good afternoon instead of discussing how to build super intelligence and instead of giving you the secret formula of how to do it I will simply because it's too late in the day to day rather discuss whether there can be whether they can be super intelligent and so we are cut all used to these spectacular headlines and I won't comment them in detail here because you've seen them all but again and again we see as well as humans can or we see something like better than humans can so in the case of MT we know the stuff very well because I'm from language processing and the people at Google are former SAS people of ours some of them at least so int voice computing is another case where people say that now some systems for some applications are better than humans and because it I gets better than humans in some cases for some tasks people are immediately afraid will they take all jobs away but I won't discuss this I will rather discuss today what
super intelligence could really mean if it exists and where can we expect it and where is it pretty much impossible in the foreseeable food future so this was a runner wings theory or that was his prediction in 1993 within 30 years we will have the technological means to create superhuman intelligence shortly after the human era will be ended if you look at the time very close or already do and Stephen Hawking warns that robots could evolve faster than human race and could be hard to stop I guess all these things have been discussed in the morning I will take a slightly different angle ayuh and then of course we know that in heah every eye specialist knows who's really behind our researcher okay so and here also it's Tesla's mask also it could it has to be sighted so I have started to penetrate all sectors of manufacturing and service industry actually and is now in many places in much less spectacular forms than alphago and deep blue but and the magic is of course
that in AI enables machines to do things that up to now only humans could do so we think immediately if only humans can do it that must be AI yeah there are some things that humans can do like reproduce other humans that may not be called AI but most of the things that only humans can do our AI so but AI extends and enhances humans perception and cognition so far as clear and I think you saw this in many of the talks but the question is not does it only enhance human cognition or can it really be human cognition and AI helps humans and machines to interact so almost every respectable large company now has these speakers in your living room and and tries to get in into next into the you saw the latest announcement into the bedroom the bathroom is coming and so all these things so you can talk so why does the whole thing happen now and I hope that has become clear from the morning talks it is actually because
five different things come together there are five developments and each of them is spectacular enough but if they come together something new comes up so it is big data both the availability of the big data but also the availability of the technology for handling it and then secondly it's the breakthroughs in machine learning you heard several talks in the morning on deep learning but deep learning is not the only game in town that's maybe the strongest one but there's many many powerful new methods of learning and then but then there is something I want to that shouldn't be forgotten in at the same time when there are these spectacular success stories of machine learning deep learning there's also very big knowledge accumulation yeah think of the big knowledge graph of wiki data the Google knowledge graph and and other big collections of formal knowledge I will come back to that then it and and some of the some of the breaks loose are due to the success in this area then there is of course
Internet technology IOT and and web services and then of course without the hardware without the progress and hardware all of this would not be possible today that all we know and so it's mainly to go back I will not because there were several talks already I will not now go into the deep learning itself but much of the stuff that I will say actually has to do a lot with deep learning and so it's actually it addresses deep learning directly so now our question is is the superhuman performance and of course the answer is easy in certain tasks it has been proven with deep blue and chess for quite a while ago and IBM Watson clearly IBM Watson was better than the to human quiz champions everybody could see that on television and it was an extremely fair competition because one of the three candidates were connected to the Internet there so that was completely fair there was one human in the middle was the machine and there was another human of course the machine
didn't fit into the seat so actually the next floor below was also filled by the machine and the machine was not connected to the Internet but it had downloaded half of the internet before so was a complete was a complete fair was a complete share competition so and then you have also things like alphago chess world master chess world champion was beaten and Google's driverless cars you know that the American authorities already determined two years ago that now the level of human driving has been reached and they may not be driving as Michael Schumacher or settle but they are driving much better than the average guy so there is clearly superhuman performance so is there then also super intelligence and is what these machines do really intelligent or is it just big is it just like in the case of IBM Watson tons of machines that can do what humans do without intelligence the same thing has been said about deep blue of course if you can build a computer that can calculate many many
draws many moves and advance and there can calculate all the consequences of these moves then maybe you don't need to be intelligent yeah you just need need to calculate so this is an interesting question node is it intelligence really or is it just pure computing power so this is a driverless car and this is the let's come back to the doctor a little later this is the doctor but I'll come back to the doctor a little later so they are actually the three big pros of artificial intelligence and in your companies you have to decide what to do and the three pros are propaganda processes and products so on the left is propaganda this is things like yeah this is things like alphago IBM Watson and in the beginning its lots of propaganda but now IBM Watson has become a core product and Google has declared a is earth so it's moving from propaganda into products and processes for most companies it won't be the case that in the near future they may have a
AI product maybe a company that produces chairs or handkerchiefs or glasses they may not put AI in there in the in the near future but even there I can help quite a bit in cost savings in automation in customer relationship management and business analytics and logistics production maintenance market research and all these areas and then of course its products in AI on devices and software and and cloud services so but still we have to answer what does it have to do with this thing here and very often people like now to show a brain and draw con draw parallels between the little cells in the brain and the neurons and neural nets and think this explains already that we are getting close to the human brain it's of course far from being true the type of neurons that we have in the neural nets do not in this in the slightest their resemble already the high differentiation that we have in the human brain so let's step back a little bit from the wetware as
we call it let's step a little deck and say what are the components of human cognition and then go and see whether the AI systems already have them or it whether they don't so theirs first when we are born we are already born with something namely with innate instincts and capabilities those we owe to millions of years of evolution yet before the humans were humans so we even a child is afraid of snakes insects there we discussed body excrements we have attraction to the other sex you have emotions and we have very sophisticated means for dealing with human faces the baby will know that the mother laughs before it was trained like a neural nets on millions of laughters so the baby knows this already was born with the innate capability so the second thing is acquired explicit knowledge those effects theories methods and then symbolic sequences and patterns what's the difference so the facts are the things that we have stored in the brain we may not know exactly how to phrase it each
time we repeat the fact we may use different words so it's not that we have stored the symbols but we have stored something like symbols namely concepts your pointers to certain things in the world and then we have can also remember a little bit of symbolic sequences some songs some formula some sequences some phone numbers but we are very bad at this error so this is where you're very bad at and you can already think about what the story will be or the computer is very good at these things so but the computer is also very good at storing facts but we will come back and see what the differences and then the third one is that we have other acquired implicit knowledge like our first language so try not to listen to the next sentence I'm telling you please switch off your understanding you cannot because it's not voluntary you cannot switch it off and many people speak a language very well who don't know what an object or a subject or verb phrases
so that means it's also implicitly image at least the first languages and then lots of procedural behavior let's describe to our children how we do swimming breaststroke and after we've described it three times we push them into the water and go away because they can swim them now so this doesn't work of course we know in this case something like a neural net would be good there something training training training so this is the this is the this is the part of human brain now we can put this the different we can split up little more sophisticated ways we have the sensory processing first and we interpret then we do inference then we plan then we communicate move physically and so and then there's also some reactive system and this is actually the part that resembles most closely they learnt behavior by neural nets but I'd come back to that yeah up up there the stuff this is more the symbolic AI and in the human brain we have both and the power of human
cognition is that these are connected to this perception as a section this is unplanned this is the inferencing part this is language and so on here so and this is the effective system the reactive system is most closely connected to the limbic system to our emotions here doesn't have to be that way but it is that we're here so it is the way so now let's look at different types of systems in AI and machine learning this is an old base an old expert system for credit check as American Express and others have used it for a while it is built on knowledge has rules lots of rules in their logical systems and this is a classical reactive robot control programmed line-by-line now machine learning comes into play and again we get now systems on both sides on the side on the on the left side we get things like IBM Watson storing many books tables schedules storing storing lots of the and chatbots are actually maybe a mix because the knowledge they look up
maybe in tables and texts and rules but when they use speech understanding the speech understanding completely works with statistical and neural methods today and for instance our Big Data text analytics is here but machine translation is on this side that strange to language things on this side but I mentioned earlier already also humans use quite a bit the non symbolic way of processing language or the non explicit knowledge and so right now the best method so but Google machine translation does not understand a single sentence not even a word but it translates because it has been trained on millions of sentence pairs translated by humans and it is just like a parrot mimicking this human cell so it's a very very good mechanical parrot so why doesn't that get together because in the human brain that will come together and actually many people who work in this field will know that it will come together why because some are in our labs all groups are working on it they are not showing it yet at
the big conferences because this is the next wave here so right now people get a paper if they can improve the F score in some convolution you will net or GRE you net by a few percentage points but this is all not the important stuff the important stuff usually people don't talk about the important stuff is who who will be able to combine these things in a reasonable way this will be the next revolution not the improvements in F score so so right now there are really different systems as I showed you already but this machine learned behavior systems have no reusable knowledge at all that's why we can build on the one hand systems that beat the Google that beat the go world champion and can drive but the driving system cannot translate the translator cannot play chess the chess player cannot do anything cannot speak or understand speech each of these things are idiots they can do one certain type of behavior now it is possible to connect two of them people have
done that and tried that you get a huge neural net and you connect to but this is not what humans do if I'm driving and somebody tells me that the road is wet or that somebody spilled oil here or that there is a new traffic law in effect then I can adapt my driving because my learnt behavior is connected to the knowledge I have but this is not true for most of the statistical train systems and for nearly all of the neural net so there are some in the lab that has some influence where you can steer and go in and change but in it doesn't scale yet so this is they are completely free of knowledge and they cannot explain why they do something there's no grounding our knowledge is grounded we only know something if we know the lower part sell its components and can ground it in the end and everyday knowledge so think of a doctor I told you I will come back to the doctor because we do lots of
medical work in AI I like doctor examples so you you know the situation that the doctor looks and tries to find out what's there and the doctor cannot remember a single case and the doctor cannot remember scientific literature describing something but all the doctors have very good intuition especially general practitioners who have seen everything under the Sun many times so they have an intuition and they cannot explain but they say must be this I don't know why and I knew some doctors of that sort and such people I also met in companies there so usually you get some old hands who are very experienced who have very very good intuition of things there now our neural Nets only have intuition they have no knowledge they only have learnt intuition so imagine somebody who was a doctor who was deprived of going to medical school who never was allowed to read a medical book but only learn from symptoms and diagnosis from pairs of symptoms and diagnosis and he got millions of these in the end
he will be doing what the intuition doctor has it's no knowledge not in the sense that we talk about knowledge it can also not explain why it doesn't there so now there's a whole branch in neural networks of course that tries to get this knowledge out because implicitly it must be in there if we could now learn from these neural nets if we could learn then the neural Nets would be able to maybe crank out the rules better rules for the other doctors and this is a whole branch and machine learning actually a very fruitful idea right now it doesn't work so the difference is for solving intellectually challenging problems people need understandings and creativity if we have understand understood if we have understood a difficult problem we have almost solved it this is the first step to solving usually we cannot solve it because we don't understand it if we have understood it fully usually the solution is there but these systems that don't understand a little bit because right now their knowledge is
not connected with the learned behavior have a break at have a barrier at this point so this is the problem now so if we look at these components of human intelligence then we know we have statistical learning of behavior in some systems those are the neural net systems we get large reusable knowledge and other systems like IBM Watson or systems that are even more built on on symbolic knowledge on huge amounts of formal knowledge and then we get a few inference capabilities not yet far as good as humans but we get some inference and we get some knowledge acquisition and intentions and planning a little bit most conscious low self consciousness more analogy reasoning or sentiment emotions new creativity and actually a very good analogy reasoning is used for creativity because if you look at any reasonable theory of creativity analogy really doing with analogy is the basic building block is a very basic building block of creativity so this is not there yet but on the other hand in machine learning I don't know
if somebody talked this morning about transfer learning because I couldn't for I had so fought such a full day I wish I could have listened to the talks I could not transfer learning is a discipline within machine learning also used heavily a neural net learning that actually builds the mechanisms that can be used for analogy formation so it is not that was hopeless learn analogy it's not hopeless but they are still something missing and this is the drive that humans have for creativity because if you think of my earlier picture of the different or list of the different components of human cognition we are not born we are not living by goals that are given to us intellectually only but we have deep desires if desires come from our instinct they drive they determine our goals there's the limbic rewarding system we get chemical rewards yes without these chemical rewards we would be yeah we would be like some people who are very depressed there who need these rewards given in form of medicines because
otherwise we couldn't function at all anymore we would be without any driver we would kill ourselves so we are really dependent on all of that and there's no equivalent for that and the machine and there won't be in the very near future not even in the distant future so but assistants need to understand their master so it actually bringing these two worlds together is absolutely necessary this is and I think many people visionaries in the field they know and the reasons that you don't get more talks about it is because people cannot tell you yet they want to be the one who tells you how to do it it's not that they are not aware of that need so if you look now a little bit and then I will conclude already uh if you look a little bit in in the in the growth of knowledge that we have now have we because you all talked about neural nets on learned behavior the whole morning look at so you know this is the Wikipedia
compared to the Encyclopedia Britannica and then some of that knowledge gets formalized this is an picture of the Google knowledge graph also some of the knowledge of the Wikipedia gets transformed and here you get your documents and here you get your knowledge here you'll get your knowledge from the knowledge graph and these knowledge graphs as we are setting here are growing day by day buying not just knowledge communities but we and others develop software that also automatically now from texts and there are some just some number for instance the dbpedia we are working in a project with them where their stuff they are now more than a billion facts the Google knowledge mold this is less secure this is more secure knowledge this is less secure knowledge the knowledge graph at Google has the secure knowledge that knowledge fold has lots of and secure knowledge that gets even higher but it's not secured so you get lots of this stuff so now you can imagine how this stuff is coming in because you not only
have the machine learning and the knowledge and the behavior that I learned but you get something else you get systems that can use data at runtime now they can use data at runtime the data from the outside and the first step with neural nets will not be although some people are thinking that you put all databases into the big neuron that you'd put it all in and then it says use the biggest database and so but there's a more effective way of right now because the other thing is architectural II and from the efficiency not solve there is namely using these huge amounts of data at runtime where many systems can do that and if the system has the access to this huge amount of knowledge and as people we only have this tiny head yeah like these poor guys who were beaten by Watson here we only have this tiny little ball here here with very little in here and there is a machine here that has all that stuff in the world in
there has more than a billion Texas all the books in there all the tables or so you can imagine at this point we're going to lose yeah and the system will still not have the drive yeah so this is actually where the superhuman capacity will come up and it is already coming up and this is where things will happen very very fast in the next years still we won't have my prediction is we won't have the Hollywood bad evil AI a-- that comes and are looking where does the human have the switch off button yeah so how to switch off the humans and the evil AI that ones because I told you already earlier all these things self consciousness on Drive all these things are not there the systems are not reflective enough they don't have models of the world and of themselves yet but in their decision-making and in their support for human decision-making they will be much more powerful than us in a hopefully very benign way why because they have access to
all the knowledge in the world yeah we can already build robots that go around and can tap into the knowledge vault or into dbpedia know everything about the thing and but there's one more thing that maybe was mentioned in the morning but I'll repeat it in case the speakers in the morning ran out of time before mentioning it there could be and and this is the following in our case one part of our knowledge has undergone evolution our knowledge has undergone a couple of thousands of years of evolution because the symbolic knowledge gets passed on from generation to generation and can be added to but if the pianist dies or a very world-famous driver race driver then it's behavioral knowledge is gone yeah it's gone with a at the point of death and so that means there is when it comes to the dancer a driver there's always a limit that has to do with the life cycle pursuit over the life cycle at some point you cannot learn more you can only learn from
and and also you have limitations of your own neural system but the thing is with these neural nets that were discussed in the morning you don't have this anymore because you can copy them now you can copy them like we copy our symbolic knowledge shell books in tables and so on so that means it can grow over generations or it can be implanted in many devices and learn at many places in the world so the learning actually is also much much more powerful than in any other system okay so I think I'll finish with that because I think that explains a little bit what I have to say this is modified slide by both convice that was the CEO of the GSK I he sees this as the main phases of AI because we tend to think a is something very new and cool but actually it is old and much cooler so it's soft and so this is the phases it started with the heuristic systems put everything into the program then the knowledge-based
systems then the machine learning and I just adapted the numbers because I disagree with what converts go under times but I agree on the phases and and now the next system is and I don't mean now IBM here with cognitive systems IBM seems to have bought the term cognitive someplace here but I don't mean that I really mean what I showed you earlier namely the integration of the two worlds okay and let me finish them I think if we managed to get to this point then we can activate immense economic reserves and because all of our processes and society are very wasteful still and it will be a sweeping augmentation not a deletion and augmentation of human cognition and probably also discussed it will be liberation from alienated labor and to transfer labor to non alienated labor is a difficult task but we cannot discuss it today here it will perfect service knowledge society and accelerate the evolution of knowledge we already have very fast acceleration of knowledge but this will double and triple again
and it will lead to better longer and safer life thank you very much you