Skip to main content

tv   Jennifer Neville Discusses Artificial Intelligence  CSPAN  December 23, 2016 9:07am-10:13am EST

9:07 am
obviously a story they are not going to want to push. they want to make these layoffs and keep it moving but when i talked to him he was like, how did you find out about this so quickly, i just told the employees area and i gave credit to my sources for that one but those are the type of stories that i think to go back to what you were saying about cheerleading, we need to cover the good stories, we need to cover the companies that are hot and have a lot of promise but we also need to chronicle the ones that are struggling or that fail to really make sure we're getting both sides. >> we will make one more question and make it quick . >> thank you so much for attending. my question is on, sometimes when you look at tech news it's like chasing everything that school. as a bystander, from laugh yet indiana who has no involvement in terms of what uber is doing, do you think text have a responsibility when there is new technology
9:08 am
coming out to refusing it, that is and extent of people in somewhere in indiana. for example, instagram or snap chat. they come up with new features. it's really like a new? >> and then if you are talking to a 35-year-old male he's like, what should i care about? that's a question that reporters refuse it because sometimes when i read tech news it's like everything that is new but less in terms of how do i make sense of how i say it? >> that's a good question, it gets down towhat you are saying about now you have all the different audiences . >> people say the snap chat news comes out and we have this population that you use is snap chat can't wait for these features to come out and they are up on the latest ones in before seeing it, they probably understand how they are going to use it.
9:09 am
we really need a news tory says this is exactly how you your physical app is going to change and how you're going to need totouch it instead to do new things and were going to need to tell someone, there's a thing called snap chat . >> but all three are very valid stories that need to happen because the people that are learning that even people that don't understand how snap chat works are missing out on the fact that there is a segment of its population that is mutating with disappearing messages and the impact that's going to have down lot the line is going to be important so it's important to figure out how to break down the story for all these. i guess the answer is yes, that's super important and we are trying to do it all the time. >> i'm afraid that's all the time we have, would you join me in thanking our panelists as well as our graphic
9:10 am
artists . thank you all area. [applause] >> now another panel from the purdue university dawn or doom technology conference with a presentation on the history and future of artificial intelligence. >>. [inaudible conversation] well, good morning ladies and gentlemen and welcome. it's my pleasure, my great pleasure to introduce jonathan neville, jonathan is an associate professor and
9:11 am
chair of statistics here at purdue university. professor neville, she's come to us even though her department is enjoying an external review right now so it took her quite a lot to be here and i appreciate that. her research interests lie in the development and analysis of relational learning algorithms and the application of those algorithms to real-world pathways. today she will present a talk entitled ai easy, this is ai hard. if you haven't already done so, silence your electronic devices but don't put them away. as we learned in the last action , 15 years ago, 10 years ago, if you are heads down on yourphone during a presentation that was regarded as a bad thing . without paying attention to the talk, now it's regarded as a proud thing if you are
9:12 am
not on your phone because you're obviously tweeting about it all the time. snap chatting or whatnot. but please do so on your devices, we hope to see the hashtag dawn or doom and any equivalent postings of facebook, instagram, snap chat or whatever else, whatever other social means you prefer. please welcome me, join me in welcoming doctor jennifer neville. [applause] >> thanks jerry, i have a mic on so you can hear me okay area can you hear me now? okay. okay. thanks for having me here again at the symposium, this is one of my favorite places to come and talk because i can think at a higher level about talking about what i do to more general audiences in my typical research conferences so as jerry said, i have a joint appointment to computer science and statistics, my area of
9:13 am
research is human the learning and pacifically i focus on very complex relational demands where there are unique to taking counts between many entities in order to make efficient predictions, accurate predictions but today i'm not going to talk about my own research, i'm going to really try to give you a sense of what's going on in the ai community right now, there's been a lot of interesting through's over the last few years, there's a lot of excitement about ai and what i'm going to try to do today is compare and contrast which happened earlier this year and talk to you about what part and what is easy given those two occurrences. so really, i want to contrast these two reasons that happened earlier. the first major breakthrough that happened was, you can see here marks the team was the event where google's
9:14 am
computer, game play system that plays go one over a professional go player and it was a major breakthrough for the community, for something that at least when i started was something that i thought about doing, i taught myself to play go and i realized i couldn't even learn to play go let alone program a computer to play it . so it's, the success happened much earlier than people anticipated and so in the news, this was touted as a major breakthrough and really something that was going to transform what we saw in terms of this ai. however, less than 10 days later, we had another event happened which you may or may not have heard about in the news area here it is, march 24.
9:15 am
microsoft released nai chatbots onto twitter and it was designed to mimic a young teenage girl and the robot was supposed to interact with millennial's and learn how to engage them in conversations on twitter and this experiment went horribly wrong, if you didn't see that in the news. within 24 hours, this pleasant teenage chatbot turned into a racist, abusive, sexist entity and microsoft removed it from the internet within 24 hours of releasing it and apologized profusely about the event . and the important thing here is that they didn't anticipate this was going to happen. some of the news that explained what was going on blamed us as humans for being horrible people on twitter and turning this cute little
9:16 am
chatbot into a horrible mess. what you might not realize is that microsoft didn't do this unknowingly . they had a similar type of chatbot that was being used for two years previously, very successfully in china, interacting with as many as 40 million people so what was really the difference between when they rolled it out in china versus when they rolled something out on the general internet on twitter, that's what i will talk about. but the important thing here that i want to contrast from a technological standpoint is that out of the weekly in the machine learning space there's a lot of shared technology between these two systems so you might think that a computer program that going to play go would be very different than a robot, a chatbot that is going to interact conversationally
9:17 am
with people but many of the methods underlying these are the same and so why was one system a major success and the other a major catastrophe? to understand this, let's go back to the beginning of ai. to learn a little bit about the history.ai has been around for about 60 years now. it was tarted as a field in 1956 when john mccarthy, mark kaminsky and claude shannon propose to have a two-month conference at dartmouth to really explore and define what this field should be under the conjectures that all aspects of human intelligence should be included in a computer system including many aspects of learning so when they jumpstarted the field with this conference, there was a lot of excitement, a lot of effort in many different directions in the field, i will focus mainly on machine learning the cause that is my area, there was also a lot of
9:18 am
other areas included in ai that i won't touch on today. the early work in the field was really inspired a lot by considering how we as humans might be learning in our biological system. one of the first methods, the electrons between fibrous and glass in 1915 and this was inspired by the neural conductivity in the brain. this was a mathematical model trying to mimic what they thought happened with neurons which were coming into the neuron and it was accumulating up to a certain point and once it reached the threshold it would fire a signal and so they tried to do this mathematically. another early program was samuels program that he invented in 1959, it started with game playing, one of the first game playing systems where it would formulate this
9:19 am
problem of how to learn to play games where you have to look ahead as to what the outcome of the game might be in order to decide a strategy of what to play and the methods that he used to develop this system really were the precursors to the subfield of machine learning right now called reinforcement learning and in that case there's not an immediate signal that tells you whether you are right or wrong in terms of what decision you made but you have to wait for some amount of time before you get some feedback as to whether you are headed in the right direction or not and that is the case. and finally, the third thing i want to point out here is that at the same time there was another area of ai that was focused on developing dialogue systems and so they were precursors to what now are being used as chatbots and in these systems there was input from users and the ai system would have to
9:20 am
figure out how to make a response to them in one of the first examples of this as a successful system was the lifo system which was in 1964 and this was really don't as a parody initially of a ruggeri and psychotherapist and what would happen, what the system would do is it would interpret some information about what the user had put in from pattern matching and processing the language that they had inputted and then take responses based on certain inputs that they had in the system and you are linked to a parody of psychotherapist, that meant a lot of the answers were things how do you feel about that? and tell me more and things like that so at first when wise and buck created this system he created it to show people how difficult their problem of interacting with humans would be and he was
9:21 am
surprised at how many people actually worth fooled into thinking this was a human and so this jumpstarted work in that. so let's see. so how did these basic systems work? let's consider the scenario, the way that we frame these as learning systems reflects how you might think if you self reflected about how you your self learn how to play a game. this is what you put into your computer how the rhythms as well so if you are going to teach somebody how to play checkers, you would tell them about the board. you would tell them about the pieces and what kind of moves they can make, what are the rules of the game, how do you win the game and so on and what you would do is you might start off by watching other people play the game to
9:22 am
figure out choices of move, strategies and particular scenarios where you might play yourself and lose a lot at the beginning but eventually you would start to understand which moves worth these choices in which moves were bad choices so you can improve your strategy over time so that's what we do when we put that into an algorithm and the way that you, i messed up my order here. let me tell you how the algorithm would do that, let me tell you about the history of the successes we've had. this is the area where a lot of the general public is very excited to see about what is being achieved by these computer programs and in 1994, the checkers program that was developed at the university of alberta was the first major success at a difficult game where the software was declared a world champion. in 1995, the backgammon program developed in ibm also
9:23 am
got to a grand master level. the most surprising thing that happened with td gammon is they found the moves that the computer program would make were different than what the human players would make and particularly they paid opening positions at the backgammon board differently than humans would and not only was not surprising but eventually was incorporated into human play as well so that humans learned from the computer program about better strategies that they hadn't thought of before. some of you may be old enough to remember in 1997, deep blue, the program from ibm, also achieved this level with chess. much of the success from the deep blue system was really from architectural computational improvements that allowed them to improve searches faster than we had
9:24 am
previously so that success was very different than the ability to learn backgammon but recently we've seen with out the go, those two types of successes have been combined together and so the system, the system developed by deep mines and google has just now, there's a session player in go and again had a surprising move that people initially thought wasa mistake . they moved 3 to 7 in game two if you want to look it up but the professional go player said no human would ever have made that move but at the time when they made the move, it was very surprising, when the system made the move, it turned out to have turned the tide of the game and resulted in a loss for the human player so again, they were surprised by what the program ended up doing area they framed this problem of learning how to play a game in a search tree like this and so what this search tree,
9:25 am
this game tree shows is all possible board configurations and so this shows you part of a search, a game tree for tic-tac-toe and up at the top you can see the board is completely empty. everybody knows the rulesof tic-tac-toe, you have exes and owes . you choose to move somewhere on the board, these three positions represent all possible conditions of where ex should go because this is the middle and you could be here or here and you have to face the outcome and the corner position. so ex has to choose where to go and then after that at the next level , oh needs to make a choice based on the available available spaces left on the board so following through this game tree you have a sequence of choices by the two players who form a branchin the tree , all the way to the bottom of the tree where you would end up winning or losing so this is how we frame this as a computational problem.
9:26 am
in the problem in terms of deciding how to, what strategies to play comes from the fact that when you make this move right here at the beginning of the game, you have no idea. there's no signal as to whether that's a good move or not until you get to the end of the game and figure out whether you've won or lost soul of the algorithm needs to do is make choices all the way through the game somehow and get signals as to whether they win, lose or tie and propagate those signals back up in order to change their evaluation or moves in order to develop a strategy based on their experience so the complexity of this algorithmically comes from the size of the game tree and how easy it is to encode it all and get to the bottom and propagate those signals back and forth so the way we look at the complexity is to look at the size of the game tree,
9:27 am
how big are the was game trees? tic-tac-toe is a relatively simple game that you probably don't play anymore unless you have children because you know that you can't win and has to fill 100,000 nodes, on the order of 100,000 notes in the tree. so that's still fairly large. . you can see with go here that this is an order of magnitude larger than checkers when we had 660 here and so to give you a sense of how big these numbers are, 10 to the 28 is about the number of atoms in your body. and then to the 80 is the number of atoms needed in the entire universe so these numbers are astronomically big and if you think about exhaustively searching the game space, there's no hope so this gives you a sense of to the envelope calculation
9:28 am
of these trees. let's consider the situation where you could evaluate 1 million board positions every second . which is a lot. if you could do that, maybe our current state of cpus would be able to evaluate and you would be able to search over the entire tic-tac-toe tree in a 10th of the second area but when you move on to checkers, it would still take you 10 to 18 years to still research that tree. so it's entirely impractical to think about searching these entire trees so what we need to do to be able to solve these problems is to resort to learning, we need to be able to learn whether a particular move in a particular board configuration is a good move without having seen the entire tree and the way we're going to do that is, how does the author go system work? for speakers and any talks on this, i'll try to characterize it very simply for you. it's a combination of the
9:29 am
ideas that have evolved from these two historical methods. the learning is really a very complicated version of a perceptron so it's many perceptron's combined together in many layers and so learning is much more obligated that was with a single node in the middle but that's where the learning has grown out of. what alpha go does is it combines learning and reinforcement learning so the ideas that come from the checkers playing game where we have to wait and understand what the reward is going to be after a game is won or lost are also combined into their system and the way they do this is by learning multiple models and combining them together. in particular they have to neural networks that they learn with. here is the result from their paper. so one neural network, you can see there's a network here and one neural network is designed to predict the
9:30 am
next move given the current state of the board and so that's a very immediate feedback because you're not trying to predict whether they are going to win or lose, they are going to predict the next move and they changed their model on 30 million positions that they acquired from traces of certain games so human decisions about what they make and they learned the second model is going to predict the likelihood that they are going to win a game given the current state of the board so this is a much longer range prediction of how good is the situation at that point. and then they update these two models using reinforcement learning and combine the two together in a complicated monte carlo search procedure which, let me tell you that it's not a simple approach, it's a complicated system to solve the problem but it's using many of the simple foundational methods that we built up in that area of ai.
9:31 am
so when you think about these gameplay systems, the two dimensions of complexity that decide whether the problem is easy or hard depend on here the size of the search base as we go from small trackable things with tic-tac-toe to things that are intractable like the size of go, they get harder and also the amount of the way you have until you get feedback that we can use in the algorithm to do learning, that's also another dimension of complexity. i think i may have forgotten to say that one of the factors in these games is the impacting the amount of delay you have depends on how long the game is. so if your game only requires 10 or 15 moves, you're going to get feedback after 10 or 15 moves but if it takes you on average 115 moves, you
9:32 am
have to go much further down in the game tree and your reward is much more delayed area so now let's contrast this with dialogue systems and what went on with that today.so in figuring out how to learn, how to interact with humans in a dialogue system, the basic learning procedure is very much the same as what i described for the game but in this space you need to learn about language and behaviors and interactions with other humans and so what are the experiences you're going to have? they're not going to be games with a fixed length of time and a very clear system of where you won or lost in the end. if you were to have conversations with people, you're going to have subtle feedback about whether they like you and wants to engage with you or not and so something that maybe even humans are not all that good at, we are not uniformly good at because you have factors
9:33 am
like social intelligence, people who are able to interpret those signals from other people more effectively are valued as having more emotionalintelligence . from those singles, once you interpret them, you learn from the feedback new strategies of how to engage with people and how to extend and make that engagement more effective. so in terms of the history of dialogue systems, we've also had a number of successes along the way and the ai community, so if the history started with you liza, with psychotherapist models that i talked about before and the 1995, we had the alice system with is able to convincingly mimic conversational patterns through natural language processing and pattern masking area this is the administration for the
9:34 am
breakdown movie her that came out in 2013 and was aired at the dawn or doom conference in 2015 and then you might think this next one should have gone in the game playing rick, in 2011 ibm won the top players at jefferson but much of the technology that was using to play jeopardy are the same types of technologies they need for dialogue systems so they use the input from the clue and figure out how to answer the question using a lot of natural language processing, information retrieval so although there wasn't the same sort of interaction, it was really a big success for the natural language processing community. that is something you may have heard about in 2014, there was a system called the guzman which nominally passed the turing test. it was the system developed by russians in a competition
9:35 am
on trying to convince judges that it was human, it was able to convince 33 percent of the judges that it was human so they claimed that past the turing test. there's some people that disagree with that claim because this chatbot was developed as a ukrainian boy and they used a lot of computer tricks to hide the limitations of this process system so because it was a young boy who was from another country, it was very forgiving to not quite understand the language or questions people were asking and also to the more curious it would be able to ask the questions and deflect the conversation when they didn't know what was going on. so this system, while people think maybe it hasn't quite
9:36 am
passed the turing test, it did use techniques like humor and deflection in conversations in an interesting way that was able to convince the judges that it was human so really, that should be considered as kind of a human area so in 2016, i guess you could disagree whether this is a success or not but when k was released by microsoft, really it was successful in that maybe it was a victim of its own success. so take acquired 2000 followers and he did hundred thousand times in the first few hours that tay was released. there was resistance from tay from many groups on the internet but it was vulnerable to this coordinated troll attack, hundreds of users identified a way to interact with tay in order to pack the language structures the chatbot was using in a way that was not desirable.
9:37 am
so how do you, how do these systems work? i can't describe them in a simple representation for you. it's a much more complicated architecture that has a lot of components, the main components are that a user has a message, either detected or verbally said, there's some sort of technician and natural language processing that tries to understand the context and the intent and the task that the user is trying to achieve and this goes into some sort of dialog management system that has all the guts of the natural language processing and response generation and there's two basic ways that methods have tried to figure out how to generate responses , the older method is a rule-based matching sort of system that works from the
9:38 am
input and tries to match them to certain kinds of templates or the types of response that they should generate and goes to a database of responses and tries to generate those and sends it back to the user. the more current methods are single based methods based on what happens in the information retrieval system that are used at google and being to identify which documents to return to their queries area if the same basic technology and it takes the input as information needs and uses language models to match it to some database of responses and figure out from the relevant responses how to decide what to return to the user. so how do we develop these dialogues? there's two major dimensions of complexity contract with the game playing systems are that although language has
9:39 am
structure and rules like games, it's not as clear-cut as the actions and the outcomes in games . for one reason, it's continually of evolving. words are being added to our language all the time. rules are broken, the way people speak doesn't necessarily follow those rules. the intent slang or colloquial terms or uses of sarcasm and irony and humor in ways that we use words in ways they were not intended. this means what the algorithms have to consider and how they learn how to behave with user input is effectively unbounded. so there is an infinite number of possible scenarios the users can put in and you have to develop a scenario, a system that's going to be robust and able to adapt to scenarios they haven't seen before. the second type of issue is that the feedback about whether you are doing something correct is much more vague and unclear and
9:40 am
you wouldn't know that from your own social interactions and how you try to interpret who's happy with you, who's not, who likes you, who doesn't. the feedback is often unclear and there's a much longer delay before you get feedback. so you can see this is an issue in terms of what we put as suggestive functions into our systems. right now they would be trying to optimize something like the length of conversations that you will have with a chatbot in order to use that as a proxy to something like chat back or user engagement.so what does tay do as a system? just like without a go, it's uniquely training. the microsoft people have not written papers that publish the nature describing the entire system so the information i have here is gleaned from many thoughts that i've seen from people at
9:41 am
microsoft research about component technologies that they are developing as well as some of microsoft's information about tay. it's pretty clear they are using complex natural language processing that's based on deep learning techniques as well so what they've really done is move beyond the simple masking and retrieval based systems to use complicated deep learning models to predict what kind of responses are likely to be a good match to particular user input and again, they are using massive amounts of training data, just like the alpha go system so they are using millions of examples of interactions with users that get they get from twitter or other online places where you can see these overtime. but the big difference is that in the chatbot system, is what i would call an open system versus a closed system. there is no clear balance to the type of interactions they might have, the type of
9:42 am
behaviors they might see from people and this makes it vulnerable to attack. he never precipitated this kind ofcoordinated trolling attack , what happened and they were developing the system. okay, so the dimensions of complexity here with respect to dialog systems are smaller to the things we talked about with games and in this case now we have a third size and space which is an open system or an unbounded search space and now we have feedback that is not just delayed but it's very unclear and so we see as we go on in the world in the development of ai, we tackle harder and harder problems. so when we compare the two, the gameplay systems and the chatbot conceive that when a problem can be formulated for
9:43 am
immediate feedback in a trackable base, we are able to follow it pretty easily. so those are the situations where it's considered easy and we have our major successes in the community but as we go further up this upper right-hand corner, there are problems that are still hard and work can happen. if we go back to discussing the difference between alpha go and tay, when i said they were built on the same underlying technology , they both use learning methods that have shown a very significant and impressive improvement in some areas of mission learning as well as they have massive amounts of data to frame them on but what one major difference between what our ago did and what tay could do is that tay can't learn from playing against herself which is what
9:44 am
our ago did. so in and out of the go system because of the rules of the game are very clear and whether you win or lose is very clear, they could take different versions of their system, play them against each other to generate even more experience of the different types of sequences and moves and outcomes you would see but that's impossible to do with tay . you could do it, but it wouldn't be successful to make the system robust to the kinds of interactions that you would get in the real world with real users because of the variety of engagements you would see from users so what tay needs to do to get the kind of feedback and experience that alpha go is getting, is that tay should be out in the real world and tracking with scissors and trying to figure out how to behave so that meant when tay
9:45 am
was released onto twitter, she was still learning. what happened was this coordinated troll attack was that users created conversations with tay and use language in a particular way that eventually made tay think this is what i should say and so eventually after enough interactions with these kind of people, tay was turned into a horrible person and what i like to explain to you as a sample to explain to students is that if you can think of, well, although tay was very very successful as an algorithm and has learned a lot of things about interacting with humans, you might think her amounts of knowledge would be equivalent to a seven or eight-year-old child who had had some experience with this batch of people in a limited environment and you don't let the seven or eight-year-old child interacts with the general public on twitter.
9:46 am
so the difference when ice was operational in china is that in china there's a lot more restrictions on what kinds of information is on the internet because of the cultural control so they don't get really the same kind of interactions as you might on twitter. so if you think about this, you might say as i did from an algorithmic perspective, we should have known that this could have happened and the algorithm should have been able to detect that people were specifically changing their behavior and how they talk in order to make this chatbot behave differently because of the atmosphere setting where they know that they can change input in a way to change the output in a terrible way but this very complicated to detect algorithmically because it's very easy to say in hindsight there are hundreds of people that
9:47 am
participated in this according to an attack but when it started it was very subtle and it's very hard to detect and know that one individual interaction is not a valid one, it's an adversarial one so it's difficult to code that to the program to figure out how to let ed identify it, and this is why when we had our young children and we are teaching them how to interact with people, we don't let them go on twitter right away because in kindergarten, the interacting it constrained environment people who are kind and loving to them and maybe there's bullies in the class but eventually they get only small doses of it until they figure out how to interact in those systems. so in terms of what we need now for future research, really what we have to think about is that we have to push in this direction where we are moving to unbounded kinds of situations where we are going to have our systems they and this is something
9:48 am
that humans are able to do it very easily. maybe not easily but some people do that easily but we expect people to be able to adapt and use situations in order to behave with relatively few examples after they have developed over time into adults and is so the way microsoft is trying to encode this information into a chatbot system is they are working with improv actors and so this might be surprising to you that this is something very non-computer science see that they found that what improv actors have to figure out to do is they are thrown into situations all the time and they have to figure out how to keep the act going which is exactly what they want to do with chatbots so this is something they are working with with humans who have the skill to figure out how to
9:49 am
algorithm that ability and maybe if we learned more from how they do their interactions, we will be able to put that into our algorithm easily to deal with this open ended system. the second issue is with the dimension here of feedback and so we really have to figure out how to deal with this subtle inconsistent, maybe long delayed feedback but that's also something that we do fairly well, maybe to people with emotional intelligence do it better than others that's something we can figure out how to do ourselves so it should be something we can figure out how to do in the algorithm so the two areas of computational science or computational humor are the kinds of research directions that are trying to take these into account so computational social sciences are nominally the area that i'm in. when i tried to come up with algorithms that will take into account interactions
9:50 am
between users, what i want to lose our complicated situations where many people are interacting with many other people and we'd like to learn how to predict their behavior in these situations but it's not as fixed environment as a game play where you have continual one-on-one interactions over time, people interact with different sets of people and a move and change groups and things like that. early ideas from social science of interpersonal communication and management, all of these kinds of ideas need to be put into our algorithms and our learning methods. computational humor is another area where we are going to have at least a few speakers later on in this conference including bob this afternoon talking about how we understand what's funny, how do we know what's funny and how might we put that into an algorithm, is another interesting direction to try to deal with the situations that are not clear.
9:51 am
okay. so to wrap this up, we've made great progress in ai over the last six years. they have things now that are starting to be based on reality where we have self driving cars that are able to identify from sensors and figure out how to adjust the car area that we have smart buildings that are able to automatically detect temperatures and airflows based on optimizing energy consumption and we have the beginnings of personalized medicine where instead of deciding one treatment plan for a large population of the general public, we are starting to have methods that are being certain lies to someone's genome or the history of illnesses over time. so this is really an exciting time for this field in
9:52 am
computer science in general, ai and machine learning specifically so you might be tempted to say well, these mentions of complexity i thought about are not too hard, we will eventually go into that upper right-hand corner without much effort but i would caution you hear that as a community, we've always underestimated the difficulty of the problems that we are addressing and so what i will end with here is this story that all ai students know about that you might not know about. in 1966, 10 years after the beginning of ai, seymour pepper who was one of this group of mit faculty that started the field pull his graduate students to solve the computer vision problem as a summer project and why did he think that they could solve it over the summer? they said because unlike many
9:53 am
other problems in ai, computer vision was easy because it was clear what we were trying to solve and it was easy to code that algorithmically. six years later, we are still working on the problem. it was supposed to be a summer project. but maybe we are close to solving computer vision fairly soon based on all the work going on in ai. so i'll close here with this image from the facebook vision system which is able to take images that are uploaded on facebook and they automatically identify the components of the image and know what they are so it can produce a textual representation of what's in the image and that can be used to produce a verbal description for blind people. and so it shows you the capabilities that we have with our systems now in terms
9:54 am
of how much can be identified and these are going to be identified as feet so all the little blobs are surrounding the object and not only can we identify and fragment the image and identify the components, we can also decide what those things are and produce labels for them so that the image can be described to people who can't see it area obviously we can use that information for many other aspects of the facebook system but we benefit to the largest society is clear. so i will stop there, thank you very much. [applause] >> that was a great presentation, now we can take some questions from anybody who would like to ask. also, i need you to come up like i did to the microphone. we will go around here. hi, so you mentioned
9:55 am
emotional intelligence and i have heard some scientists talking specifically this biologist george church who is a futurist, he's mentioned before the idea to teach ai emotionalintelligence , it seems you can teach them to have morality or a sense of emotion, that maybe she wouldn't have fallen into the trap and maybe i was wondering is there a way to preprogrammed morality or is that something we still don't understand that we can't do that? >> that's a very good question. i think that issue of morality also comes into play when you think about the self driving cars and the current requirements on how safe our self driving cars will have to be before they are allowed on the road is a threshold of
9:56 am
safety that we as humans can't even conceive because somehow we think that the system has to be a much safer, we can't possibly release a system out in the world that we think will eventually end up killing someone even though we as systems might go out and get into a car accident and have somebody died. so i think the complexity in that, absolutely there are people that are thinking about how to algorithm the ties that and i think it's likely that we can but we have to think very differently about it then we would with a game playing system. i think with gameplaying systems, everybody agrees on which outcomes are good and which outcomes are bad because you've either won or lost and when you think of these supervised learning systems that we developed to predict credit card fraud or to detect expanding your email, there's still a clear signal as to whether a transaction is fraudulent but when it comes to something as
9:57 am
complex as morality, we don't all agree on what the answer is and so that is something that probably we couldn't and code very easily if we were willing to be satisfied with that kind of encoding so if you somehow decided on how to value human life and what kind of action you would take if you are about to get into a car wreck with say, 10 people in front of you, if you are driving, you would get into a car wreck and possibly killed 10 peopleor there's one person on side of the road but it's your grandmother . what would you decide to do? we would all make different decisions and if we could answer what decision we would make, we probably like to think that we don't want to come up with the answer but we could look at it as a distribution of what humans would find acceptable and then probabilistically, we could put that into the algorithm but we have to think a lot more carefully
9:58 am
about acquiring that feedback and encoding it in order to use it. so i think it will, but i think there's going to be some tough philosophical discussions amongst us as humans as to how do we actually value these because what the algorithms are doing is, they are valuing different kinds of outcomes so once you put a value on things, you can take it off so if you as the general public canvalue human life , then we can start developing systems to optimize it. so in various points in your talk, you have the possibility of probability of things like computational propaganda and computational surveillance, i'm wondering if you could comment on that. >> absolutely, those are two good issues to bring up. so computational propaganda would be the situation where
9:59 am
we would inject box into the world that would try to persuade people to do what we wanted them to, pretending to be you know, just regular people just like the way humans would try to propagate that propaganda area trying to be robust to that or detect that is as difficult as detecting a troll attack because we really have to understand what is valid behavior and what's propaganda and the line between what's valid and propaganda is very difficult to define because i could try to pursue someone to vote for a particular person in the election but i wouldn't necessarily think of that as propaganda but eventually if i try, if i go too far in doing that persuasion, you
10:00 am
might think of it as propaganda. so it's related to what we just talked about, trying to probabilistically identify these scenarios is going to be much more robust than making a quick decision yes, sir no and having the system self monitor or interact with humans to say this is a situation where maybe i think this is going on, let's detect it or not. that is something that happens with fraud detection right now, when some of the systems are detecting automatically whether a particular credit card transition is fraudulent, they might shut down your card if it's clear that it's a fraudulent transaction but in other cases, they will notify you as a user to get your feedback about that. what was the second topic? >>
10:01 am
travel to the bay area a lot, look at machine learning topics and are vegan. the very quickly from just a few aspects of my profile you would be able to narrow down that it's probably me from that circuitry. and so -- history. it's sort of a scary thing, and i guess my answer to that is when that date is out there we can choose to use our powers for good or evil. there are always people that are going to be trying to use them for evil but that doesn't mean we shouldn't be developing a mess, computational methods. if we understand the things are
10:02 am
happening, maybe we can identify this type of thing is going on. i would say the privacy preserving machine learning and data mining community is really focused on trying to have these, developing methods that can be applied for learning in a situation where personal information is aggregated but yet you can guarantee there's no leakage of information about any one individual being in that setting. i think those are the types of methods that should give some solace to people. i also say this to my students. you enjoy being targeted. in your system, when it gets personalized to your own specific behavior, you enjoy the analysis all the time. your e-mail is figured at which people are valid senders and receivers for you so that they can identify what is spam. the surge engines are looking at
10:03 am
your history so that once you type a few letters and already knows what you are searching for. this is making your life easier without maybe you even knowing it. so that is the good from it. if people get to the point where they deny you health insurance because you've searched for something particular online and they have identified you might like to have a pre-existing condition, that would be a very bad scenario. spirit that was a really great talk, thank you. to me, you're describing a kind of fifty-year process of interaction between data and competition/algorithm for the data set and the rules moved from relatively finite area to an increasingly more complex city and richness in terms of the data sets and the competition comes up between a better algorithms. that dynamic colds and then it explodes with search which gives birth to cloud computing but also an incredible level of
10:04 am
unstructured data that you constructively with in social media also does that. without visually take watersheds over the last 10 years. you get into this whole other interesting dynamic which you touched on in terms of the social research which is not simply the activity of speech but speech as it is received and speech in social context contexh is this whole next level of complexity we are going to have to take on algorithms to handle. there's almost tw two fixes to this. one is the elementary one is recognized that it's bad to be a nazi, and so cross that stuff out. the other one is to look at social circumstances. this is coming from the gamer guys, suspected or even more create countervailing messages to that that will correct tay. part of the province once you
10:05 am
start to get a little nazi, all the nice people stopped talking to her so became a self reinforcing rule to one count would be more nice people come here talk to this poor lady. i guess the question is as we move into this kind of ai art we also reprogramming ourselves? >> yeah, that's a very good observation. i think that the observation you made about the wish of more nice people talk to tay i think affects not only the algorithms by humans as well. you hear about people leaving twitter after they have horrible interactions with people. we can't fix that problem for humans but maybe just maybe if we understood how to fix it for the algorithm, we would be able to improve things for ourselves as well. and so that's some of the social research, social science research i was alluding to, you
10:06 am
know, focus focus on things like encouragement, positivity, creativity and how to foster those kinds of feelings and behaviors and people. if we understood better how that works in humans we might be then have to put into algorithms here but at the same time if we come up with solutions algorithmically that would fix tay divergence into this horrible state that might also inform the social science to think about the structures that they been looking at for computationally. something like 10 positive interactions a day might make everybody feel happier, i don't know. it's an important thing to think about, and i guess my point of interacting more with people in social science and philosophy and humanity is that as engineers we can to not think about these things.
10:07 am
we like to put it into a mathematical equation. i guess i did not put any math into this talk, but usually every talk i have has a place. i just want to distill everything in a very precise mathematical equation but all of these issues are maybe hard to quantify in a mathematical equation. and so it's really from discussion with these people with more social emotional intelligence that we as engineers have, that would be able to help us put that into our algorithms and our systems. i think we will have, if we have the right kind of back-and-forth between those two communities, i think we will make great progress in the systems while at the same time learning a lot about ourselves. >> professor, i was wondering part of your opinion on tay,
10:08 am
what if she was trained, isolated from isolating it to just conversations like kind people talk to it and then learns from them? then you can track progress like first with people that usually talk to as a kid and then you go up, teenager and finally adult, what would happen if we try to train tay as a perfect human rather than just letting it out open? >> well, so, i guess i'm not sure quite sure how to answer that. absolutely one of the issues with tay is that tay is doing continuous learning. and so there are two things we could have done. we could've had more training to reinforce tay use behavior and kind, trustworthy and vibrant before running out to the general public, and didn't stop
10:09 am
her learning. we could say we've done only in these very fixed environment and that we will stop so will be robust to these attacks. moving forward. but in a kind of system would be brutal because it would not be able to adapt to new situations that it hasn't seen before. we really need to be able to adaptively learn from the new situations that we see but what of the issues that could be adjusted and probably is being adjusted in new developments of the systems is that the algorithms on based on the data that they have to. and so you could imagine a scenario where we wait, the training data coming to us, based on trustworthiness. so that for particular kinds of interactions, words like nazi, genocide, maybe we say that's very untrustworthy so we should wait those very low. we won't update the aspects of
10:10 am
the model very much when we see this kind of attractions, but if this particular people that we trust like our parents or siblings or our relatives and we see no interactions with them, we would up wait those and say not just to the two kinds of interactions because those are probably value interactions. so work on figuring out, there is work in the area with reviews and online systems and trying to automatically identify how trustworthy something is. we could potentially use those kind of things to make the system a adaptable and more robustly by basically understanding what kind of data is coming up. so yes. >> good morning, professor. i had a question. the different artificial intelligence and machine is very great. you never know exactly where the line is drawn. so, for example, people here are like discussing about tay. we have some ideas like constraining her to specific
10:11 am
talks, people, situations. but if we do that can't it will be called artificial intelligence? because if we're constraining it, it's just an intelligent machine i'm not exactly making room for itself spin the definition of what is intelligence, what is learning is very fuzzy. so these early systems like eliza was not doing anything that you would think of today as being intelligent, but at the time i talked about it as an ai system. there are things we do in machine learning how to even have that discussion with my grad students last week, that voice of right out and equation and optimize it, handcrafted that equation to reflect a particular scenario, are we learning? i don't know. we have a system that does something based on optimizing that equation but maybe it hasn't learned anything about the environment. and so it everything that we
10:12 am
develop as computer scientists, we can go back and forth between that you really when we focus more on the engineering and making a system that works with often putting things that are hand coded or manually specified in order to just get the kind of outcome will looking for and then we tweak it to behave well in the snares we wanted to. as we try to extract that theoretically we tried to move to a more general concept. i would say those two kinds of things are always there mutually hand-in-hand in the scenarios and when we're making better progress of understanding of the higher level we are pushing to these more theoretical abstractions but we also make a huge amount of progress by just making specific decisions. i would think they are both useful systems but maybe a philosophical issue to decide whether you think of them as actual intelligence or

11 Views

info Stream Only

Uploaded by TV Archive on