Kai-Fu Lee | Full Address & Q&A | Oxford Union

[Music] [Music] thank you very much it is such an honor to be standing here where so many great speakers have spoken and today I'm here to share with you my personal journey in artificial intelligence make some predictions about the future and point out some issues that we as mankind might face and give you a few thoughts on how we think how I think we can perhaps overcome that ok so my personal journey started with artificial intelligence 35 years ago this yellow paper not nearly as HTS that this university was my application to Carnegie Mellon University when I decided to dedicate my life to the pursuit of artificial intelligence because to me at the time this was a study and an attempt to understand how the human brain works and I could think of nothing grander than understanding how the human brain and I thought that would be the final step in the human journey to understand itself so that was my essay and then later in the 80s I had the fortune to work on the first computer game that defeated a the world champion not go or chess but a much smaller game called Othello and then I had the fortune to work on speech recognition delivering the first speech recognition at a speaker independent level at my as my PhD thesis later I went to work at Apple Computer and the team that I found that there continued and is currently managing this product Siri that many of you use so I felt very fortunate to have been a part of the AI revolution but the air revolution wasn't always very smooth a high ran into a lot of problems at first people thought eh I was figuring out how people thought that if then else statements in our head when we made decisions that turned out not to work then people thought AI was all about statistics collecting data but then we found that we could not collect or save enough data to make proper predictions and when I did my PhD thesis we were we stored a whole hundred megabytes of data and and that costs a fortune that costs almost a hundred thousand dollars for the computer science department to pay for that and now of course we can buy a hundred megabytes of data with less than one cents so a lot of the second wave also ran into problems but recently we've hit the third wave so now a lot of questionable are questioning whether the third wave is real or not because the last two times us AI scientists said we're close to solving the AI problem we proved ourselves to be wrong is this time the third time going to be different and my answer is our resoundingly yes this time will be different this time the algorithm is based on a set of algorithms but fundamentally the most core one is called deep learning and deep learning basically works similar to statistical learning if you take one single domain a huge amount of data and a number of outcome labels and the system trains itself to make stronger predictions for the future so it is a mathematical system which is learned by looking at examples of data so when you apply it to the domain of go you show it many people playing each other or in fact alpha goes zero which played itself then it learned from the patterns of playing and then the outcome of winning or losing and then abstracted a set of mathematical functions that may help that make a move that maximizes its chance of winning so whenever you have a problem where the outcome can be stated as good or bad right or wrong or classified as a or B or synthesized as something realistic or not then it's solvable by deep learning so when you apply it to let's say the problem underwriting loans you give it examples of people who with their characteristics financially and otherwise on whether or not they've returned the money they borrow then the system will learn to differentiate those who returned the money from those who don't if you train it on stocks that go up versus stocks that go down with all sorts of data including social network quarterly reports it will learn which stocks you can buy that will make you money versus lose your money and for those of you who are Chinese if you use a product called Mei - have you heard it's sometime in the in the Western user is sometimes called them make up + or cosmetic + it is a picture that beautifies you it's an Instagram that works on selfies and have you wondered how you've become more beautiful over the years it is because five years ago it didn't have enough training data and where does it get training data every time you take a selfie and beautify yourself what might you do you might share it on social network you might save it or you might delete it then every example of sharing and saving is viewed as a positive exemplar we want to make more people look like those photos that got saved and then we want to make people less like those photos that got deleted so you didn't know it but you were contributing training data all the time - May - and the saying goes with Amazon Google every time you buy a product Amazon learns more about people like you what you might buy every time you talk to your echo it also learns from you about you and about people like you so that is an accumulation of data that helps that deep learning gain experience from training now this this experience from training is not infinite it is limited in five ways it's not like we can be infinitely smart and throw data and it becomes omniscient it works under five conditions first you have a huge quantities of training data secondly you have to label that data with the desired outcome third you have to apply it to a single domain this is actually very limiting so a alphago cannot just go to become your Amazon buyer and your Amazon echo cannot just be turned into your ma - beautifier every product is one single domain so it's not a one size fit all there is no common sense there's no cross domain there's no true understanding but if it's one domain at a time number four you need a lot of computing resources and number five you need some AI experts to tune the system so there you have it that's basically the capabilities and limitations of AI some people separate weak AI from strong AI weak AI is as I described single domain optimized on data better than people strong AI is more like artificial general intelligence or AGI in the sense of what we see in science fiction where the computer does everything we do and can go cross-domain for common-sense understanding and the litical planning and some people even think self-awareness and emotion so when it exhibits all that it is a pretty scary future we have sometimes been brainwashed by science fiction that is an eventual outcome but in fact today it is more like fantasy it's not that it can never be achieved it's just that a number of engineering breakthroughs needed to achieve it appears to be decades if not centuries away so I think we're going to focus today on the weak AI or deep learning like algorithms which are already very very powerful so these types are so singularity is something I'm not going to talk about I don't believe in it's not near term but single domain a optimization can already do wonderful things create a lot of value and also replace a lot of jobs and tasks here are some examples you can see in customer service and telemarketing and accounting and reporters these are many of these jobs can and will be done better by machines that optimize single domain at a time let me tell you a little bit about the phases of our artificial intelligence and how it's going to hit this world I think there are four waves that will hit simultaneously some earlier some later the first wave is called Internet AI that is the internet giants who have more data than anybody else and they use that data to know everything about us then we're labeling data for free for them every day The Economist covered this week has the 3G American giant's Google Facebook and Amazon on the powers that they have because they have more data than anyone else they understand more about you then your government probably than your spouse and arguably more than yourselves and don't doubt that the amount of data is huge and these internet companies will continue to come up that is the first type of AI because Internet has the most data and they are automatically labeled through usage of the apps the number two type of applications are business oriented applications this is when a bank and hospital and insurance company accumulates data about you and applies that data to its business process to make better business decisions should it give you a long or not is this credit-card usage fraudulent potentially how should it recommend your asset allocation and whether you're the type of customer to recommend this new fund or not so those types of financial determination whether to give you an insurance whether to whether this insurance claim is fraudulent or legitimate whether a hospital scan shows signs of a cancer or cardio problems those kinds of recognition tasks within the domains of business using data that's been accumulated is the second type of AI and that's surely going to generate a lot of value and replace a lot of jobs the third type of AI is doing data getting data that didn't exist before contrasted with the first two two waves of AI the third wave is digitizing the physical world collecting data where there wasn't data before and creating usage scenarios that didn't exist before think about Amazon echo collecting speech from our living rooms we didn't think we were going to let them do that but we are and that data can be used for interesting applications imagine look at airports that have lots of cameras those cameras are and are recognizing faces and the face recognition can be used to prevent terrorists from entering the plane it can also be used to approve your spending so those types of collection it can also be used by retail stores to better understand you you know online every click and behavior is captured for better understanding to target you now offline with cameras and sensors you can also be captured and tracked and that can be applied to real uses that will help the companies make money that has applications in retail clinics hospitals schools everywhere it turns offline into online and people are tracked every minute of every day that certainly brings up lots of privacy concerns but it also gives people the opportunity to trade their privacy for convenience so that is a broad herring process that will happen at different paces in different countries probably faster in China then let's say in here but really creating new usage scenarios never before possible that our one-on-one targeting because the systems knows who you are it knows what you did and knows what you want and therefore it can make money for the merchants and it can give you a better product or service now if you think that's scary the fourth level is even scarier that's really science fiction while I continue to assert that machines will have no feelings and no self-awareness they will move movement is not impossible it's a little bit difficult it's harder than software but machines that move will go from household robots that can wash your dishes to commercial robots that can be a security guard to industrial robots that will make the next iPhone with no human participation to autonomous vehicles to drones and these will provide incredible convenience free us from routine jobs and also take away so many jobs that are manual labor related so these four waves each will create an economic value in the order of trillions and add to our GDP by 2030 on the order of 10 percent per wave and it will remove jobs also at similar rates so these are the positive and negative impacts of AI in terms of transformative wealth aggregation don't take my word for it let's go for the most conservative people possible PwC they tell us that hey I will create about seven trillion dollars of value for China and about 20 trillion for the world so let's just take that as a minimum point next and then in terms of job job replacement a lot of optimists will tell us that AI will not take away jobs because look at history every time there's a technology invention it creates more jobs and it replaces but do not believe this economists because it is not a simple of repetition in history the invention of AI is at at least as large as three of the largest revolutions technology revolutions ever the creation of a steam engine the invention of electricity and the computer revolution each of those hats has had a different impact on jobs electricity and assembly lines definitely created jobs but it actually made jobs lower pain same with computerization it uplifted people's skill sets but it also lowered the average pain so each one had a different impact and AI is different yet from those two again because it is a server-based technology running on amazing NVIDIA and Google TPU power that can be applied simultaneously to thousands of industries that can instantaneously do a better job than humans do in thousands of professions certainly not every job will instantly be replaced some jobs will have tasks that are replaced not the whole job for example face recognition is very accurate but it won't replace the security security guard completely but surely there will be fewer security guards with them face recognition and ID checking being taken care of long under writing will be completely gone because machines are just better at the quantitative problem secondary stock trading will largely be gone because the job that research analysts are doing recommending stocks are an unnatural task for human beings it is a quantitative task better than by machines some jobs will be human machines Simba the combination a doctor's job may involve the doctor applying its humans his or her human skills and communication and extracting symptoms that the patient didn't volunteer and offering that to a machine-learning that gives recommendations that the human combines with his or her skill set for a higher degree of survival rate and treatment but not most jobs are sheer replacement a few jobs are human machine symbiotic so don't let the economists or the optimist fool you that the AI revolution will just hand behind can be earthly hand waved in terms of job creation so I think the biggest benefit of AI revolution is the wealth creation the biggest challenge is the removal and displacement of jobs and when jobs are replaced it is not simply a matter of how to find someone a new job it is not just the loss of income but the loss of meaning because throughout the Industrial Revolution the capitalists have trained our and brainwashed us to believe that the work ethic of working hard and getting at your reward your reward being having wealth and having respect in the community is the essence of life and that work ethic cannot be the central reason of existence for human beings but most people on this earth believe that and now when you say well you work hard and get reward but if you do routine job sorry you don't get reward because robots would do a better job that a lot of people will not only lose the income they will lose the meaning so there so we're facing a serious crisis in terms of the loss of meaning and that to me not the dystopia is the largest challenge for the for the AI revolution and very much related to that is the rich getting richer the techno elites will have so much power the techno giants will have so much power I understand you're going to debate the weather have the technologies are a good thing or not for the world and I surely stand on the side of not I I think I think powerful companies that have so much data and that can turn data into power into value into money into more servers into more smart people into more data that virtuous cycle somehow must be disrupted otherwise they will have so much power they'll be more powerful than countries countries won't know how to manage them and they will know more about us than we know about ourselves now that's perhaps unavoidable that kind of a diversified level there could be a hundred companies each knows a lot about us about something but if one company or five companies owns all the data in the world that is very dangerous and that also relates back to the why the disparity of the techno elites and what they own in the society in wealth and power versus the poor person whose job was replaced by AI or robot so that is the issue that we face and so how how will I suggest a solution for the job displacement I have not been much help in the first 50 years of my life because I have been the poster child of the work ethic I just described I was the hardest worker that I know I wrote three books each of which describes the very work ethic that I am saying is dangerous for mankind I believe if you work really hard you'd be the best you can be you'll change the world and I wrote books that sold millions of copies that many of you perhaps read when you were in elementary schools I still believe that we want to change the world but the work ethic under which that I did to the expense of my own personal life and my family is something I deeply regrets back when I was incredibly working hard I would measure every little thing by how much impact can I make on the world I would judge should I go to Oxford or Cambridge where can I make a bigger impact I would judge should I write a book on this topic or that poll topic okay so I can sell more copies I would measure should I talk to this entrepreneur that on trip who's got the higher IQ and it's become a very very pragmatic and heartless way of running my life similarly I would put work above family I would wake up every day in the middle of the night so I can answer all my emails answer email during the day within five minutes I answer emails at night within three hours because I wake up automatically at 2:00 and 5:00 to answer all my emails so that my colleagues in the US when I was at Google would feel that I work hard and that I don't slow down the process so that my employees would feel they have to work equally hard and that has been the work ethic that drove me until one day when something terrible happened to me I was discovered discovered to have lymphoma fourth stage lymphoma this is my MRI from about four years ago there were over 20 tumors discovered and the doctors that wasn't sure I had very long to live and when that moment hits instantly my life basically played like a movie in my eyes and I felt every move I've made that I thought was so clever was completely wrong that I had ignored my family my father had passed away without me ever saying to him how much I loved him and what he meant to me my mother my mother no longer remembered me we I could spend time with her but we could never have a meaningful conversation again and that every time I had a vacation time I used some of it to accompany my parents so that I would lessen my guilt not so that I would do my duty as as their son I treated my wife and kids too as family members that would give me a passing grade and then I feel that was enough everything else I gave to work my priorities were always with work I remember the day that my first daughter was born I was struggling on whether to leave the delivery room because I had the presentation to the CEO of Apple I was only lucky that she was born a little bit early enough otherwise some I honestly would not have been present so such was the life of a workaholic and get hacked but getting lymphoma turned out to be the most useful lesson that I received in my life because I got the priorities right there's a book that I read during my illness written by brownie where someone who took care of people in their deathbeds and what she said is very clear watching 2,000 people facing death none of them had a wish that they worked harder none of not one of them wish they spend more time working became more famous or wealthy the number one wish that people had in their deathbed was that they wish they had more time to spend or they had spent more time with the people they loved or the people who loved them so that brought meaning to my life that gave me a strong realization to determine that if I could be cured this is how I'm going to change and live my life so fortunately my treatment was successful so I can be with you here tonight and that I live my life differently I still work very hard but I always spent at least one week per month with my family and that when I spend time with them I'm not on my phone I'm not answering we had an email I am with them at a times when they needed me when my daughter takes off time from her college I take the month off or at least a week two weeks off with her I spend time with my friends and I feel that the love that I give and the love that I receive has made me much more complete and it gave me a real meaning of life compared to the previous robot that I was in turning myself into iron man working machine so in this realization came a belief that if we all had that realization maybe the AI based world won't be so bad there might be the people who in 15 years will lose the jobs that they do today that routine jobs would be replaced by AI but there will be something else that they can do that will not only pay give them the means but also gives them the meaning and that meaning is about compassion and love this is something that AI cannot do we saw that when alphago defeated Korea several months ago Korea came to his tears and everybody was on his side because of his humanity and we also realized that alphago had no happiness from winning and nor did they want to hug anybody after it won the games and that sense of self-awareness and love and happiness is something machines can never have and that we have to come to realization that before we worry about machines becoming human we have to be careful that we don't get turned into machines and if we think about a profession like a doctor if we if someone goes through the experience I did imagine a cold robotic doctor that says you have fourth degree limb for stage lymphoma and you have a twenty six percent chance of living five years how would the person the patient feel when that happens what we need is a humanistic doctor a doctor with compassion a doctor that talks about look this can be treated carefully had the same lymphoma and he's well if you follow the same regimen that treatment he did you should have a good shot and the doctor should visit the patient at home give him/her her confidence and we know through the placebo effect that when people are given that support and confidence they have a higher chance of recovery so that really gives us points the way that we need more people to be doctors we need more people to be compassionate doctors we need to infuse all the professions with in caring for people we need to create new jobs that are about caring for people if you read about the number-one job shortage in America it is about care take curse it is a job that is very much needed when people that ask me when can we could provide when can we make a robot that can take care of the elderly I think that is an immoral question to ask how is it not our responsibility to take care of our parents and even if we cannot take care of our parents 24 hours a day we should at least afford a human to take care of our parents I had a friend who once created a small robot for for elderly a companionship and it had all kinds of functions you know watching movies getting movie tickets and ordering takeout all kinds of capabilities with the robot and guess what function got used the most it was customer service and guess what customer service the elderly wanted they just wanted someone to talk to and that is the abandonment that we humans do not deserve to live if all we can do is put a robot in front of the elderly so that's prop job of the caretaker is a noble profession it's a profession that should be honored in our society it should not be looked down upon as a low-paying service job so we need to create more jobs like that and also jobs of happiness a wonderful concierge a good tour guide a wonderful masseuse someone who comes and helps clean up your your closet every season someone who comes to your home and cooks a wonderful meal these are the human jobs that human touching jobs that we can once finally afford to have because AI will make all this money and then we can turn that money and tax those techno elites and create jobs like this and I think that will be a world that we can all look forward to is a peaceful coexistence of AI tools that we use for creativity for value creation for smart jobs for smart for the innovation for the scientific engineering jobs for storytelling for great PR marketing entertainment movies so that will be one segment on the top of what we call creative jobs using AI as tools and then there will be many more service jobs that are really jobs of compassion and that we need to make sure these jobs get created we need government policies we need a venture capital ecosystem that fund these companies they are for-profit but they may not become unicorns and they will create these jobs that will employ people and most importantly not just give people a payment for their services but bring meaning back to their lives so it is with this realization that I understood thirty-five years ago I was totally mistaken I went after trying to understand how our brain works and I thought that was what AI is about in the last 35 years I am most proud of my colleagues who have made AI incredibly capable in single domains are able to do job and tasks routinely for us and I think that is an important thing for mankind because it is never the intent for human beings to exist in this life for routine jobs once we are freed from routine jobs we can go back to thinking deep thoughts to love our hobbies to share compassion with each other and to make the world a place with much more positive energy and finally the realization I have that I was so wrong is that the most important organ that we have really is in our brain but our hearts thank you thank you doctor leave that incredibly inspiring speech someone who played such a vital role in Microsoft Apple and Google what do you think are the key factors behind creating a successful tech giant hmm well every tech giant is different so I think it's um not it's too simplistic to think there's one formula there are some commonalities there's always a founder with who is not a normal professional manager there's a founder who does not believe in defeat does not accept no for an answer insist that his or her vision is correct and goes after it with a strong degree of confidence and vehemence until he or she succeeds that part is common in all of these companies and then each company has maybe a little bit different culture I think it's important for each company to know what its culture is so that you can build a company around it you can hire people who are like that and then the people will bond and work with each other through thick or thin so for example Microsoft's culture is building an incredibly complex product so complex with so many people that others can't do it that's very admirable Apple's culture is around wowing the user and really putting the user needs in front of everything else very admirable Google's belief is open sharing of information through technologies done by small teams of super smart people so that they don't have to go through inefficiencies of bureaucracy so all very different beliefs so I would advise when you go out and seek the tech giant you may want to work for think not only about how great their products are how high their stock prices are but whether their core values and their culture matches you thank you so after your distinguished careers at these firms you have now founded your an investment firm what impact do you hope that this firm will have on the technology world I found the sign of Asian ventures eight and a half years ago when I felt the ecosystem was lacking in the early stage that entrepreneurs needed help to get started and they were taken advantage of by mid late stage VCS and I wanted to build a company that will be Pro entrepreneurial that will supply the much-needed know-how by early entrepreneurs who often have a lot of shortcomings I wanted to help fill those gaps and increase their chance of success and accelerate their speed to build to build a team and then build the products and that continues to be something that drives us and in fact I think coming to the UK this time I also see so much brilliance and energy but I also see certain gaps in the VC ecosystem so I hope there's someone here also hoping to that's able to fill those gaps so you've spoken about Chinese innovative industries as rivaling those of the US do you think that government censorship puts this position at risk censorship for different countries no within China how do you feel about the fact that government censorship might put the innovation within China at risk when rivaling big companies in the US yeah well I grew up in the West I obviously would prefer environments where I can speak my thoughts and I think countries which I think we sometimes assume countries have laws for bad reasons China is a developing country the level of awareness understanding is not very high so I think it's views based on cultural tradition how to teach the kids growing up is very different I think it's hard to put simply right or wrong label it is a different culture different form of governments but to answer your question about innovation I actually think most people in Silicon Valley and the West are overstating dramatically the requirement for freedom of expression you know as as a fundamental requirement to innovate and create business value I think if you have an environment in which you know what is legal and what is not every environment is like that you just don't venture into the things that are not legal that's the same way in US UK or China and that is not a fundamental suppression for the company to succeed it is perhaps correlated with innovation but it is not a hundred percent correlation I think the existence proof of so many innovative companies in China I think is I think this proves that statement as a fundamental assumption in theorem if you look at the Chinese eco system today China has the most innovative payment system it allows 650 million people to pay each other without Commission peer to peer instantaneously and at 15 cents per trend as low as 15 cent level of micro transaction and that is incredibly innovative and uplifting in in in helping other people create more more companies I just saw a mobike outside the Oxford campus that's one of our portfolio companies it wouldn't be possible without this method of payments China has a strong personalizable News service that is addictive that people are on at 70 minutes a day that is more personalizing and more capable than Western products China has live-streaming and many new business models so I think there's no shortage of innovation and I think it would be wrong to assume that the lower level of freedom of expression has a directly negative outcome in terms of product innovation okay how do you think that regulation can be best tailored to favor innovation and where do you think the West does go wrong where Eastern companies have been able to innovate quicker and develop faster yeah that's another good question about policies I when I I think people in the UK love to talk policies wherever I go people talk about policies regarding privacy and security and bias and how does each person control his or her own data how do we deal with the trolley problem and autonomous autonomous driving dangerous how do we minimize that and health records and and the like and I think that is the culture of UK and that is why UK is the great country that it is but in China there is a techno utilitarian approach that basically says let's get a product out there and iterate and collect data and fix things when they break and be very fast iterating and willing to change the laws when things prove are proven some other way and again neither one is right or wrong but I think in terms of making AI progress I think analysis paralysis will be the West's work with the Western world's biggest danger to fall behind in the technology competition okay so in your speech you mentioned the threats of people losing their jobs through technological progress and in recent years we've seen the rise of populist s-- running on a platform of protecting people whose jobs are threatened by this technological progress how do you suppose that governments and businesses I guess might be able to tackle these issues whilst preventing entrepreneurship from being sacrificed I think we have to look to the future not to the past I think we need to protect jobs but not jobs of the past if a job is going to be threatened by a technology it is going to be replaced sooner or later the later you let it happen by artificial means will only make your country fall behind countries that are willing to take the big hit right now so I think a proactive government should support technology let the jobs that need to be replaced be replaced but put programs in place that create other jobs that are longer lasting and sustainable and help the transition and the social welfare and the support and the retraining so that the tree good transition of the workforce takes place not simple preservation of the old jobs if the job is meant to be gone it's better to let it be gone right now and get to the efficient technological solution and and help those people move on to their new jobs I think that is the program while I like to see in progressive governments and at the moment do you feel that that that's thing lacking around the world or do you feel that one particular region is doing it better than another I think Europe and well you're not part of Europe but III think I think Europe and UK tend to be tend to be a little bit conservative in accepting the inevitability of technology revolutions I heard some surveys where half the people want to slow down technology I think that is not only impossible to do but detrimental for any any country to do because they will set back your economy so I think Silicon Valley is forward-thinking I think China is forward-thinking I think the whole world should be forward-thinking because their pace of technological changes so fast that the moment you think you want the stifle innovation is slow down technological progress it's going to pull back the whole economy and that's not something that any country can afford right now do you think that this fear of innovation within Europe and the UK stems from government or business or maybe just kind of culture of society well I think the UK is an incredibly innovative our country going back from philosophy innovation are the scientists and even in the field of AI I mean deep mine is the best company in the world probably in terms of AI innovation so the roots are definitely here I think what the UK currently is challenged are one is I think a relatively protectionist conservative mindset that is dangerous to progress as I mentioned - I think too much government policy and control that may slow things down 3 there is not a strong enough VC ecosystem to help the new companies get started an entrepreneurial mindset and I think the potential is definitely here and I think those issues need to be addressed and I think there's a fourth one maybe the fact that UK is a small market it's it's actually unfortunate that the UK is not that small if UK were a smaller market as Israel or Singapore people would start thinking global would start thinking about building products for China or for us but UK is kind of a medium market and that gets people to think okay I'm gonna start the company for UK first then we'll figure out how to go abroad but the pace of change is so fast if you don't start your product on a large market then you're going to be copied or you're gonna have competitors by the time you enter the large market so I think switching the mindset to maybe thinking more like Israel or Singapore and less like the British Empire that would be helpful okay so what would be if you had to give a piece of advice to students here today who may be interested in pursuing a career in technology what would be advice that you'd give them I think the first thing is given there isn't a strong ecosystem and aren't a bunch of great startups to join I think that people want to go into tech should go into a great company a company like Google company like Apple our Facebook or deepmind and learn all that you can and then I think when you become surrounded by brilliant people and culture you will figure out what it is next that you will want to do because I think universities you know Oxford there's an amazing place in the university but it is a little bit far from the hubs of Technology so I think get yourself closer by going to a great technical company to work and probably slightly preferable to go to the headquarters of that company that means for many of you since language is in the problem going to work in the United States and with full intent of coming back but I think the Silicon Valley really has a lot of magic I think that is worth to be surrounded by it learn from it and then if you're destined to start the company that kind of surrounding will give you just the kind of training you need and also consider if you're really going to go to Silicon Valley you might join a startup there I think the ecosystem there is strong and you will allow the I think you were giving me example at dinner that sometimes these internships and first jobs are as for telling you what you don't want to do as it is for what you do want to do so I think use all your internships to go to the best technical companies in the world and not limit yourselves to the UK and that will give you a lot stronger hints than what you can learn from the media internet or even the professor's here great so on that note I think you've got time for a couple of questions from the audience if that's okay if you'd like to ask a question raise your hand the member in the aisle right there someone was to fly thank you dr. Lee and I got one question for the technical part so you said a lot about the data-driven algorithms about leap deep learning but like most of the well-established learning now it depends on the huge amount of data set like these are two major drawbacks maymay coming like first of all we find if I'm looking at a cat a single cat I'll be able to recognize at a second time but like for this kind of small data sets impossible traders to train good and you know network or deep learning stuff to recognize it and also like because you need a huge data set the data set main itself not be accurate for example I'm working on medical data set and some of the matters of the label may be like 80% accurate almost 70 percent of accurate so if you're training on this kind of data set the algorithm the output the bass output could be just lower than that kind of accuracy so how do you think that a future like whatever it is deep learning or whatever it is whatever we may call it this kind of a new algorithm will tackle these kind of problems thank you okay okay so it's about having small data sets no good you need large data sets but if they're not accurate what to do there are many applications that are self labeling that don't require any labeling for example if we were going to make I didn't give you this example but we invested in the company that's going to give about 25 they gave about 25 million loans last year so it's a phenomenally successful company and the loans were made by an air algorithm not by a human but how would you train such an algorithm if you have loan officers label each person as vulnerable not loanable then there could be a lot of errors and judgment and bias all built in but if you actually train the system based on actual outcome that is if you really have the guts to give out those 25 million loans and realizing you're gonna lose a bunch of money by people who don't pay you back that is the best kind of data because the people who don't pay back are the negative exemplars the people who pay back are the positive exemplars so find ways within your domain that the data can be automatically labeled by an outcome not by some human expert applying subjective and but potentially biased judgment but some things still have to be done by hand let's say you know face recognition you have to label this is John this mary's face and so on for that there are many data sets available on the internet that can be used for example the koko competition the imagenet competition the stanford a reading comprehension competition and also we funded something called the AI challenger competition in each of these competitions we spent hundreds of thousands if not millions of dollars cleansing the database purely for the purpose that we don't think students and researchers should waste their time label the data so I think it is better that you use a standard data sets that's already been clean paid by us and that you don't have to waste time labeling yet and it's already pre labeled and accurate and that you also have a benchmark of other research who views the same data set so you know how you stand so I would strongly advocate that for your research you pick a standard data set and not take something new because the cost to label it for you is not worth it as a students and then there are these standard sets that exist and you might as well use it thank you should we go to the question over here hi do you think it will never be possible for us to create robots that's able to love robots that are able to love us or the end that we're able to love robots that are able to love us love each other like love robots I see there is the pretense of love right I think copying human emotion can be done I think you already see little robots made by Sony and Softbank that are pretty cute are somewhat lovable they exhibit somewhat human emotions but it really isn't real see I think for the human part of it is the human human spiritual connection that's connected as in boundedness for thousands of years it's just not there with a robot partly it's because we know in our hearts even if they fake it pretty good they're still faking it it is not real it is just pretending plus I think even if they're good do a good job faking it they're going to mess up maybe one time out of let's say 250 and imagine if you had a beloved boyfriend who is a wonderfully who's a robot who's a wonderful tender supportive but one time out of 50 he just goes nuts and says something totally out of the blue and it shows totally no sensibilities and and no love and the ridiculous well will you keep that boyfriend probably not so so so I would say the bar is pretty high for love right if I have a customer service engine that makes one mistake out of 50 that's okay if I make a loan officer that makes one mistake out of 50 that's okay but for love it's something that needs to be real and from the heart and and an eternal and you cannot make silly mistakes that are ridiculous so I think it's really not possible in to do it on a general basis having said that I have absolutely no doubts that the movie her will come true for certain people who crave that kind of conversational imaginative non anti human kind of love so there will be certain people who will fall in love with machines there's no doubt that will happen but I would argue it's a very small minority thank you I think we've got time for one more question Katie the lady in the aisle over there yeah so as a student in the Oxford internet Institute there are the past half year have been hearing so many one-sided and biased criticism saying that many the social credit system in China such as sesame credit by Alibaba have the issue of discriminant people from accessing different public services and also assigning people identity and score without giving the right of explainations so what is your opinion on these kind of AI based social credit system and what do you want to respond to these kind of bias so one-sided it criticism by the West it's against sesame credit in particular yes and it's also saying I think there's also some bias report by the West News saying that this kind of the third party system is a tie to the government serving as I think these are pretty buyers and not you so what is a response and what's your opinion on our base okay well well in part I sympathized with the because that shows you the power of the Internet Giants I don't think any one company should have that much absolute power over the future of people's credits on the other hand I I haven't checked my possess me credit I don't know how good it is but I actually believe Alibaba can do a very credible job that's better than say the American credit card credit score why is that because well it's all numbers right if you look at an American credit score that gives you a high likelihood of getting a mortgage or not or getting a credit card or not it is based on arbitrary numeric addition based on your previous behavior of whether you pay back something or whether you how much income you have how many times you change a dress they have some ridiculously humanly humanly created numerix that basically linearly add up a bunch of random numbers that cannot be a better credit score than a deep learning machine that is trained on real data so can the can each have some bias yes can each algorithm be improved yes but I think the existing credit score is ridiculously simple and I think we should be ashamed as AI scientists if we can't do a hundred times better than them and now having said that you know would sesame credit or somesuch credit system have flaws or biases built in and how can we remove them I think there can be efforts made to remove the bias the largest bias is human subjectivity so to the extent there's human labeling that's what will perpetuate bias because humans are subjective and they make mistakes so the more it is based on objective data the less bias there can be you could also make a point that certain credit scores will discriminate against certain race gender age category and so on because those are provided as features that the credit can determine they get against for example if older people are used at an example because I fall into the category if older people don't repay long as much then it gives less loans to other people well that seems biased well if you really really feel that way if this is a societal consensus that age discrimination is hurting the progress of the human race then remove age as a parameter the system will still operate so I think the the machine learning gives us the ability to keep everything objective and quantitative and remove the items that we think we need to do so-called reverse discrimination to ensure the lack of bias so it's all very doable within our quantitative system and yes any study can look into specific companies and products and say they're flawed but we know the system is better and it can can be fixed I would also warn finally warned against removing too many features in order to remove bias because if there are 2,000 mm features and you start saying well races needs to be removed gender needs to be removed age needs to be removed address needs to be removed sip code needs to be removed and when you remove the system into nothingness then the deep learning won't work at all so I think we have to strike a balance with removing some features that you think are truly harmful but really letting machine learning run its full power and as long as things are based on objective real outcomes and data and not human subjectivity I believe the the systems will prevail and do better than people in terms of less bias well thank you for your question and indeed please could you all join me in thanking God you

Channel: OxfordUnion

Views: 49,922

Rating: 4.8272424 out of 5

Keywords: Oxford, Union, Oxford Union, Oxford Union Society, debate, debating, The Oxford Union, Oxford University

Id: 8wOqP6fZuto

Length: 57min 53sec (3473 seconds)

Published: Mon Feb 19 2018