Coming up, tomorrow night ross park will read from his new novel, who are you, thomas bledsoe. Stephen kinzer will present a history of the cia. We also have some other events next weekend tickets are available for mondays conversation and another talk on wednesday. Tonight we welcome gary marcus, author of rebooting ai building Artificial Intelligence we can trust which argues a computer being a human in jeopardy does not signal we are on the doorstep of solely Autonomous Cars or super intelligent machines. Taking inspiration from the human mind, the book explains if we need to advance Artificial Intelligence to the next level, suggests that if we are wise along the way we will not need to worry about a future of machine overlords. Finally a book that tells us what ai is, what it is not and what a i could become if only we are ambitious and creative enough. A deeply informed account. Gary marcus is founder and ceo of robert robots. A ing metric intelligence and published in journals including science and nature, the youngest Professor Emeritus at nyu. Along with ernest davis, rebooting ai building Artificial Intelligence we can trust. Thank you very much. [applause] this is not what we wanted to see. This is not good. Maybe it will be all right. We had technical difficulties. Im here to talk about this new book rebooting ai building Artificial Intelligence we can trust. You might have seen an oped in the New York Times called how to build Artificial Intelligence we can trust. I think we should all be worried about that question because people are building a lot of Artificial Intelligence they dont think we can yet trust. The way we put it in the piece, Artificial Intelligence has a trust problem. We are relying on ai more and more. It hasnt turned our confidence. We also suggested i want to suggest theres a hate problem. A lot of ais overhyped these days by people who are prominent in the field. Andrew yang is one of the leaders of deep learning, a major approach to ai use days. A typical person can do a mental task with only one fog, we can probably automate using ai now or in the near future. A profound claim. Anything you as a person can do a second we can get i i to do. If it were true the world would be on the verge of changing altogether. It may be true 20 or 50 years or 100 years from now but it is not remotely true now. The trust problem is this. We have things like Driverless Cars the people think they can trust that they shouldnt and sometimes they die in the process. This is a picture from a few weeks ago in which a tesla crashed into a stopped emergency vehicle. This happened five times in the last year that a tesla on autopilot crashed into a vehicle on the side of the road. A systematic problem. Here is another example. I hope this never happens to my robot. This robot is a security robot that committed suicide by walking into a puddle. You say machines can do anything a person can do. A person can look at the puddle and say i shouldnt go in there but robots cant. We have other problems like bias. A lot of people are talking about this. You can do a google image search for the word professor and get back Something Like this where almost all the professors are white males even though statistics in the united states, only 40 of professors are white males and if you look around the world it is lower than that. You have systems taking a lot of data but dont know if the data is any good and are reflecting that now and that is perpetuating a cultural stereotype. The underlying problem with Artificial Intelligence is the techniques people are using are too brittle. Everybody is excited about deep learning which is good for a few things, actually many things. Object recognition, get deep learning to recognize this is a bottle or this is a microphone or get it to recognize my face or distinguish from michael teds face. I hope you can do that. Deep learning can help with radiology. It turns out all the things it is good at fall into one category of human intelligence. The category they fall into are things called purpose up to a classification. You see examples of something and have to identify further examples of things that look the same and sound the same and so forth but that doesnt mean that one technique is useful for everything. I wrote a critique of deep learning a year and a half ago you can find online called deep learning, cortical appraisal. You can find it online for free. Deep learning is greedy, brittle, opaque and shallow. Everybody is excited about it but that doesnt mean it is perfect and i will give you some examples but first i will give you a real counterpart to the claim that if you are running a business and wanted to use ai you need to know what ai can actually do and what it cannot do. If you are thinking of ai ethics and what machines might do soon or might not do soon it is important to realize their are limits on the current system. Of a typical person can do a mental task, with one second of thought and we gather enormous amount of data directly relevant we have a fighting chance to get our ai to work with that if the test data that we make the system work on are different from things we taught the system on or are not too different from what we talked about in the system doesnt change much over time. The problem youre trying to solve doesnt change much over time. This is a recipe for gain. What ai is good at is fundamentally things like games. Alpha go is the best in the world, better than any human and fits in what i specify the system is good at. The system hasnt changed, the domain, the game hasnt changed in 2500 years. We have a set of rules that you can gather as much data as you like for free or almost for free. There are different versions of itself in order to make the worlds best player. Compare that, you dont want to robot the does elder care to collect an infinite amount of data through trial and error and and fitting grandpa into bed 5 of the time. When it works the way deep learning works, it is fundamentally taking big data and statistical approximation and label data to label a bunch of pictures of tiger woods and golf balls and pictures of angelina jolie. And all pictures of tiger woods, identifies it, tiger woods and angelina jolie. This is a sweet spot of deep learning. Wired magazine had an article saying deep learning will soon give a supersmart robot as well. We have seen an example of a robot that is not all that smart. It has not been delivered on. Lots of things even in perception and something of which is reading. On the right are some training examples, you teach the system, another elephant that looks a lot like those on the right, there is no problem at all and say it knows what it opened is. On the left, the picture on the left the way the deep learning system responds, it says person. It mistakes the silhouette of an elephant for a person, it does not do what you were able to do which is recognize the silhouette and the trunk is really salient. It is extrapolation or generalization and deep learning cant do this. We trust deep learning more and more in systems that make judgment whether people should stay in jail or get particular jobs and so forth and it is quite limited so heres another example. Kind of making the same point about unusual cases. If you show a picture of a school bus on its side in a snow bank it is with great confidence that is a snowplow in the system cares about things like the texture of the road and the snow, no idea the difference between a snowplow in the school bus and what therefore. Fundamentally mindless statistical summation and correlation. This thing on the right was made by some people at mit. If youre in a deep learning system you save an espresso because there is a phone there, not supervisible but picks up the texture of the phone and says espresso because that is the salient thing about espresso. It doesnt understand another example, show deep learning systems a banana and put the sticker in front of the banana which is a kind of psychedelic toaster and because theres more color variation in the sticker the deep learning system goes from calling the top one a banana to calling the bottom one a toaster. It doesnt have a way of doing what you do which is to say it is a banana with a sticker in front of it. It is too complicated. I can do is say which category something belongs to, that is all the learning does it identify categories. You are not worried that this is starting to control our society, youre not paying attention. The next slide out of this . Maybe not. Going to have to go without slides because of technical difficulties. I will continue though. This is something we can do. One second here. I was next going to show you a picture of a parking sign with stickers on it. It would be better to show you the actual picture, is not going to work. And it is called a refrigerator filled with a lot of food and drink. It is about colors and textures but doesnt realize what is going on. That i show you a picture of a dog doing a benchpress with a barbell. Something has gone wrong. Thank you for that. I would need a mac laptop and i couldnt do it fast. How did the dog get so gripped that it could lift the guard dog. When it comes to reading im going to read you a short little story that Laura Ingle Wilder wrote. He finds a wallet full of money that is dropped on the street. That the mullet might be that the wallet might be someones. Did you lose up pocketbook. Guess i had 1,500 in it what about it. He opens it and council money he breathes a sigh of relief. That darn boy didnt steal any of it. You form a mental image of it. You know and infer a lot of things that the boy the wallet that you will have it. You can answer questions whats going on. There is no ai system that can actually do that. First of all it is famous because of elon musk and he found it. There to give away all of their ai for free. They gave away all of their ai for free until they made at this this thing called gpt two. So dangerous that we cant give it away. It was a system that was so good at human language that they did not want the world to have it. People figured out how it worked. Now you can use it on the internet. My collaborator said in the story into it. The guy has counted the money. He now has it and hes super happy. It took a lot of time our him to get the money from the safe place where he hit it. It is perfectly grammatical. If he found his wallet what is he doing there. The word safe place and what are coordinated here. It is called looking for clues. The first clue i think we need to do as we develop ais to realize that it is just part of what intelligence is. There is verbal intelligence. As a psychologist i would say things like common sense, planning and tension. Its where we what we had right now. It is good at doing things that fit with that. Its good at doing perception. And certain times of gameplaying. It doesnt mean it can do everything else. The way i think about this is the deep learning is a great hammer and we have a lot of people looking around saint because i have it everybody must be a nail. And some things actually work with that. There has been much less progress on language. An exponential progress in how will they play games but there has been zero progress in getting them to understand the conversation. The second thing i wanted to say is that there is no substitute for common sense. The picture i wanted to show you right now is up i robot on a tree with a chain saw. Its cutting down the wrong side if you can picture that. This would be very bad you would not want to solve it with a popular technique with reinforcement learning. He would not want a fleet of a hundred thousand robots. Then i was gonna show you the really cool picture of something called the yarn feeder. Its a little bowl in this yarn. Weve enough common sense about it. You can recognize this. You get the basic concept. Thats what common sense is about. How can it show you a picture a picture. Then i was gonna show you the apocalypse. Maybe they should clean up. All right. Then what i really wish i could show you the most is my daughter climbing through chairs like the ones you have now. At the time my daughter was four years old. There was space between the bottom of the chair and the back of the chair. She was small enough to fit through. She did not do what we would call reinforcement learning. She did not do it by imitation. She have never watched the Television Show dukes of hazzard. She just invented for herself a goal and i think this is the essence of how human children learn things. They say can i do this i wonder if i can do that. Can i walk on that small ridge on the side of the road. All day long they just make up games where they were like what if it was like that. She learned essentially in one minute she squeezed through the chair and got a little bit stuck. This is very different from collecting a lot of data with a lot of labels. They need to take some clues from kids and how they do it. The next thing i was gonna do was to quote elizabeth spell view teaches at harvard down the street. She has made the argument if you are born knowing that the objects are in place. Then you can learn about particular objects. If you just know about pixels and videos. People dont want to think that humans are built with notions of space and causality. Nobody has any problem thinking the animals might do this. A few hours after its board. Anybody that has to see this video has to realize it is something built into the brain of the baby i ducks. An understanding of three dimensional geometry. I must know something about physics in its own body. It doesnt mean it cant calibrate it and figure out just how strong the legs are and so forth. As soon as it is born and knows that. It shows a bunch of robots doing things like Opening Doors and falling over. Are trying to get into a car and falling over. You get the point. The video i was gonna show was things that have been simulated. Everybody knew exactly what the events were going to be they were just can have to have the robots open the doors they have all done them in computer simulation. They could not deal with things like friction and wind and so forth. To sum up i know a lot of people are worried about ai right now. There is a line in the book. It would be like in the 14th century. With highway fatalities. What we should really be worried about right now is not some vast future scenario in which ai is much smarter than people and can do whatever it wants. The fact that we are using current ai. On the topic of robots. They are just closing the door. They cant actually open doors. If that doesnt work lock the door. It will be like another seven or ten years before people start working on doors. Just lock the door or put up one of the stickers that i showed you. If that doesnt work talk to in a funny accent. They dont get any of this stuff. The second thing i wanted to say is deep learning is a better ladder than we had built before. Just because something is a better ladder it does it doesnt mean that its necessarily gonna get you to the moon. We have to discern as listeners and readers the difference between a little bit of ai and some magical form of ai that has not been invented yet. We will take as many as you have. If we want to build machines that are as hard as people. We need to start by setting small people. [applause]. Im a retired orthopedic surgeon. Now theyre coming out with robotic surgery. Do you have any information about where that is headed. Right now most of that stuff kind of an extension to the surgeon. Like any other tool. They need to understand the relation. And our ability to do that right now is limited for the kinds of reasons that we are talking about. There will be advances in those fields but i wouldnt expect that when we send people to mars whenever that is we are nowhere near to that point. It will happen someday. In cognitive development. Trade and raised in a human environment. Would it learn language and the answer is no. Other questions. See mac in the current limitations do they also violate that. Self driving cars are really interesting test case. It seems like maybe its logically possible to build them. The problem you get is called outlier cases. If you had Training Data and the things you teach the model are too different from what you see in the real world they dont work very well. The case of the tow trucks in the fire trucks. It is in part. All of the cars are moving fast on the highway. The system see something he hasnt seen before. It doesnt understand how to respond. I dont know if they are going to prove to be closer to Something Like chess or go. Its good to be more like language which to me seems completely outside the range. People have been working on it for 30 or 40 years. Its relatively slow progress. People sought one problem and it causes another problem. The first fatality from a driverless car was a tesla that ran underneath a trail. Or took a left turn onto a highway. You have the problem that it was outside the training center. I havent told but i dont had proof of this. What happened is the tesla thought the tractor trailer was a billboard. It had been programmed to ignore billboards. The Driverless Cars are built a lot like whack a mole. People make little bit of progress. I will just use more data. We need to get a lot better. The cars need human intervention about every 12,000 miles last i checked. Humans only have a fatality every 3400 miles on average. If you want to get to human level give a lot more work to do. And its just not clear grinding out the same techniques and if they are going to get us there. My question is about Machine Learning. If you are just doing pattern recognition you just dont learn anything. Do you think we are making progress on having Machine Learning as a program to be able to tell us how they are making decisions. There is a lot of interest in that. There is a tension between techniques that are relatively efficient in techniques that produce interpretable results. Does this look like another asteroid that ive seen before. It is as far from interpretable as you can possibly imagine. People are making little incremental progress to make that a little bit better. You get better results. Ive not seen any great solution to it so far. Right now there is a ratio with how good it works and how little we understand. Someone is gonna die and had to tell the parent of a child. It happens to be the perimeter number three. It will be on satisfying and meaningless. On the application to healthcare diagnostics. In the racial bias that people are concerned about. In the fact that we cant afford to have any misdiagnosis. I guess theres three different questions. The first is and i might have to look at it here. It relates to the last. Its also the case that human doctors arent completely reliable. They are pretty good at pattern recognition. At least in careful laboratory conditions. Nobody really has as far as i know a working realworld system that does radiology in general. In principle deep learning has an advantage over people there. As a disadvantage you cant read the medical charts. Like doctors notes and stuff like that. It is just written in english rather than being a picture of a chart. They cant read that stuff at all. A really good radiologist is kind of like a captive. I realize that there is missing symmetry there. They tried to put together the pieces. The story about whats going on. Current techniques just dont do that. As i a second rule out next week. The first cases are probably going to be radiology that you can do on a cell phone radio have a radiologist available. The systems might not be perfect but you can try to reduce the false alarms to some degree and decent results where you couldnt get any results at all. We will start to see that. They have not been digital and starting to do it. Radiologist had been digital for a while. Youre trying to put together some complex diagnosis of a weird rare disease. Symptoms are going to be able to do that for a long time. I made an attempt with that at watson. It would miss heart disease, when it was obvious to a first year medical student. And there it goes back to the difference between having a lot of data and having understanding. You dont understand the underlying biology. Then you cant really do that much. We just had the tools yet to do really highquality medical diagnosis. Im working now as out as a data analyst in scientist part of what im working on is scoping what are small and discrete task. It would be helpful for and other tasks like forecasting like you were saying not separable right now with current methods. See mac i am always interested in ways to explain those things. Founded versus unfounded. I think the fundamental difference is the closed world. Some problems are open ended they could involve arbitrary knowledge. We only drive on the roads if youre talking about ordinary circumstance. But its openended. The Police Officer could say this bridge is out. There is some new possibilities in not way. And what you end up finding is the Driverless Cars work while in the stuff that is closed. They have a comfort zone. Do they have that move of evolution. That we are using the details the wrong way. I dont see it that way. The billion years of evolution they altered genome. They built a rough draft of the brain. It is clear that it is not a blank slate. You can do it a deprivation experiment where animals dont have any exposures to the environment and they understand various things and so forth. What evolution has done is shaped a rough draft of your brain. You can think about ducklings looking for something to imprint on. The minute that they are born. Evolution has given us a really good toolkit for assimilating the data that you get. More and more data and time can i get the same thing. And maybe but were not very good at replicating that. We could just replicate that. There is another approach to engineering which is called where you try to look to nature and try to solve the problems. They try to take clues. That is fundamentally what i would suggest by another name. We should look at how biology in the form of human brains or other animals. That it manages to solve problems not because we want to build literal identity and we dont need to build more people i had two small people, they want to compute really fast with the best of what people do. To be able to read and so forth. So we can do things like problems that no human beings can solve right now. The paper is published every day. No dr. Could read them all. It is impossible for humans. They cant do it either. If we built machines that could read. We could revolutionize medicine. I think to do that we will build in basic things like time and space and so forth so that the machines can make sense of what they read. See mac how you think about fixing the problem. And building the new models and what form they will take. Are they going to be that same structure or is it something completely different. The first thing i would say is we dont have the answers but we try to pinpoint the problems. We try to identify in different domains in space like time. Where the current systems work and where they dont. The second thing i will say that the most fundamental thing is we need ways of representing knowledge in our learning systems. There is a history of things called expert systems. If this thing is true then do this other thing. Its likely that such and such is happening and this is happening. The knowledge looks a little bit like that. And then we have deep learning which is very good at representing correlations between individual pixels and labels and things like that but very poor at representing that kind of knowledge. We need a synthesis of that. So learning techniques that allow you to be responsive to data we have systems that can do a little tiny bit of this. But we dont really had a system of any anyway where we have something just like wallets occupies. We just dont have a way of even telling the machine right now. We try to close some of this knowledge, then are we playing a different game. We need to import something about space. What i would say is that there is a lot of knowledge that needs to be included it doesnt all had to be hand included. They can learn some of it for themselves. But there is some court domains and ive borrowed this from them. If you have a framework for representing time then you can represent things that happen in time. And just seen a lot of correlations between pixels its not really going to give you that. The number of words in the english language. It Something Like 50,000. And maybe there is a hundred pieces of common sense that goes with each of those words. You talking about talk about millions of pieces of knowledge. It will be a lot of work to encode them all. It is not an unbounded task. Its one that they dont want to do right now. Its so much fun to play around with this learning. And get good approximation. Nobody has the appetite to do it right now. Theres just no there is just no way to get there otherwise. There is a long tradition of nativist saying that the way you get into the game is just something that allows you to bootstrap the rest. I dont think it is a walkable problem i workable problem i think is a scoping problem we need to pick the right court domains so we can boost the rest. If you look at the development list. Babies have some basic knowledge for each of these. We could start with the kinds of things that they had identified. And work on those. See where they are. If not, thank you very much. I think we have a book signing somewhere thank you all for coming out. You are watching the tv on cspan two. With top nonfiction books and authors every weekend. Television for serious readers. The cities to her is on the road explored the american story with support from our cable partners. This weekend we traveled to rapid city