 So now to this talk. This is Luke Gotchling. He's standing there and he will talk about Yes, you know, Turing tests. That's probably something everybody should know in here, but maybe in the meanwhile some tools are good enough to simulate something where what for example a three-year one cannot beat so There are maybe some more interesting questions what is intelligence and yes the same tool may not be able to translate a book or something And yes, so we want to get a general thing about this and this is the talk The quest for artificial general intelligence Thank you I think we should all get some sort of award for surviving the weather of the past few days And thank you for showing up out for the first talk of today So my name is Luke I spend most of my time getting computers to try to understand documents and this is the quest for artificial general Intelligence beyond the terrain test. So a good starting point, of course is how do we define artificial intelligence? There is no one single definition One way we can define it is by saying intelligent behavior in artifacts This works pretty well. What about a longer definition? Well, rich and night say the study of how to make computers do things at which at the moment people are better But there's a little bit of controversy about this because not everyone agrees that we should be comparing them to To people and so Benjamin Bratton says that we would wish to define the very existence of AI in Relation to its ability to mimic how humans think that humans think will be looked back upon as a weird sort of species What about Artificial general intelligence or a GI well one definition for this by Gertzel and Panachin is Systems that possess a reasonable degree of self-understanding Autonomous self-control and have the ability to solve a variety of complex problems in a variety of contexts and to learn to solve new problems that they didn't know about at the time of their creation and this whole learning process I think is really fundamental for AI and AGI So let's try to break this down if something is going to be generally intelligent. What must it have? Well, it has to have the ability to solve problems general problems in a non-domain restricted way in the same sense that a human can and Kind of the flip side of this it also has the have to have the ability to solve problems in particular domains and Particular contexts with particular efficiency. So for example, it has to be able to be really good at chess and maybe really good at summarizing a book and then combining the previous two so it has to be both generalized and Have specialized intelligence capabilities and use these things in a unified way Going back to the previous definition the ability to learn from the environment from other intelligence systems and teachers It's this learning thing. I think that's really key It's that it gets better over time and really the only way to do that from an intelligence basis is to learn And of course to become better at solving novel types of problems as a gains experience with them so The year is 1950 Alan Turing is at the University of Manchester and he has a question his question is can machines think? so, I mean, it's a very philosophical question obviously, but Before he goes on to define what eventually becomes known as the Turing test He actually has an interesting analogy and that is a new form of the problem Which is known as the imitation game and this is a very simple game played without any computers or any machines It's three people a man a a woman be and an interrogator See who may be of either sex and the interrogator is isolated from the other two Now the interrogator is going to ask questions of a and b and going to try to determine which of them is the man And which is the woman? And obviously the key to make this game interesting is that the man and the woman have to agree in advance about Who who they're both going to try to be so one of them is going to be lying So he can't just say like hey, are you a man? Because they both should answer the same way in which case this that's what makes this game interesting and of course the objective is for the interrogator to determine which of the two people is the man and which is the woman and Then so starting with this original test He defines what becomes known as the Turing tasks And so he says we now ask the question What will happen when a machine takes the part of a in this game? Will the interrogator decide wrongly as often when the game is played like this as He does when the game is played between a man and a woman these questions replace our original can machines think and so It should be noted that this is not a like a formal Problem definition from a mathematical point of view. This is more of a guideline, but it's still a very very Widely accepted guideline for for machine intelligence or general intelligence now the field of AI has suffered a lot of hype and Setbacks over many decades One of the original failed predictions for AI is this one I believe that in about the year 2000 it will be possible to program computers with a storage capacity of 125 megabytes To make them play so well that an average interrogator will not have more than a 70% chance of making the right identification After five minutes of questioning Who said this and what year did they say this? Well, it was actually Alan Turing in 1950 and this is one of the first failed predictions about artificial intelligence So another kind of Challenge for general intelligence is people will say something like well, okay So the computer can solve certain types of problems, but well, you know, it can't drive a car And then the 80s happened And then at Carnegie Mellon and Bundeswehr University Munich We had the first Autonomous vehicles that were able to navigate a course You'll notice that both of these vehicles are trucks And of course the reason for that is because all of the computing hardware had to be have to fit in the car And the computing hardware in those days was so large that the smallest vehicle they could get was a truck That would fit everything and of course we've we've come a little bit since then We now have cars that are able to navigate public roads alongside human drivers and do so more safely than the human drivers and then cars that are even able to go on 130 mile Journeys completely autonomously in the desert now. We can't buy these cars yet, but I think it's only a matter of time before we'll be These cars are gonna be on the roads more so than they are now and we're gonna be able to buy one Because they are gonna be safer than human drivers Now they used to say that well, okay, so AI can do certain things, but it can't beat a chess world champion Well, we know what happened in 1996 and that is Gary Kasparov lost to IBM's deep blue Being the first time that a computer Defeated a chess world champion The hardware and deep blue is kind of interesting because it was 30 120 megahertz ships and 480 custom VLSI ships You can't really you know and this was one of the top 500 super computers at the time But if you look at the limb pack benchmark, it's really funny because the capacity or the computational power was about 11 iPhones worth Now, what about Jeopardy? I think if we remember back a few years ago that this Jeopardy Was defeated by a machine IBM's Watson defeated the top Jeopardy players back in 2011. So that's that problem has been solved as well Now what about recognizing things inside of photographs So there has been some progress with this a zoo and others have for example shown that You can take a scene and the computer can say it's a woman throwing a frisbee in a park or a dog standing on a hardwood floor But it's not really quite there yet because in this case the top It identifies it as a man wearing a hat and a hat on a skateboard So I don't know how it got that from that picture and then the bottom one is a man is talking on his cell phone Well, another man watches. Well, he's eating a sandwich. So I think we still have some work to do So let's take this image of a panda and the computer says well, it's a panda with fifty seven point seven percent confidence Now a good fellow and others have shown that you can actually add some noise to this image In this case the noise is identified as a Nemo toad with eight point two percent confidence It didn't really do a good job there, but the purpose of this noise is to actually to trick the computer. So The end result still looks like a panda to us, but to the computer It has been deliberately tricked it no longer looks like a panda to the computer In fact, it looks like a gibbon to the computer with ninety nine point three percent confidence And so it's been it's been fooled. So yeah, we still have some work to do there now a lot of development for intelligence has been influenced by animal brains Neural networks obviously or originally influenced by biology and so actually just very very recently We've been able to look at what does the actual human brain look like or in this case What does the mouse brain look like which is at a very low level is actually going to be similar to a human brain? And so here we have one axon in blue which is used for transmitting signals connecting to a dendrite Which is in green which receives signals through five separate synapses which are in orange And so you multiply this by a hundred billion and you have the human brain So let's zoom out a little bit It's remarkable how we can do all of these tasks when at the end of the day it If you look at our brain at a low level, it just looks like a really big jumble of molecules So let's let's do a let's do a comparison between a computer and a human brain. So computation Computer about a billion transistors human brain about a hundred billion neurons Storage, okay We have some RAM and some hard drive space in the computer the human brain again has neurons So I think one of the more remarkable things about I guess the way that we or animals function is that neurons are used for both computation and storage And we obviously we we have a separation of this stuff for computers, but we don't have this for for biology What about the cycle time well 10 to the negative 9 seconds for computers and Unfortunately, we're not that quick 10 to the negative 3 seconds for the human brain Operations per second about 10 to the 10th for the computer However, 10 to the 17th for the human brain, which I think is remarkable that we Perform seven orders of magnitude more computations per second more operations per second than a computer Obviously, we're not really good at doing floating point operations at that rate But it's still kind of fascinating and then memory updates per second similar about four orders Four orders of magnitude in the human brain than in a computer so Let's let's take that concept and what if maybe we can bridge the gap a little bit Can we form a network with multiple animal brains that co-operate an exchange information? Using brain-to-brain interfaces So possibly era actually this year have shown that yes it is this is in fact possible So this is you know I think a really good stepping stone and they actually did it did this with rats and so they had electrical Nodes that were connected to these rats brains that they would stimulate and these these rats were separated and then they Would basically if they synchronized their brain activity they would get rewarded with water And this was really fascinating because it turns out that for connected rat brains are better at Solving problems than a single rat brain So you actually do get an amplification of intelligence in this way and they even did use them for really interesting things like for example weather Forecasting it turns out four rats are better at forecasting the weather than a single rat really I don't look a really fascinating kind of result So what about what about science what about science questions? Well This is the sort of question that a Ten-year-old may be expected to answer what form of energy causes a nice cube to melt and then you have four choices And in this case obviously it's heat and so Hicks and then others have shown that you can actually build a knowledge graph and It'll have enough information that it's actually gonna be able to answer this question Here's another question in which environment would a white rabbit be best protected from predators now the word white Doesn't appear in any of the questions so The system has to be able to infer that one the color of the rabbits fur is actually relevant and the color of snow is similar enough to the color of the fur and then that is important in terms of Camouflage and protection from predators so the naive or completely unintelligent Result for this is to basically get 25% of the questions right because you get four choices. It turns out that the The software is able to get 57% of these questions correct. So it's actually doing reasonably well Now let's take a hypothetical person this credit for this goes to Josh Tanonbaum We take a hypothetical person called Boris so you know going back to our previous example where we have the knowledge graph Maybe we just build a huge knowledge graph and then we have solved intelligence. Is the mother of Boris's father his grandmother Yes, is the daughter of Boris's sister his grandmother? No, is the mother of Boris's sister his mother well if she's a biological sister and the answer is yes Now what about this question is the son of Boris's sister his son? Now you want to say no But Boris and his family were stranded on a desert island when he was a young boy So that obviously changes the probability of this answer But what if they were stranded alone or with one other family two other five other 50 other families? What if they were rescued when he was 12 or 15 or 20 or 30 or 45 years old? And so you can't just rely on a knowledge graph because all this other context is important in terms of Influencing the probability of something like this. Let's look at another one is the son of Boris's wife his son Well Boris is an international businessman who often takes long trips away from home But he gets very jealous so he locks his wife in the basement during his trips Unbeknownst to him his wife used to be a professional escape artist So in this whole thing around context It's not what I'm just I'm quoting I'm quoting Josh Tenenbaum on I understand your concern I'm just quoting Josh Tenenbaum on that stuff. I'm not saying that like those things are what they are I just Those were the those are a standard set of questions that he came up with or a standard standard of probability Influencing things that he came up with. Yeah, but everybody laughs at least too many people. So thanks. That's not funny Okay, so in the in this whole sphere of context We have some questions that are very easy for us to answer that are hard for Computers to answer so the trophy would not fit in the Brown suitcase because it was too big What was too big? Well, we know right away that the answer is the trophy But it's actually very very difficult for computers to know this Answer similarly the town councilors refused to give the demonstrators a permit because they feared violence Who feared violence? Well, it's clear to us that it is the town councilors who fear violence but because we understand Demonstrations and permits and town councilors and who may want who may incite violence And it's actually very very difficult to to program all of this additional kind of human understanding and context Into the software so that it's able to effectively answer these questions so vinyls and Lee have developed an interesting program where they take They take help desk conversations and they try to basically build a software that is able to answer questions within a help desk kind of context and so for example You may not be talking to you may not know you're talking to a machine But in this example you are so hi, this is name from help desk connect What can I help you with today? And then the human says Hi, I forgot my password Okay, great. So it's really funny that The machine in this case has actually done a pretty interesting job of replicating and helped us conversation What about adaptability Siri play some cool music, please? You don't have any cool music obviously it doesn't Doesn't realize the context ambiguity is is Mickey larger than Pluto Perhaps I mean we know what we're talking about here Is Mars larger than Pluto? well Different different things here. So now how big is Pluto? There is a chatbot called Mitsuko. That's one of the one of the current State of the art chatbots. So for example, you can ask something like what is bigger an elephant or a person I would say an elephant is a little bigger than a person Well, I mean, I guess most of us would not say little but that is correct How many presidents of the US were called Bush? No idea. I am not really into politics Should Greece leave the Euro? I think they need to seriously consider it. It just isn't working Should fool leave the bar? Maybe you should ask someone older Now what about the proliferation of some of these systems and the impact on privacy? For example, Facebook's new photo app won't launch in Europe because of facial recognition So there's a question here of regulation and AI This is a feature. I mean the fact that you may say like okay Well, it doesn't really matter what is behind it. It's it's the privacy aspect of it but I think the proliferation of AI is Is related to this Where we actually are seeing AI being regulated obviously from the feature from what it does But I think it's interesting. I think we're gonna see more of this in the future There's been a lot of talk lately about evil AI Elon Musk Stephen Hawking have spoken about this Benjamin Bratton says something interesting what we fear more than a big machine that wants to kill us is one that sees us as Irrelevant worse than being seen as an enemy is not being seen at all The ultra intelligent machine is a machine that believes that people cannot think check good. So is this our future? I don't know How do we find out how good we're getting and or how close to that potential evil future? We are well. We can apply some other tests for example the reverse Turing test where instead of Person trying to determine if someone is a computer One of the people tries to convince the person that they're actually not human and they're a computer and obviously turns out the better You are at understanding Computer science the better you are at this turn as at that test Subject matter expert Turing test can a computer Replicate an expert in a particular subject matter The coffee test I think is interesting can a machine go into an average home find the coffee machine find the coffee and Go through the entire set of steps so that there's a cup of coffee at the end I think it's it. That's that's a pretty interesting one The eber test is can a machine tell a joke that causes people to laugh and on the subject of computer generated jokes Just Johnson built a list program that generates jokes based off of homonyms So for example, what do you get when you cross support with frosted flakes? I don't know, but it's cereal What do you get when you cross an alien with a chicken an extra terrestrial? Other tests the robot University student by Gord soul can a robot Go to college go to university and roll in classes pass those classes and get a degree Obviously, we're very far away from that employment test can a robot or machine Have a job and perform that job as well as the human in that job and obviously for certain highly automated tasks This has already been done and the total Turing test where It relies on manipulating physical objects and being able to see the world So here's some more thoughts for the future about what it can't currently do or maybe it already does Can it beat humans in the stock market? Maybe if someone was able if someone is already doing this would they tell you probably not Can it make this presentation? Can it make any presentation? Can it advocate against regulating AI thereby ensuring its future survival? Maybe make an argument for more funding for AI interesting What about devising a secure cryptographic scheme? What about defining human rights? What about promoting human rights? What about improving the quality of human life? Now, I think it's interesting that you know, we kind of think about like building these perfect machines George Dyson has a great thing where he says instead of trying to build infallible machines We should be developing fallible machines that are able to learn from their mistakes So I'd like to leave off with some fun with image recognition This is the Wolfram Alpha image recognition system. And so what what's in this picture? An AT-AT. Nope. It's a SkiPool What about this one? It's a cement mixer How about this one a Flying boat What about this one a Hunting dog How about this one a stealth bomber and How about this one a Person and That is all. Thank you. Oh, yes. So this was Luke Gotchling and If you have some questions, you can ask them at the microphone there and there I Have to excuse myself because I forgot to announce that this talk was translated to German as every other talk No is translated and you could have heard it when probably everybody who need this already knows this numbers and Yes, sorry that I did not announce it, but you can clap for them because they did it anyway So I don't see anybody or for the questions So well, yes, then There's someone coming up left microphone first, please or the right for you Hey, thanks for your talk. A lot of the examples you gave were actually rather narrow AI right like Programs that solve specific Problems, but you're gonna talk about ADI. So what efforts do you see in that direction? So I think IBM's Watson is kind of the one of the most Remarkable ones and then the kind of it like the chap I like Mitsuko for example, you know asking anywhere from political questions to Is an elephant larger than a person? I think things that kind of do general knowledge are are kind of the what we're currently You know considering the closest to artificial general intelligence So I think it's really just things like that I mean obviously we're still very far from the coffee test where a machine is able to enter a home and make coffee a Lot of the obviously a lot of research is focused on improving things and you know It turns out that it's easier to make improvements on very very specific domains But things like IBM's Watson answering general questions including questions with word play and things Like the chatbot where you can ask all manners of questions I think are basically as as close as we're getting and there's some interesting research in those areas I have a small question concerning the red-brained thingy Did you check if the significance of that is like valid? I mean probably you don't know is easy significance ever but was it really tested or is it just some kind of Social science statistics no no it was it was statistically significant and in in all of my slides I have references, so if you download the slides I actually have a link to that original paper so you can actually read that whole paper yourself It was actually no it was it was a statistically significant. Okay. Thanks And the last question also from that microphone. I had Yeah, I tried some General game playing. I I think you know what it is and I wanted to know what you think about this field of Artific art if he yeah into artificial intelligence Well, the game playing you say yeah general game playing like so like like the the recent thing about like playing Mario Yes, but general game playing is you have a artificial intelligence that gets the rules of a game and When plays the game so it doesn't know what game it plays before the play starts so I think that's a really interesting field of Artificial intelligence. Yeah, I mean you could do some of that stuff with with unsupervised learning In the case of there's a some more famous example lately of playing Mario where basically they just try Different things and then like basically replay the game like thousands or millions of times And then they learn up to a certain point and then they try something and then if you die Then they restart and then they try something else So it's interesting because it's effectively completely unsupervised It's like if if you die in the game, then you just Then the software restarts and they try something different going forward so we can actually learn to play that game in a way that That you know like then once it's fully trained on that game when you see how it plays It actually looks like a human player playing of a game because it's doing all of the moves to get through a given level Yeah, thanks So again a nice applause for Luke Gotchling. Thank you