it couldn’t solve before just like your kids too and like you did and and one of the best methods that is now widely used is called long short term memory loss of short-term memory is a method that we have developed in our labs and munich and in Switzerland so European tax money since since the beginning of the nineties and they are much better than previous networks able to learn to deal with t problems where you have to memorize things for a long time and and and launch short-term memory back then was just a curiosity but today it’s widely used by the world’s most valuable companies such as Google and Apple and so on which makes us happy and some people ask me do you have a demo and I just have to ask do you have a smartphone because you have a little piece of us in your pocket.
Whenever you take out your smartphone and you don’t want to type into it but instead you press on the little microphone there in the google voice icon then you can speak to it and and and the google white speech recognition is based now on our long short term memory which was developed since the yeah that’s the early nineties with the first author back then set pork right on my first you never been feeding skiers Alex graves and a couple of other great students in my lab and it reached the current form or less in the year two thousand six that’s what is now on your smartphone so I’m basically basically its 10-year old technology and and this thing has learned from experience to recognize speech.
So how does that did that work lots of people spoke and then in these training examples there was a teacher who knew what they meant so in comes speech signals which is every 10 milliseconds another number vector coming from the microphone so you get a hundred inputs / per second and then these go in through these input units and they suck around along these weak connections in the network and later the network is supposed to generate a sequence of letters which correspond to the recognize speech and until 2015 this didn’t work well and we had another system for doing the speech recognition and it was really annoying but then they replace it by a lunch order memory and since then speech recognition is not only five percent better a ten percent which would have been great but almost fifty percent and now even a noisy restaurant you can talk to it and it can recognize what you’re saying now this lcms all Universal powerful you can also use it to translate from one language into the other for example you give it lots of here we on in Brussels so you give it lots of examples of texts from from the European Parliament which is written in English and then translation & French in the beginning the network is totally stupid and knows nothing about English it knows nothing about French then it sees lots of examples of in a sentence as being fed in and be force it to output the corresponding translations in French in between the network is randomly pretty wide it has no idea what to do however through allowing out with him all these connections change their strength such that after a long period of training the thing has discovered the rules of English and syntactic rules are English the structure friends and how to go from the meaning and English to the meaning in French and the best system currently for translating automatically from one language to the other is based on this LCM and we had to have greatly profit of class from the fact that every 10 years we are gaining a fact of one under it when you started that computers were a million times lower than today for the same price now there are a million times faster and that makes a big difference I don’t network itself hours and feed forward networks that you can use for computer vision and 2011 for the first time we’re able to achieve even super human performance when it came to a pattern recognition pattern recognition is something that kids are doing very well and it’s much harder than playing chance how do I know that because in 20 years ago and the year nineteen 97 the best chess player on this planet was not human anymore it was a machine however back then no computer was able to recognize patterns just glasses or traffic signs or microphones or faces as well as humans to him but then in 2011 in a traffic sign recognition competition for the first time we had a super human performance there and three times better than the nearest competitor so the shot was that we are slowly creeping in there and into that level of human performance in more and more domains and remember every 10 years we are getting effect of 100 at the moment it’s scaling linearly and soon we will have rather big networks which can not only achieve one super human performance on a particular task task but on many many different to us we can apply the same techniques to medical imaging the brain image let me skip that for example when you see here is a slice through a female breasts and you see all these little cells there and this microscopic image and some of them are good cells and some of them are dangerous they are in the pre cancer stage it ourselves as their card normally you need a trained doctor a histology just who looks at these images and then says for each of these little details that that’s a good one that’s a bad one that’s a bad one i’m not a doctor but we can train our artificial neural networks on lots of theta to achieve a performance of recognition performance which is comparable to the one of human doctors that’s how it was and 12 for the first time we were able to win pattern recognition competitions and that really really important domain so this type of AI a little a small type of AI is already good enough to replace certain things that doctors are good and doesn’t mean that doctors are going to be is abolished not at all just means that the same guy the same doctor will be able to treat ten times as many patrons in the same time with high quality and we’ll have more time for the thing which is often neglected today which is into it which is doctor-patient discussions and stuff like that this is super important because the wall GP is about 70 trillion is it already 80 trillion armature and ten percent of that is for healthcare which is about seven trillion and at least ten percent of that per year is just for a medical diagnosis like that which is seven hundred billion per year but apart from the numbers and finance an economy that’s much more important is that lots of people who at the moment don’t have any significant access to health care at all through artificial doctors like that are going to have decent health care we were able to win additional competitions along these lines here is a thing that google is doing with our LCM combined with the technique on commercial networks there you in the lower left corner you see an image where where you see a text below the image the text says a herd of elephants walking across a dry grass field and you look at the image yes room and the interesting thing is this was automatically generated so there was a and an LLC and network combined with another few for network which has learned from many many training examples of images and captions to recognize what is in the image and then give a short summary in an English paragraph what you can see there no teacher just from lots of training examples like that sometimes it goes bad for example the second the second image in the top roll says two dogs play in the grass and you look at it and it’s actually three dogs but it’s not complete completely on before I came here I thought this is going to be just a little tech talk and there won’t be much of an audience but you are actually a large audience by my standards the other day i gave a talk and there was just a single person in the audience it was a young lady and I said young lady it’s very embarrassing but apparently today i’m going to give this talk just to you and she said okay but please hurry I’m the next speaker recently recently it google deepmind made a program that became the best goal player in the world and it wasn’t pre-programmed it learned that from lots of games playing against itself it’s not a totally new thing in 1994 already a backgammon program well I’m to become the best back and play and the wild by playing against itself using very similar principles however go is more complex and backgammon and a couple of additional tracks were employed there and it received a lot of attention these are neural networks which one over time to become better and go players and I’m proud of that because deep mind is a company which doesn’t even exist five years ago and then in 2014 was bought by google and they were heavily influenced by my students actually the first two guys a deep mind who are doing one of mine is doing which is artificial intelligence and and machine learning they were both students in my lab where they met and they were the first at the mind who had really phd’s in that field then later they hired a couple of additional guys from my lab so this shows that there is now a lot of commercial interest in in this stuff and Google and many communication companies are massively using artificial intelligence or at least these artificial neural networks all the time to place better and whenever you are searching for something so these are marketing companies communication companies and they have taken i think half the advertising business after the world google and facebook ugh which is using similar techniques by just being better able to tailor ads by looking at what kind of data can i get from these users search that i can increase the probability that they will click at these ads so the techniques that I’ve mentioned so far can be used for things like that and are part of the money making machine behind these search engines what we will see in the near future I extensions of stuff that we did maybe you already also 10 years ago where you also can control robots them to that sex we had lstm networks that learn to control the surgery robots like that to tie knots into in environment fine settings and artificial pics in this example no real pics harmed some people think that creativity and curiosity are something that will always remain a domain of humans but this is natural and we have a formal theory we are fun and curiosity and and and creativity which allows us to already build simple artificial scientists and artists and here we have a little robot a baby robot which in the beginning you nothing but then over time going through experiments to to interact with the world usually when you have in systems like that as two systems one is the neural network which is interacting with the world and then there’s another one it’s friend you you might say which lines to predict what happens if I do is add on that what happens if I do is add on that and then this production machine and in the beginning knows nothing but over time like a baby learns how work that gravity works how do the apples fall to the ground if i push them from the table and so on and so on so it lasts to become a better and better predictor of what’s going to happen if i do that and that and then we can measure the insights office second module office world model if you will we can measure the depth of these in science as it learns something that I didn’t know and that’s a number and we get that to the first guy who is creating the experiments that lead to the data that has the property that the one model can become better and now the first guy is motivated to maximize its all these rewards all these curiosity rewards these intrinsic joy signals which motivated to come up with additional experiments that tell it even more about how the world works so I’d official scientists in a certain sense that we already have had running for her a couple of times for a couple of years how much more do I have not so many additional minutes left however let me let me again point out that we are currently greatly profiting from the fact that every five years we are getting a factor of 10 now we have 75 years after tues er which means we have now even we have now lstm networks long short term memory networks after type that we development switzerland and yannick with about a billion connections your brains have about 100,000 billion connections one hand thousand means 25 years because it’s five to the 10 which means we have to meet way for 25 more years and for the same prize we will for the first time have LSD and networks that have the size of a human brain and they would be much faster than human brains because they have electronic connections not the slow connections that we have things are going to change the price on many people don’t realize how quickly this is now moving forward what was the be the next thing I think in the not-so-distant future we will have something we don’t have that yet which is like a little animal like intelligence like a little monkey little monkeys at the moment cans do many many things that our best robot cannot do it on can learn lots of things that machines cannot be at learned however we think we understand how to get there and within not so many years we will have little and artificial intelligences on the level of arm of a simple of a small animal like a crow or monkey capuchin monkey and once we have that at the step towards human level intelligence won’t be that huge because look at evolution it took billions of years to come up with a little monkey but then only a few millions or tens of millions of years to add human level intelligence on top of an because technological evolution is a million times faster than biological evolution because the dead ends are weeded out much faster so it to me it would be super surprising if within a decade in within a few decades we won’t have a human level intelligence of the artificial kind I don’t want to deny that we have a company not only the academic lab but also a company which is called nations which is trying to make that a reality in a sentence is pronounced like birth Mason’s but it’s spelled in a different way nn4 neural networks hey I for artificial intelligence what will be the far future of course once a eyes are going to be smarter than humans and it there is no doubt in my mind that this will come within the century what will they do they will not stick to this thin film of biosphere around the third planet because almost all resources in the solar system out there in space less than 1 billions of the solar energy is hitting our planet and the rest at the moment is wasted it’s not going to stay like that and they will move out there and they will build billions and billions of self-replicating robot factories and the asteroid belt and spread from there in a way that is completely impossible physically impossible for humans space is hostile to you mine humans but it’s really friendly to appropriately designed robots and they are going to spread slowly out through the Milky Way and within a couple of millions of years completely within the limits of physics and light speed and so on they were established a network of senders and receivers all over the galaxy and of course from then on they were a eyes will travel the way is
always have trouble namely by litespeed by radio from send us to receive us in a way completely infeasible for humans we are currently witnessing the beginning of something that is HUGE this is not just another industrial revolution this is more than all of civilization this is a step and you step on the path of the universe towards higher and higher complexity and the last time we had a step of that significant I think was about 3.5 billion years ago with the invention of life so this goes beyond human kind this transcends human kind and and it’s a privilege to be part of that and fitness the beginnings and with that final font i would like to point out that we shouldn’t think of us versus them as humans versus those super uber robots of the future but view all of us including human kind of civilization and these future beings as part of one grand scheme that allows the universe to go from small complete from from simple States towards more complex States and its it’s great to be a part of an thank you very much for your attention thank you very much you’re doing some exciting predictions for true artificial intelligence coming at us within our own lifetimes basically what you’re telling us we’re going to say I’m true just run it by us again what classifies true in that definition as opposed to what we can already see how you doing around us today what can be always easy when you’re talking to your smartphone and it’s mostly pure pattern recognition your smart phone doesn’t have arms it doesn’t shape the world it can influence you by giving you advice I said for example it says now you are on this fine City and you can and i know there is a secondhand shop not far from here which is after type you like and they have a special thing on offer and you go there because that’s a good deal for your but they don’t have a robot arms and the moment what we see is that robotics and mechanics are lagging behind what we can do in pattern recognition it’s not going to stay like that so i think within the next year’s and and few decades more number of decades we will see very sophisticated robots that will be able to solve all kinds of problems that humans at the moment have to solve by themselves including strawberry plucking which is much harder than most people think simply because there is no really good strawberry plucking robot it’s not going to stay like that and then of pants hey I and general is really not just pattern recognition but interaction with the wilds are you you act you perceive you act you perceive you get a stream of input state
28:01on that you’re shaking yourself on your way to solving goals because all of AI is about problem solving and this is currently becoming a reality although most commercial stuff is just pattern recognition just better speech recognition better gesture recognition better prediction of the stock market another thing that our company is pretty good at and and better prediction of what you want to do next given the data that you have on your smartphone I’m just going to pick up on that point of strawberry plucking or picking I’m simply because that leads to the question obviously there are lots of concerns about the implications of artificial intelligence for employment for instance we’ve already seen robot journalism happening just this week or last week it was perhaps how worried should we be that those of us who are involved in in a worthy kinds of kinds of professions are are simply going to be replaced i mean many professions obviously strawberry plucking being amongst them yeah so in the eighties there already said always armed it is very easy to predict which jobs are going to go like taxi driver and stuff like that it’s very hard to predict all the new jobs which are being created all the time and that’s this playing mounting homo Luden’s homo Luden’s the playing man is inventing new professions all the time and most of these professions are really luxury and professions for example although the best chess players in the world are not humans anymore you have so humans making money by its playing chess against each other well he was involved is much slower than the fastest machines but he’s still making hundreds of millions just by running against other humans and all these new types of interactions with other people that you see on social networks bloggers youtubers and so on who could have predicted that 20 years ago so if you look at the unemployment ates today they are pretty much the same that we had back then so adaptation is obviously the secret there you know and like Matt peacocks and earlier in the earlier session of your kids are wondering what they should do they should become data analysts are amongst other things questions from the floor for your continued to bomb we have over here get you a microphone thank you i was interested by the comparison you made about parenting and teaching machines to as you teach a child and I recall when I he first became apparent never once said well you know there’s no manual for you need to figure it out and I realized very quickly that it’s not that easy and some people do it better than others so using that comparison how do you sort of respond to the future that you’ve sent out yeah so you are worried that some pounds are going to teach them the wrong things for example in military applications where some of those guys will be taught to do military jobs where for example self-driving cars are going to be used as a self-driving landmines seekers which of course every general will want because he wants to protect the soldiers and so on and so you’re worried about educating them to do things that are detrimental at least to the lives of certain people however on the other hand it is clear that almost on and off the commercial research in this field is driven towards making a is artificial neural networks that alarm to help humans to make us happier such as a smarter friend in your pocket your smartphone which is even better at understanding you weren’t talking back to you and and giving you advice and so on or in health care where people will just live longer because of this partially automated health care that we are going to see so am i correct that you are worried about the relation between on the one hand these military applications as opposed to these much larger and much more valuable commercial and health care or Internet applications well not quite actually what I’m worried about is the fact that as a parent again I do the best I can raising my children but there are things that I might do at the age of five that might have an implication for 10 1015 for the choices they make so sure that a lot of people who are who are the lots of effort to do the right things but actually you know take the comparison of a smartphone certain people now you we talked about having digital detox weekends or whatever it might be because we realize that there are consequences that only become apparent later on and its technology is a fantastic thing but the constant these consequences and that’s really what i’m getting at because you’re taking sort of science and you’re going into room where it like you say it’s an unpredictable thing like parenting yes but I’m not the first on this path it reminds me a little bit of the discussion that we have six hundred thousand years ago when fire was invented and and back then and ethics committee was established which which which weigh the pros and cons and some people said yeah it’s going to keep us warm at night but the other set but you can also use it to blind people and then at some point the Commission came to a conclusion and they said we are not going to stop that development because we can’t even stop it and let’s move forward and I think weezy the same thing happening now there with the thank you very much for that for those questions certainly inviting analog hello thank you for your talk we hear a lot about trust in society and we had about society splitting in two and basically apart from the discussion on artificial intelligence intelligence per se is not equally distributed in society do you think that the rich well educated people are replacing the other half with robots and what the robots have voted for breaks it so in my profession it’s not unusual that people like the idea of unconditional base salary which recently we had as a discussion in Switzerland so I’m not twist but i’m living down and so on I see what’s going on so I think almost was it a third of someone a third of the population would support support that idea and I think in a couple of tens of years we will have many more are so many people in in this profession of building machines that become smarter over time I think it’s a good idea to have robots pay taxes and to have robot owners pay taxes and of class society will have to come up with systems like that with a social response to the technological advances and it will happen as it always happens otherwise we will get a revolution thanks for that question we have another question over here in the second row yes professor you made it clear calculation and beginning about how much time it will take until this or that happened and now my question is if you extend that calculation how long will it take until the robots will definitely take over the earth yeah well they really take over the earth that is a very debatable seen inspired by arnold schwarzenegger movies and and it has not so much to do with reality we can see you are being taken over or you are being enslaved only by others who are like yourself who have similar goals and share the same goals so that’s why humans usually quarrelled with other humans but not so much with some with kangaroos and and it is the case that almost all people are interested in other people who are similar to themselves because either they can collaborate with them or compete with them or sometimes both and knit one nation for one company competes against another company each of them being the collection of humans and and and the fundamental condition for that is that you shared goals because you are similar now the super smarty is the future will not be so interested in humans just like humans are not so interested in the ants the super smart eyes of the future will mostly be interested in other super smart the eyes of the future simply because those will be much more interesting and share similar goals in an environment which may be quite disconnected from what we have here in this little white sphere we ask much smarter than the arm and Santi but only when they invite our houses we take measures against them but most of the arms in the world they are happily living in the forest and be i’m glad they are doing that and the weights of all time it’s still comparable to the weight of all humans simply because we don’t have too many old conflicts with each other and that’s going to be the same thing with the super smart Romans we have time I’m going to I’m going to make allowances for one more question here in the front because they were going to have to a very short follow-up you were mentioning a movie i mentioned another 1 2001 a Space Odyssey the Durham always not again the humor he just has a logic has a mission to follow and man is on his way so he’s had different goals these things differently than a human but it soon becomes a danger and if potentially and interactive intelligence says for instance there are about to begin to many people so logical step will be to kill for instant to be other people to make sure that the seven other ppl billion can survive that work for instance a different kind of logic than a human logic so i don’t know if it’s that clear that the machine will not be a problem so what you are saying sounds more like human logic to me rather than machine lodging and we do have people are in the history of mankind there have been people who had ideas like that let’s kill all the others that such as just we remain and then beyond the achieves but again it’s always about similar guys against other similar guys because they share the same goals if you don’t share the same goals then there’s no interest in fighting for example what we do the you fights off the past come from because this country has something that this country also ones are like oil or land or whatever and and and then there are these fights but generally speaking as soon as you have disconnected life and goals of a different very different type which you can expect from future suppose my eyes then you don’t have to worry too much about the things that you see in scientific in science fiction novels for example and in some some of you may have seen the film matrix matrix huh has a silly plot it has great computer graphics and the coach are great the blackboard a great but but the plotters the silliest plot ever so they have the way eyes of the future they live RC energy after human brains so each of each brain produces maybe 30 watts of energy and the coal power plant that you need to keep thinking man alive produces much more energy than that so all of these plants are still legal conflicts invented by film producers who just wanted to have a clash between robots and and and machine and horses and joins and it’s are very unrealistic as seems clear that this is not the future thank you very much you’re going to meet her but I’m going to have to stop it there the bad news is we’re stopping it now but the good news is that your guns – <operand> twenty </operand> has decided to stay with us and I believe also attend the evening festivities so he also be around here in the ensuing moments after this session for some one-on-one questioning please give my hand thank you very much yeah.