Ron so there are children alive today that would not have been born if not for machinery so maybe it is not so monday and after all. Another example. The smart phone in your pocket right now is chock full of dirty algorithms using a Machine Learning to address to read you are saying to correct your spelling errors, predicted he will do and make suggestions using gps and you can compare that with your calendar to see if you will be late for meetings. All this is happening today with the help of Machine Learning and as we progress cell phones and what Companies Develop will learn more and more and as a result with all of this economic value is no wonder that Tech Companies are all over it they are investing heavily in Machine Learning so google recently paid half a billion dollars for a company that has no customers and no product pitches had better learn the of the rape of rhythms that led to predict what people will click on. That alone is worth half a billion dollars every year as another example of just this month with the medical Imaging Company so they can trade in those learning algorithms just like for Breast Cancer today it takes doctors to do initiate a and Machine Learning with the director of research at microsoft says what you pay is a hot area in Machine Learning is comparable to a top nfl quarterback. Also Machine Learning has a dark side which we need to be aware of. Increased automation of whitecollar jobs it could cost you your job. Some people say nsa uses that to spy on you. We dont know because it is secret but also it may lead to a the terminator of evil robots taking over the world. Mitt could be your best friend or worst enemy depending on what you do with that which is why were at the point of Everybody Needs to have a basic understanding of what it is and what it does and why i have wrote a book about it. You dont have to under state and the gory details it is like driving a car even after a understand how oh the injured works the you do need to understand what happens when you push the pedal. But today you dont even know that it exist. So we as a society have a lot of decisions that we need to make if we need to be informed. Even more so than in the past. So how is it the computers do this thing . It may be surprising it is even possible. Reprogram them to do the same repetitive task over and over and discovering new knowledge is supposed to require intelligence and creativity they said computers are useless because they cannot they give us answers the Machine Learning is what happens when you ask questions. It continues to ask is how to return input into output . For example, so pretend the input is the xray of a prestigious trying to learn there is a tumor here. And it does this by looking at data. In the old days before a Machine Learning data and the algorithm went into the computer again think of the Breast Cancer example the output may be the diagnosis. And the algorithm is a secret a sequence of instructions telling it exactly what it needs to do. It is like a recipe for rigo. Use these ingredients. Except it has to be more detail because computers are down but with Machine Learning it turns around. Now the al point also goes in and with the Machine Learning of rhythm does is figured out once they learn that algorithm that i need need, if this is the image of the breast how do i get to one from the other . The amazing thing of machine and learning is when you program the traditional way it takes human programmers and for every application to write a Different Program one of her soaring, a diagnosis, etc. , etc. But Machine Learning algorithm can learn all knowledge that have been learned by learning of a rhythm so the whole holy grail is to discover the master algorithms one that makes others the single algorithm to discover of knowledge if you give it the appropriate data if it is about credit risk it will run through credit scoring press cancer diagnosis it will do that. What could this possibly be . Then we can talk about the future consequences next part of what makes this interesting these algorithms covered interesting origens for each of these there is the school of thought the five kinds of Machine Learning what idea is it is built, and what the algorithm dust does. There is the single list they have their origins and logic and their master algorithm of induction as the inverse then those connection is to use big idea is to reverse engineer the brain so they tried to implement the brain on a computer and then that is called back propagation. But there grid in the rest is you. A redo regard evolution as a kind of the algorithm than there are those that have origins is statistics in may will see what that is as well. And those that have their roots and all learning are intelligence am probably the most famous is well see what those do. So here are three of the most prominent in the world from carnegiemellon. He was also on my ph. D. Committee so what is the idea of basing the list . The most direct version by filling the gaps in our knowledge by hypophysis in the basic idea is Interest Deduction and with the idea how do we deduce things . Deduction goes from the general to the specific this goes on the specific to the general so think of thats lower integration is part of differentiation. For example, edition is the answer to an end to what do i get . Attraction is the answer what do i need to ask in order to get for . And in exact parallel with that deduction follows the you go to humans are more toll to guess what . This comes from general knowledge here but the inverse says it is it my data then what am i missing . And of course, the answer is i need to know that humans are mortal. Here the rules are written in english but for the computer is written in formal logic so the idea is the same then you run these rules and you can do something that you cannot combine them and arbitrary ways that has never been done before so this is the basic idea behind symbolism. Here is an example of what you cant do. Guess to the biologist in this picture is . Not just a guy in the lab coat. The other guy is not a biologist either. The biologist is the machine. This machine is a robot that news about dna and molecular biology and is making discoveries about metabolism , diseases and runs the process all by itself with inverse deduction to make a hypothesis and it literally carries out the experiment with dna and sequencing to test the hypothesis. Then maybe it throws them away or keeps going. Right now there only to appease computers in the world there called adam and eve and he just discovered a new malaria drug that is being tested. And is like having 1 Million People doing medical research. This is a very powerful thing to do. I am not big fans of this approach if it is too abstract or too rigid for them to be making deductions on paper here are the most prominent one the steadied as a psychologist his goal in life is to understand how the brain works he has persisted through thick and thin and has made a lot of progress in and in fact, he tells a story of coming home one day from work very excited to say i have done it i have figured out how the brain works and his daughter said not again. [laughter] another famous connection test it is very much in the news literally on the front page because of the success under the name of deepening. That is the name of connectionism is the name of deep learning. So how do we do this . How do we build these learning algorithms . The brain is made of marylands better kind id like trees then their roots are called dendrites and branches into more dendrites so it is a little bit like the forest with a difference the roots of one tree connect with the branches of others then it is like a of a big electrical storm happening. They are fired by the neurons and it is happening. So the connections might be stronger or weaker. The structure the connection the more like the the first run can make the second fire. With that amount of electricity coming in as one fires it could make others fire that is a network of 10 billion. So number one we will build the simplest mathematical model that we can it comes as it and put it could be a pixel on an image on a of a graph of a cat or of a face and that represents how strong it is. What is the weighted average of the input . If it is the threshold of the output is one. And now we take this to connected to the deep network with layers deep learning comes this way because it is all done in layers. But another problem you have to solve is how did you learn that . It is tricky because here is an image of a cat it is supposed to be one but it is 0. Two. The question is how do make it go up . There is the search and errol and now i tried to seek damages responsible . And then you turn each of these as it comes from the previous letter. The one that causes it to be too low so that propagation that this is some evidence some parts of the brain work like this algorithm. When something goes wrong who do you blame . When it gives the wrong answer something needs to change so it gets a errors back to the network to see where the synapses need to change. Now the states in the last few years it has amazing processing like recognition and translation to this type of Machine Learning. El google does the image schurzs search to understand the simultaneous translation you can talk live in a different language theyll use this. But a famous example of the first pages of the New York Times was the Google Network at the time was the Largest Network ever built and has of the order of 1 billion synapses which is still small compared to your brain but human best compared to those who went before. And basically its that theyre watching you to videos for of longtime media should a couch potato. [laughter] then people like to upload videos of their cats it could be known as the google cat Network Although it could recognize dogs and people and others. The of the missionaries have a different idea that say where does that trading come from . But somebody had to invent the brain to have that accomplishment and evolution hope to accomplish that. The big pioneer of revolutionary learning just died recently and in fact, for the first couple of decades people would say evolutionary computing was john holland and his students then in the 80s he proposed genetic programming to the master algorithm and then there were doing fun things with evolutionary learning. So the basic idea of evolution there relearning . The process of the evolutionary nature. Ever since darwin because know how genetics work so it happens in this thematic way anytime you have a population that has a genome then they get killed or not and they have a logger neck so the individuals in particular cross over sexual reproduction of to a successful individuals when the evolution also happens if the dna turns out not to be good. So after that happens you have a new generation of genome and the process repeats again that they transfer from the envy but to a human being but now it is just bits from the computer it is like having dna. But otherwise the process could do quite Amazing Things like have circuits to do radio reception or to invent a low pass filter these are better than design by the cuban ventures for those circuits so they could make the path for the invention. The bar of this genetic algorithms or genetic programming why do this with strings . It is low level representation that the program is really a treat of operations except that to cross over into two strings so if you pick this point for crossover then one of the offspring is the tree in white the other is the tree in black the one in white is kept alive the formula for the duration of a planet here ebullition of the sun. It is the square root of the distance and you can actually discover this using genetic programming. Now perhaps the most interesting and ambitious area of application of genetic algorithms is actually not to do this in the computer but in the real physical world. So where you literally have these robots at cornell there performing whatever perform the function we want them to perform them the robot to the highest fitness can program with a three the computer to create the next generation of robots finally may have robots under very robust to crawl around that can adapt. Traditionally they did not. So the next nine eupepsia spider it could be one of these. Hopefully we would keep this robot under control so then that brings up a bunch of other issues. The most fanatical of all learning their fanatical for good reasons but that paradigm comes from statistics the they were a persecuted minority so they have to give very hard core to survive and it is good that they did. But now but now the most famous who actually won the award of Computer Science other famous bayesians of those pictured. So bayesians takes its learning they think there is a formula for the world war Machine Learning like e e mc2. For so it actually has a base pair of. Than they put the picture outside of their office. So what is this gobbledygook . Let me give you basic intuition of what is going on. They have a problem of the Machine Learning that it tries to solve. For missing the last of compensable knowledge or the problem of uncertainty that you never know anything for sure. But a moment is never certain proposed else could have been the case so we consider all possible hypothesis that the same time the net changes have much to believe in each hypophysis. So you start off with a prior that is how much you believe in each hypothesis before you see any data. Even before you know, anything then when you see data there is the likelihood if the hypothesis is true how likely a by to see this . Then the idea is more likely make the world that use the bend the world make sure hypothesis value multiplied at by a the likelihood that is how much you believe in your hypothesis after you have seen the data if you see more in board data then they should become very unlikely that eventually theres just one hypothesis but in the meantime you may consider a bunch at the same time. There is also marginal probability but we will not worry about that. Doing this in practice can be very difficult the computations can become exposed to inexpensive but there are clever ways to do this and very interesting applications exist one they were familiar red the first most prevalent way is to use a vision of learning with that tries to do is to decide when a log that email if this ban or not. The priority is even before it sees your in box how likely . Ninetynine . Take your pick then the evidence is the words in the mail if it contains more exclamation marks in sequence it is likely to be spam with it contains the words your mom or your boss you just want to be sure. Said is one big application another the google ads the algorithm is uses is part of the Vision Network with millions of notes to predict if you will . Of that or not. Then the more money rigo would make some of the happier you should be of the advertisers. Finally talked about the analogize hours theyre not as cohesive is a loose grouping of people of i analogy this is they have extensive evidence people do a lot i needed to solve a problem i will look for a similar problem in my memory look for the most similar symptoms if they have the flu then this patient could have the flu. This is just the simplest form one of the of pioneers of the of leading analogy based until it cable long this was the most important and also a Douglas Hofstadter probably the most famous analagizers of all also a scientist and he coined the term analagizers and in his most recent book is 700 pages arguing everything all intelligent thinking is just analogy. He very much believes in a logical reasoning is the math of it all. So let me propose up a simple puzzle it ended is the simplest machine running that there is it was invented many years ago but it is still pretty good i have a map the two countries over here and show with the plus and every year with a negative sign. Just of the location of the cities can you tell where the border is . That is a hard problem but probably a good idea is the following any point on the map closer to aplus city getting dash city is probably part of the positive country. They are exactly the same distance so this is just reasoning by similarity. Now instead of being a backup it could be Breast Cancer imaging. It could be used for all sorts of different things. This border is a little bit jagged too many sharp lines and straight quarters another problem if you think about it you could throw ladies examples if you throw away this guy and this does not change. Then you have millions of these examples and then it is a heavy expense. They only keep the support factors with the following exercise you will always want to walk with the positive cities on your rights day as far away as possible think of it as a land mine while step eight very far from any land mine with a margin of safety and this is the machines do. It is just the technical name how you measure a the similarities between two objects in the world. What me just give an example of the message are at work in the world. They use all sorts of algorithms is still one of the most powerful if i figured if you like a certain movie one of the best things i can do is look for a similar taste to yours. Are the ones they did not like the and this idea worked extraordinarily well i have seen in the several places onethird of amazon business comes to recommend and three chords recorders come from the recommendation and system. This is the basic idea. Is quite interesting to notice those similarities and they have all been successful of what they do. Here is the big picture that it has worked on for a long time with the symbolist knowledge competition and so forth. People believe they have the master algorithm. To say that is all we need. And then we will be done. All of them are real. The the dozens of the header problem. With your artificial brain. So what we need and what we have been working on the last decade the the single master algorithm. Looking now the standard model was were looking at Machine Learning to solve the problems of balance to do all the of learning that they could. We have not found yet but there was one way to create such an algorithm. And all looked very different the all had that same structure but is the choice of programming language in which to express what it has learned it is not like job of or c because that is too complicated. Or maybe Vision Networks the you have to use their representation because of the power of logic that this is something that we have done successful lead is a combination of logic and of the typographical model. End in essence that have the formula and the logic and the analog of sentences and english and then becomes a lot less likely. And then those high probability formulas. To start with that representation that every learning algorithm has is part of the evolution equation. If youd give me a algorithm how good is it . To diagnose Breast Cancer. And then to see how often do they give the right answer . House small are those that they generate . They can all be included in the probability that we talked about that diaper a simpler hypophysis. Braddish itd be part of that mission is to be whatever you care about. So that his luggage should be. If you want to be happy and that is what they should try to learn about. Then it goes about the feeling that. Optimization by which that finds the program and has the highest possible score of 98 to discover in the formula in the first place. It was the type of programming language so this is currently the best shot lead is so far from the answer and my suspicion with the master algorithm requires ideas we have not had yet. Said the is ideas that we talk about but there are important things we have not discovered yet end date experts in dessau hard time with al qaeda. And if not maybe i could publish that. Des those of we are working on for a long time. And also it is not good enough. With those issues the of the example to half of robots cooking for us why dont they exist yet . To be on the fly and the reason of analogy. It is a structured environment. But to do the other things to do in the home. What those sold driving cars that will generate. And then to read the web that the computers can reason with so in the future then you get the actual the answers to the questions the most famous example is the knowledge craft. End in order to do this to do without Machine Learning but to be that symbolic representation so that probability of the other thing that will not happen but it depends on the learning. Even the most important of all is cancer. The reason it is a heart problem it is not a single disease. Every buddies cancer is different and it mutates as it grows the cancer today is the the save as six months ago so it is unlikely there will be a single drug to solve all cancer but it is a large the program the patience genome and medical history to see it that it predicts this is what they need for your cancer or the culmination designed specifically for that. But it involves all of the same time. The to understand that structure look at those parameters. With the reason by analogy. But once they have an algorithm based on the data we could have such a program. There is already a large effort that tried to collect the data and people like david believe as they share the data at some point to care most cancers and if they dont then they probably wont. It depends on the of rhythms and the data i will mention in one last thing. What that will affect you in your everyday life. These the recommended system recommends the movies eric is a lot of these but what this marco in your pocket right now is trying to run the 360degree picture so from every decision they have to make whether picking a movie to see our house to buy or a date this model can propose things to help you go from a million options down to just a few. But in order to learn that model renewed war algorithms but once you have added is moore dispensable bed youre smart phone is today. The model will basically help you do away with information then your model is a track with my model then you go to find me a job they elect all possible jobs sarah model works with the rest of the country or if i want to go on a date with regard a thousand days with a million different people here are the most promising this all goes on under the hood so there is a lot of impact but he is saying that daybed that has been produced that we should be able to come up with as i describe. Thank you. [applause] there are microphones on either side of the stage. Please keep your questions concise. What are your thoughts on decomposition with a master at algorithm . Using first order logic. The simplest believe the composition is the central for intelligence to solve a problem then to solve and solve a problem then combine them together but others say no because of the big network this is one example where you do need the simplest but the other ones dont so much it is unjust for like problemsolving but understanding when a bridge that is great composition because i cannot combine into an infinite number of sentences. So that would be like the simplest. Some like Stephen Hawking sorry law and must is worried about progress of Artificial Intelligence and others believe and of the singularity. Day you have an opinion what we need to do of regulation or if you fall into one of those camps . This is a prominent question as it is all over the media there are utopians like elon musk and i believe neither side is right i think the problem people like Stephen Hawking in in elon musk are worried that Artificial Intelligence over but i dunno to many experts that have the terminator scenario seriously. This is natural they confuse intelligence with human than they think it will have the same desires that we do then compete but a. I. Is very different from human intelligence a kabir infinite the intelligent but no danger provided the only solve the problems that we give them we set the goals they figure how to achieve them. Would not be a good jab and infinitely Intelligence Program . Yes. So provided redundant anything complete of down with some safeguards on either hand there are some problems one of the reasons why Everybody Needs to have a rough understanding the technology has the potential for the good and bad. Theyre all building models. Ended is a little bit dangerous you want to make sure your part of the conversation now there are other dangers like jobs i do think there are a lot of jobs that will be automated and and a lot of half dead and more of them will be. But how is my job . If it is that i need to get on top of this to do the things the machine cannot do machines dont have common sense. Enable lot for a long time. And to understand the context those required human intelligence. I dont think in the short term but right now the ad implied rating is 5 again but people need to retrain themselves to do the new kinds of jobs that a. I. Does not make possible. I dont think this will happen tomorrow but if we do get to that point then there will the new more jobs but then that is a good thing because people look dont have to work then they find meaning in life in other ways. So how will that will be shared . If it is 1 quadrillion theres been that is not the outcome. That his wife and a democracy our boat is very important with the idea of a basic and come like in alaska everybody gets a check from the State Government every year from their short there share of oil revenue that the machines can do what they need to do for production but given the nature being what it is people will want to make money may be some are necessary. But in the end everybody would be much better off for prod finally the single biggest danger you tell it what you wanted to do and the computer does it put the computer is dumb so does the wrong thing it does what you literally said instead if you wanted it to do or does the wrong thing because a dozen understand code or just makes a mistake like with credit scoring the computer may decide i will not give you a credit card even though we should. But the cure for that is to make the computer more intelligent people worry they will get too smart to take over but they are stupid and they have already taken over the world. [laughter] thanks for your talk. I am curious, first, what do you think are some of the toughest problems yet to be solved with the master algorithm or Machine Learning . Second, how long do the anticipate of the master algorithm how long will that take . And a lot of progress has been made. First there is the barrier that combining this is extremely difficult for example, combine in logic and probability is something that others have worked on for 200 years or more. There is one problem we have made a lot of progress were well on the way to solving it but now they combine these things to read a time. To add a time so maybe they will have a completely different idea in ways that people have not thought of yet then there is the question that it can learn commonsense knowledge . How much to apply common sense . There is a school of thought that says if you read the novels and textbooks that is all the knowledge that you need to have. Like they really are trying to solve this at the end of the day but as much as i thank you can get mileage in the end you could only acquire common sense by living your life and the only way is to move around in the physical world and getting up and learning how to do things. Media brings the problems we will lot even see lings the proe will lot even see like a self driving cars. Or things that need to gradually get better like the roomba. When will rediscover the master of rhythm . You have periods of no progress then suddenly a lot this is what happens with vision and speech tender 15 years nothing got better now suddenly we have a vision in a speech that is practical. This depends on peoples creativity. It could happen tomorrow or many decades or some say it will never happen but one of the reasons i wrote the book to make it possible for other people to come up with these ideas have a scientist has an idea to solve the problem that is the best outcome i could possibly hope for. With the approach to solving problems changes so much of how theyre learning in schools and in universities and it has to change a light of it. Very much one of the biggest impacts is the thing to realize people learn very slowly it is difficult and there has to be a better way on the one thing people watching videos or alerting online and the critics of the Online Learning that doesnt have that personal interaction like the classroom but interactive component is very important. But when youre in the classroom it is not that attractive an the other Thing Companies that are starting to do this. Everyone of these interactions with a computer something that you ask about teaches the computer what are the difficulties . It is not just from you but from everybody else. So you will have of a better computer to a teacher much better than any other could. People have been working on this for a long time