Where on Earth is AI Headed?

[Music] uh welcome everyone to the second lecture in the microsoft research indian institute of science ai seminar series i'm delighted to introduce tom mitchell today who will be talking on where on earth is ai headed now tom is the founder's university professor at carnegie mellon and he's one of the original machine learning pioneers and he has a string of firsts to his name including starting some of the first major international conferences and journals on machine learning as well as starting the world's first ever machine learning department at carnegie mellon which has produced some of the finest thinkers researchers and engineers in the space of machine learning uh and of course tom has himself graduated stars such as sebastian throne and or in ezioni and i know he's had a profound impact also on thoughts on yokins yorkins tom is a elected member of the u.s national academy of engineering of the american academy of arts and sciences he's a former president of the triple ai and he's testified on numerous occasions to the u.s congress chaired committees of the u.s national academy of sciences and he's also advised many companies on their ai strategy now tom once told me that he never goes to bed without first thinking of or starting a new experiment and i think this is this curiosity and this desire to learn and grow is what has led him to make such highly impactful contributions in diverse areas ranging from never-ending language learning to brain fmri and from theory to algorithms to applications and uh this has led to numerous awards in terms of the hki uh computers and thought award the icml best tenure paper award and so on now in addition to all of his technical accomplishments uh i think tom is one of the nicest most generous and kindest people in the area and he's mentored many of us here in india for which i'm particularly grateful in fact one of the most amazing things about tom is that he started his career flipping burgers in mcdonald's and once they realized that hey this guy can do math they promoted him to the cash register and there's been no looking back after that right as they say the rest is history so i'm delighted to welcome tom but before we get started a quick note on logistics so we've reserved one and a half hours for tom's talk so there should be plenty of time for q a afterwards so if you have a clarification question or if you have any questions actually just type them into the chat and if it's a clarification question tom can take it during the talk or if it's a discussion question you can take it afterwards and helping tom today and helping him choose your questions and ask them on your behalf we have a team of moderators including uh bill thies and amit sharma from msr india and partha pratim talukdar from the indian institute of science and google research and of course we also have akshay on b and venkat patman from the previous session so with that tom thank you so much for being with us here today and over to you oh thank you so much manic and i have to um admire the ability your memory ability because some of those uh facts that you revealed were uh shared with you quite a while ago in an informal conversation so um i i admire your memory um and thank you for such a gracious introduction uh what i want to do today is kind of the opposite of what i usually do when i give a talk you know usually i'll have some algorithm or some idea like how to use unlabeled data and multiple classifiers to estimate accuracy and i'll go into deeper and deeper detail about a particular approach today i thought what would be interesting is the other extreme you know we've all been immersed in an amazing explosion of ai progress over the past 10 15 years now and i thought what would be fun would be to look at that much more broadly see where it's make some remarks about where i think it has been and importantly where it might be going and then use that as the basis for what i hope will be a very interactive discussion after the talk so that's the plan um so let's get started um of course we should start by looking at where we have been um and over the past 10 years 15 years it's easy to see that computer vision which is an important part of ai has gone from being really not very competent to being in some cases even better at humans at doing things like finding objects in scenes similarly in speech recognition um when the iphone came out around 2007 15 years ago um you couldn't talk to the iphone we don't even half of us don't even remember what it's like to not be able to speak to your intelligent agent in your iphone but again over the past decade we've seen tremendous progress up to the level where computers can now transcribe spoken language to the equivalent text stream just as well as humans they don't understand that language as well as humans but they can go from speech to the words that were spoken just as well um in the area of game playing we've seen tremendous progress where for example a.i beat the world's top goal player the world's top poker players and this is interesting because it's a different kind of intelligence from computer vision or speech recognition those two vision and speech are examples of perceptual abilities that happen really automatically and sub-second in our brain they're what some people call fast thinking instant reactions game playing on the other hand strategic reasoning is a very different kind of intelligence slow thinking that requires multiple steps chaining ahead thinking well if i do this and my opponent does that then i could do this but then they'll do that and you have to think about things much more deliberately much more consciously and again there we've seen tremendous progress and of course in the area of robots we've seen tremendous progress we're starting to see for example autonomous vehicles self-driving cars starting to drive on some of our highways so it's interesting that we've seen progress in all these different areas but of course there's a common cause to these the the same factors that work in these different areas of progress and ai it's machine learning and to a large degree deep learning algorithms that have made it possible for for ai to make this kind of progress in so many areas now that's kind of a quick glance just to set the context on the last decade or so but i want to point out that this rate of progress is not slowing down in fact in just the last few years there have been several very interesting and important trends for the field one of them has to do with training and then reusing large deep network models this started in the area of computer vision where by i think 2015 we were seeing that there were large trained models that had been trained on a data set called imagenet which had a lot of photographs labeled by hands like the ones we saw in the previous slide but then those models could be downloaded and used by other researchers and developers for other things and so for example this slide shows an example involving skin cancer diagnosis where a model was trained to look at the blemishes on your skin and diagnose those as cancerous or not and the interesting thing about this from an ai point of view is that first this model was trained on over a million standard photographs not photographs of skin blemishes but those imagenet photographs that had pictures of things like antelopes and lemons that i showed in the previous screen nothing to do with skin cancer but those models had learned enough about the lighting characteristics of images what kinds of low-level edge and contour features should be decoded from the raw pixels etc and so those models were actually a great starting point for training skin cancer model and so that model that was developed on over a million standard photographs was then retrained using only a quarter of a million skin cancer images labeled as cancerous or not and which kind and in the end that led to a computer vision model that was as good as doctors in terms of being able to look at a new blemish and diagnose it as cancerous or not so this was a really important development in machine learning started out in computer vision but then it has with a kind of with a vengeance moved on to natural language where we're seeing the same phenomena and we see systems now that are trained systems like bert actually every major tech company now seems to have at least one of these google first came out with burt but then microsoft now has turing and lg and other companies have um other versions but these are models that are trained on just text that's not labeled by a human but in a way it's self-labeled because the task that these models are trained to do is simply to predict the next word so given any kind of collection of text it's easy to mask out a word show the previous words and have the model uh trained to predict the next word in fact you can also also mask out uh not just the last word but words in the middle which some of these models do but the point is that these models turn out to be surprisingly useful at least to me i never saw this coming and many of my friends i've talked to never saw this coming but it turns out once you train these models not only are they pretty decent at predicting the next word in the sentence but you can now use those to pre-process any text in order to do a large range of natural language processing tasks ranging from question answering to labeling parts of speech which which of these words are nouns and verbs and adjectives to uh predicting not just the next word but a whole paragraph of text that may come from this and so i want to just show a couple quick examples of recent uses of these pre-trained language models one of my favorites is has to do with common sense one of the age-old problems in a.i has been how everybody agrees that common sense knowledge is essential to human intelligence and one of the questions in ai has always been how are we going to get all this common sense into a knowledge base so that ai systems can use it well it turns out these language models simply because they're learning the sequences of words that occur have implicit in the billions of trained parameters a large amount of information that's relevant to common sense and so this slide for example shows some work by yejin choi at the university of washington who's developed a system with a nice web demo if you want to try it out yourself where you can type in a sentence like tom lectured at the nasa virtual retreat and it will then uh type out a number of natural language phrases that correspond to different kinds of common sense so for example if you type in this sentence it will say well that's because person x that's tom in this case that's because tom wanted and it will say to share knowledge or to teach others but then a different kind of common sense before tom could do this tom needed to prepare for the lecture to pray for the talk as a result tom feels good about themselves as a result uh tom reed uh person x reasons that tom could be a professor at the university now notice none of these things were actually mentioned uh explicitly or in any way in any way in the input sentence but this model that yejin choi developed is able to output these kinds of common sense phrases for literally any sentence that you input that describes some event so that's what i mean when i say these models are surprisingly effective at tasks for which they weren't directly trained now these models tend to be trained on very large volumes of text here's an example google's latest paul model is trained on 780 billion tokens of text the different tasks on which they evaluated are very diverse the common sense reasoning to natural language inference here's here's an example from the google paper to show you kind of where things are right now at this point you can take one of these trained models in this case it's the palm model which was recently in the last few months released um and you can input us a question and try to get an answer in fact the way you um set up the task of question asking for the model remember the model is just trying to keep predicting the next word in a sequence so the way you set up this task is you literally given this kind of sequence of words where you say q roger has five tennis balls he buys two more cans of tennis balls each can has three tennis balls how many tennis balls does he have now and then you continue and you say a colon the answer is 11.

you can give maybe a dozen such examples of questions of arithmetic questions and their answers and then you can input your own then you can input as the rest of the input just a question and then the a colon and let the model predict what will be the next words now interestingly it's this is a an example that fails the answer is not 50. but interestingly they report in the paper that if instead of inputting just the questions with the raw answer if instead you input the question with an answer that reasons through the chain of thought so if you train if you set up the context with a question and instead of saying the answer is 11 if you say answer roger started with five balls two cans of three tennis balls each of six tennis balls five plus six is eleven the answer is eleven then that's sufficient to make it when you input your question to get the right answer and furthermore to prompt is sufficient to prompt it to output its own chain of analogous reasoning about cafeteria had 23 apples and used 20 etc okay so again this is not a talk about language models but i just want to quickly summarize the state of the art on some of these systems these same models are now beginning to write code for example on the right here is an example where the input prompt is literally a comment string um it says given a string representing musical notes parse the string compute the total length of music and beats it gives the legend and the output of the model is literally this python code which will compute that so these models are now being trained on a blend of text and code and it's being discovered that to some degree they can translate literally text to code in symbol cases they can summarize uh complex patterns this is an example from a gpt-3 which has some nice examples online you can look at but one of them that one of the its uses is you give it a paragraph of text and you ask it to summarize this for a second grade student and it does okay so you get the idea things are moving quickly in this area beyond language modeling we're starting to see the use of deep networks for a variety of ways to support basic science one of the most visible recent results here is the alpha fold system from deep mind which does the best job so far of predicting the three-dimensional structure of how a protein will fold given just its uh chemical description just its atomic descriptions its sequence and here you see i won't go into the details but these light blue bars are the error rates of the previous state-of-the-art models and this is the aerate for alpha fold one of the interesting things about alpha alpha fold as a case study is that it is um one of the first deep network models that ingests background knowledge about physics and biology as part of its solution and this is another trend that we're seeing in very recently in deep networks is a trend toward trying to find ways to train these networks where we can input background knowledge about in this case biology and physics but depending on the application of course lots of other kinds of background knowledge as part of this we're also starting to see that data that was collected previously for one purpose can sometimes be used for another purpose this slide shows an example where data that was collected on retinal images the part of your eye where the photons hit and it gets converted to neural signals that data that was collected originally for doing diagnosis of eye problems it turns out can be used surprisingly to detect for example the age of the person within three years it can be used to detect whether they are a smoker what their systolic blood pressure is within 11 points whether they're female or male and whether they've had a heart attack so this is an interesting example too because it highlights the fact that data collected for one purpose and maybe privacy considerations that were considered at the time might become out of date as the technology progresses and we discover that from that same data we can discover new things a colleague of mine laila webby is just in the midst of sending a final copy of the paper to nature that shows um that images of fmri images of brain activity uh which were collected for from anonymous well from people who were where the data was then anonymized that that data can in fact now be used to detect oh this person in this experiment is the same as this person in this other experiment so there's another example of technology progressing to the point where we can get new information that we didn't realize we're going to be able to get from old data we're also starting to see all kinds of work i mean i think a summary statement for computer perception and ai is that we're seeing superhuman perception in many ways some of it has to do with specialized perception skin cancer diagnosis but part of it is also just a change in the physics this paper shows for example the ability to look around corners here's the sensor here's an object being sensed behind a wall but by looking at just the reflection of light off this side wall it's possible to reconstruct what was this view and here's the reconstruction that's made just from the scattered light off the white wall despite the black wall being between the camera and the object so when you start combining these new technologies for sensing with deep networks and new technologies for perception i think we're well on our way to a world where superhuman perception is feasible in specific cases and in some sense already here okay so so much for [Music] where we are today it's always fun to speculate on what's coming next of course nobody knows the answer to that including me but we all have opinions and i think it's uh fun and useful to talk about those opinions recognizing that of course we could be wrong so let's see if we're going to start trying to understand what's coming next i think it's really useful to look at what are the technical trends we've been kind of doing that in the last few slides rapid progress in natural language one that i didn't mention but which i think is really important is the rise of conversational interfaces systems like um alexa amazon's uh conversational interface [Music] like the that's um used to train these language models we're also seeing a big bump in research on uh how to get humans in a in the loop with ai systems including how to make more ai systems that can do a better job of explaining um their outputs and can be more transparent so these are some of the technical trends that i think we're going to see um continue over the next five and ten years and probably even accelerate so i list these five as examples ones where i think there will be even more research in these five areas and some others than there was say five years ago so there's some of the trends but then it's also interesting to look at where the drivers and governors that are outside of ai one is just the huge amount of money coming into ai microsoft google facebook many other companies are investing and this has caused a sea change in the past decade where we've gone say in the u.s from a situation where most of the research in ai had been originally funded by the government and took place at universities to today where really most of the research takes place in companies and is funded uh privately so it's a it's a different world that'll surely influence the future of ai commodity hardware gpus famously have which were developed as hardware to support gaming systems have famously become the workhorse for deep learning and now new hardware being designed that's even more efficient for neural network style computations bias and fairness are social issues that have come up that are certainly going to guide and drive the future of ai blockchain many of many of my friends think blockchain is going to be a big impact i don't fully understand their reasoning on that but they could be right um maybe the rise of the metaverse the idea of having virtual reality mixed reality systems the importance of ai for national defense and therefore leading to international competition is certainly going to cover color the development of ai as will privacy concerns so um that's kind of a useful starting point for thinking about the future one way to get a signal about what some people think the future holds is to look at a collection of 20 million dollar research centers that the us government has funded in the last two years and planning to fund this year hear the names of these centers in the areas so there are some interesting trends here one is these centers are dominated by application driven centers so ai in education ai and agriculture ai for molecular discovery ai and next generation food systems so much of the bet that the u.s funding agencies are placing on ai many of those bets are centered around specific applications of ai in different fields and i think that's actually a good bet because for many years much of ai research has progressed by really trying to do particular applications seeing how that fails and then extending the technology another thing to notice in this list is that there are several topics for example ai in education where each year there's a new 20 million dollar center being created so these are very big bets this is literally a 60 million dollar set of centers around ai and education this is an area where i'd be happy to talk more but not not um before the discussion period um this ai in online education is an area that i'm very bullish about and i'm doing some work with some online education companies to see how can we use machine learning together with the recorded log data of literally millions of online students how can we learn from that data how to teach better the next student so education is one of those areas agriculture is also another area where each year there's a new 20 million dollar center so anyway these i think are are revealing um hints about where some people think ai is going to be highly important but what will be the next game changing technologies well again nobody knows for sure but here are a couple ideas i think one possibility for the next five years maybe 10 is that computers might surpass humans at perceiving human emotions this slide is from a company called affectiva that basically develops and markets software that will look at your face and judge your reaction now we have seen in the past decade computers going from not being very good at computer vision to being able to recognize a variety of physical objects and diagnose distinguish between cancerous non-cancerous lesions things like that there's nothing in principle it seems to me that would prevent computers from also becoming good at emotion detection and if that were to happen i think that would be a game changer you could imagine wearing those uh a newer version of the virtual reality glasses that when you're walking around at a party um reveals to you the reactions of the four people who just heard your joke that was a little um maybe not as funny as you thought but there's one technology that i think could happen what are some others um i think if we had a breakthrough in privacy preserving technologies this would be a game changer right now many of the uses of ai that could be used like applying ai to huge amounts of medical data are really bottlenecked they're the the reason we're not doing that has to do with privacy issues and if we had a breakthrough in privacy preserving technologies that would be a game changer um what about a technology that makes device theft impossible in many cities now including here in pittsburgh where i am there are bicycles just around on the streets you can pick up one of them scan your app on your phone to the bike and start riding it and nobody worries about these bikes being stolen in fact the bikes are not stolen similarly for scooters they're around and they're not being stolen um why not well they're they have gps sensors in them they have uh connectivity and so if anybody tried to steal them uh we would know where the bike or the scooter was similarly if you buy a high-end car this day these days if you buy an audi or mercedes they they know where they are they're in contact with the home station and so it's harder and harder to steal those cars because the car knows where it is it can report you so if we extrapolate this maybe we'll get to um how much can we miniaturize and drop the cost of that kind of technology if we can that would again be a game changer um one obvious one especially given the rate of progress in natural language even though today's technology is still far from being able to read and understand the morning newspaper what if computers did get to the point where they could do that and build you a summary of that morning newspaper and then have a discussion with you about that news item over breakfast um if computers get to that point that'll be a game changer partly because for the last 25 years since the advent of the web us humans have been preparing for that day by typing in all kinds of texts into computers so that once they can read it'll all be there for them to read and they will of course outread us by a factor of 10 to the 9th because we can't read all that stuff but they will um here's another quantum computers if if quantum computers work that'll be a game changer many difficult to solve computational problems will look very different one of my favorites is what if computers become programmable by everybody toby lee one of our recent phd graduates from cmu recently did a project uh for his thesis where he took a mobile device and he changed the conversational interface and he made it possible for you the user to teach new commands to the um to the phone with a combination of conversation and gestures demonstration so for example with toby's system you could say [Music] whenever i'm driving home after work and i pass the starbucks if it's hot outside order me an iced cappuccino otherwise a regular latte and the phone responds by saying i don't understand you want to teach me and then you can say yes and you can say for example if you want to know if it's hot outside open this weather amp and you tap on it you say here where it says current temperature if that number is above 80 degrees it's hot and then if you want to order the coffee open this starbucks app and you tap the starbucks in and then you demonstrate and explain how to order an iced coffee um so that kind of um programmability by changing our world basically what toby's work is is an attempt to move from today's world where the only people who can reprogram computers are the people who make the effort to actually learn the language of the computer fortran java whatever it is but maybe we'll be in a world in the next five years where computers learn the language of people well enough that each of us will become a programmer to the same degree that we can already instruct other humans we already know how to do that each of us if we can instruct computers in a similar procedure that'd be a game changer we'd suddenly have uh billions of people on earth who have the ability to reprogram their phones for example and combine this with that slide we saw earlier about language models being trained not just on natural language but also on github code and those examples that we saw where the language model was attempting to translate natural language instructions into the equivalent python code so you put these two trends together i think we really are in for it it is very plausible that in the coming decade we're going to see a sea change in programmability which would be a very liberating force for many non-computer science trained people okay here's another one i have my eye on the clock so don't worry i'm going to make sure we have plenty of time for discussion um one of my favorites is um what i'll call learning light bulbs but if you look at the trends these days about the cost of hardware um in particular the commodity hardware that goes into cell phones which are sensors you know like uh microphones and cameras and uh computation and um internet connectivity all these things that go into cell phones the we're producing those at such great scale now that the price of actually producing those is dropping rapidly and so it seems that we're on the verge of being able to produce for on the order of 15 maybe 20 us dollars uh light bulbs that have cell phone functionality and uh so what does that mean that means anywhere where you can plug in a light bulb today um in the future maybe you'll plug in your learning light bulb your ai light bulb and what would it do why would we want this well if we can [Music] if we see some of these other technology trends that i was just talking about including conversational learning um then this might be these learning light bulbs might become the equivalent of the personal computer back in the 1980s by the way there it was unclear whether personal computers were going to dominate the market general purpose personal computers or whether there would be a variety of specialized computers xerox park for example xerox was producing specialized computers ibm was producing the pc in the end the pc won out and the reason it went out was the economics of producing commodity hardware were hard to compete with any specialized market just couldn't produce enough to bring the cost down plus the technology development curve moore's law uh led to the general purpose uh pcs doubling in power every doubling in their capabilities every few years and there was no way the specialized computers could uh duplicate that kind of development curve but i think maybe the same what will be the equivalent thing in ai an ai maybe it will be something like the learning light bulb that will be a general purpose ai device that can be installed anywhere by the way you won't have to worry about very battery power anymore because it is a light bulb that's plugged into the wall but then what will we use it for well let's put one in each hospital room that has a patient in it it'll use its cameras and its microphones to monitor what's going on in that patient room it could be instructed the same way toby's prototype system allowed you sometimes to instruct your phone but for example if you were the nurse you could say to the light bulb in the patient's room make sure that the if the patient wakes up the middle of the night to let me know the light bulb would say okay and monitor the patient furthermore if it had the right machine learning software in it it could [Music] learn a sequence model it could learn what's the next thing that's likely to happen in the room what's what are the typical patterns what are atypical usually if the patient gets up out of the bed and walks to the toilet then they walk back and get back in bed and they don't usually lie down on the floor in between so you can imagine if you had a learning light bulb there that was also instructable by the hospital staff could be pretty interesting the same light bulb in your kitchen would instead help you cook dinner the same light bulb in your coffee shop would learn to for example that typically when people get up and leave the table they bring their purse with them and they bring their phone with them they don't usually leave them on the table and it could call out to you before you go out the door saying hey you left your purse on the table here um so uh the the kiock key question here is will there be the ai equivalent of the pc uh and if so what will it be i think it will uh part of the question here is what will be the combination of sensors effectors connectivity and can we come up with the right can we be creative enough to come up with the software that will allow some generic device to be instructable and reusable in many places okay um maybe that's enough for now about about technology trends i want to end on social trends um and what i'm thinking is i want to end in five minutes and i have quite a few slides on social impacts of ai but let me just [Music] do it this way i think there are really important social impacts of ai i'm not the only one most people think that um but let's go through a couple of them and uh see and talk through them one of them is just bias in machine learning uh people are rightly concerned about the potential for bias and machine learning and unfortunately bias is one of these words that actually has two quite different meanings and two communities all i'll distinguish them by saying uh socially biased that we consider it some decision procedure to be socially biased with respect to some input feature if the decisions output by that procedure are influenced by feature a so for example we could say if the decision procedure d is one that approves loan applications yes or no approve classifier and if a is the race of the applicant then if d is different depending on the race of the applicant then we could say it's socially biased with respect to race okay that's a notion the notion of statistical bias and machine learning is very different basically we consider a machine learning algorithm to be statistically biased if it outputs decisions that are not the best fit to the data to be more technical about it we consider it biased if the expected value of the prediction differs from its true value and the data distribution but in discussions of bias it's important to think about these two different concepts and what really is the problem of bias in machine learning in machine learning it's really um usually the problem is that the machine learning algorithm is designed to correctly reflect what's in the training data but the training data itself is often socially biased so for example if we have a historical set of data that we're using to train a classifier to decide whether to approve a loan application or not and if the historical data shows that more loans were approved for men and fewer for women then if the machine learning algorithm is accurate or statistically unbiased it will of course learn a classifier that continues that same that same social bias and that's usually the big problem in bias and ai another kind of issue that happens sometimes is just that the data is not biased but it's unrepresentative in the sense that it doesn't include a representative sample so for example there have been face recognition algorithms that were trained only on certain faces of certain races and so they weren't very good at for example recognizing african faces so these these are some of the sources of bias and machine learning in this case there is some interesting technical work to try to address this issue it doesn't fully solve the problem but i think it's interesting and important and the technical work in the area of bias for example includes changing the objective function that we train the machine learning system to optimize for example if we have data of loan approvals that's biased against women applicants then we don't have to use the usual training objective function which is of course the usual objective function in training a classifier is train the parameters of the classifier to maximize the number of training examples that are correctly classified but given that we're trying to avoid giving more loans to men than women if we if we want to avoid that then we can change the objective to be the objective i just said plus a second term and that second term can be a penalty for the difference between the ratio of men loans approved and the ratio of women loans approved if we want those rates of approval for men and women to be equal we can simply add that as another term in the objective function and there's a variety of research going on [Music] of this type that basically is saying let's be explicit about which biases we want to avoid let's build into the training objective of the machine learning system uh let's build in terms that prevent or penalize that kind of bias and this is leading i think to a very interesting discussion because now it turns out as this work is progressing we're beginning to understand that we don't really know what we want do we want loans to be given equally to people of different genders men and women different races different ages do we want no age discrimination income discrimination do we want income discrimination well it is a loan so the ability to repay the loan is a relevant variable and so what we're finding is people are making more explicit what these biases are and have the ability to explicitly prevent them is that it's leading to a much more in-depth and precise conversations about what kinds of biases are there and which ones are ones we want to build in the avoidance of and at some point for example if you take the loan example you can see that if if you put too many constraints you basically end up with a classifier that satisfies all the anti-bias restrictions but cannot really stay in business because the loans it's giving are not being repaid so i think the takeaway message there is that this noticing that bias is an important issue in ai and machine learning in particular has led to some technical approaches that can try to address redress that problem but that in turn has led to a very interesting and more precise conversation about biases and what we really want and that's an ongoing discussion similarly in the area of privacy as opposed to bias there are many issues and again there are technology solutions that are being developed to reduce the privacy impact of using big data in ai systems techniques like homomorphic encryption differential privacy federated learning which are effectively attempts to avoid collecting big data and then giving it out to people and instead many of the technologies that are being developed federated learning for example are trying to leave the data distributed in place for example in each of our cell phones but pass around the algorithm to do computations in each location as opposed to bringing all the data together where the privacy issues can be more difficult to deal with and again that's a long story we're not going to have time to cover it here but again we see there is technology work to try to actually change the whole privacy value trade-off curve there is a fundamental trade-off for example if you think of medical care the more privacy you want um the worse medical care you're going to be able to get less privacy you want the better care but now you're sacrificing privacy so there's this trade-off curve and there is research going on to actually not just pick the right place on that curve but to move the curve up um by using encryption and federated learning and other things so that we'll be able to achieve the same thing we used to achieve but at lower privacy cost by using things like homomorphic encryption finally i can't help mentioning um social issue is uh automation and the workforce and the future of jobs this is an area where i happen to have spent some time the last several years i'm i co-chaired a u.s national academy

study on automation in the workforce and i'm currently ramping up to chair another study in this area for the national academy they're doing five years later another study and there we what we found was that [Music] yes ai is in some cases automating tasks that humans used to do in other cases it's assisting humans um one key thing uh we found is that it's not the the best way to think about automation of jobs is to recognize that most jobs have a lot of are a bundle of tasks for example if you're a doctor you have to diagnose the patient you have to come up with the therapy you have to have a heart heart conversation with the patient about which therapy they want to use what are the trade-offs you have to build a patient now some of these tasks are amenable to computerization billing the patient diagnosis more and more is becoming an area where the computer will assist the doctor probably not replace them in the near term maybe in some areas of them but having the conversation with the patient is the kind of thing that i think [Music] patients will continue to want to uh have that conversation with their human doctor so like many jobs being a doctor is actually a bundle of tasks and if you actually look into the details of automation what you see is that automation occurs at the task level not at the job level and so our conclusion about what's going to happen is that actually there won't be that many jobs being fully automated except for single task jobs like tollbooth operator but most most jobs which are multitask what will happen is that ai will influence one or more of those tasks may be automated or maybe assist as in the case of diagnosis for doctors and what's going to happen is the job will reshape by redefining or re-weighting the bundle of tasks that makes the job anyway there's much to say there but again i think in this case the takeaway is that society and us technology developers have a voice in how job automation is going to play out partly because as technology developers we can decide whether to develop a fully automated ai system or ai system that has a human in the loop and tries to use the ai as an assistant in a human computer team that outperforms what either one could do alone and you know depending on the task uh one solution or another might be better but again we're going to have a voice in how this plays out um [Music] by choosing whether we we focus on human in the loop techniques or fully automated okay so let me end there because as i mentioned i want to have time to have a very [Music] lively discussion with you we've talked about a lot of these issues um to me the takeaway is very simple we live in one of the most interesting times in history certainly in the history of ai this is by far the most interesting time in earlier decades ai was a fun laboratory curiosity which was very interesting to discuss in philosophical discussions um with your friends but today it's out here in the world with us it's changing the world that we're in the side effects which some are intended some are not intended have raised issues around bias privacy automation of work and more democracy [Music] fairness and living in this very interesting age where ai is really reshaping to some degree our world i think puts a an opportunity and kind of a requirement on us as technology developers to be aware of these issues and to see how we can help part of what we can do to help is to educate [Music] our non-ai friends about what the technology really is what it really can do and really can't do um and part of our job is to speculate as i've tried to do here today about the future and what might be coming not guaranteeing that it's coming but what might be coming so that we can begin to consider and perhaps prepare for uh the other side effects of of those future developments so with that let me stop and let's open it up to discussion thank you thank you so much tom this is amit sharma from microsoft research uh i think your talk really gave us a lot to think about not just in terms of where ai is going but also our role in it as as technologists and and people who are making these systems uh so what i'll do is i'll just start with some audience questions maybe we'll start with the most popular question 14 people have this question so do you think the current approaches for representation learning uh suffice for real world applications there's a large discussion around the need for neurosymbolic approaches or causality for robust predictions which may be extremely important for real world scenarios what has been your experience on on these great question they do not suffice for what we really want to do um we're greedy right we want more intelligence systems than we have and representations are really key to that and um i i completely agree if people are suggesting that one of the big open problems in ai is how to bring together um there are two ways of phrasing this one way is how to bring together background knowledge with new training data a different way to phrase it is how to bring together symbolic and neural network representations um i think we need to do both of those one of to me one of the interesting questions about [Music] how this will occur is whether in the end it will all be neural so for let me just say this my brain is completely neural uh [Music] and so is yours and it manipulates symbols just fine so there's nothing in principle seems to me that would rule out that a neural system could manipulate symbols since our brains do it on the other hand the current approaches that we have for neural networks don't come close to being able to do that so i think it's a big mystery one of the things i've been trying to track recently in this area is um approaches people have been trying to use to bring background knowledge into neural networks and for example i saw a talk last month by william cohen who uh is now he happens to be at google on on using memories um to augment uh [Music] transformer style models and uh what was interesting to me about this is that the memory was storing um in this case uh as you go through as the language model was going through words in the sequence whenever it would find a noun phrase it would look it up in the memory to see what were all the mentions of this noun phrase in wikipedia but what that memory would return was not the symbolic text of dimensions but the embedding the neural encoding of those mentions um [Music] and they and those neural embeddings were trained so that it could simply add some literally some ad that embedding to the representation that it had of the sentence at the time so what he managed to do the thing that seemed interesting to me there is that here's a process memory retrieval which i think is potentially a symbolic thing for trying to get in from background information about this word but it managed to store in that memory something which was a knurling coating that could be that was in fact devised so that it could be simply added into the neural encoding that was currently being used so sorry for going into so much detail on this but the reason i wanted to is that to me this is the kind of puzzle that we face right now will it turn out that we need graph style representations explicit graphs together with neural networks or will it turn out that we'll find a way to translate uh those into neural encodings and go from there i think it's one of the great open research questions no no that was good i think it's good that you went into some details there uh another question uh that has come up is on the problems that we should be solving so the question is about how does one reason about problems that can and should be solved with ai versus problems that sound like great applications but maybe problematic so for example predicting whether a person would commit a crime based on their photography or past evidence but the questioner is asking they used to think it's it's this was not very difficult to judge but the possibility of superhuman perception uh makes it a little bit more difficult to think about it really does um a couple comments one these are really social questions not technical questions these are questions what do we want not what can we do so i think i you know as a technologist i work on the question of what can we do but the social question of what we want is at least as important in the end but it's a distinct question i agree that the ability to make these kind of predictions is likely to be greater in the future than it has been in the past and furthermore we have a choice about whether to do that um i'm i'm not really sure what more to say about this particular case except that um of a friend dale blumstein who is a very well known a researcher on uh recidivism uh people who are in jail and are led out of jail do they commit another crime and he points out that technology is a way of letting people out of jail safely and so for example you put a tracking device on somebody you can let them out of jail in the old days when we couldn't do that we would keep them in jail just because but now you can let them out of jail and you can tell them we will know if you go to this region of town and we do not want you to go to this region of town and so um al has sensitized me to the fact that technology can be used to uh in this case give people more freedom than they would have had and uh so for me that was an interesting idea and this question makes me think of that but i don't have the answer to that question yeah and it's a very hard question and i'm glad we are having this discussion through through the means of your talk uh one question which two people asked was about uh ai augmented education that that you mentioned uh and and they're just curious to know uh what are the sort of possibilities you see and what could be the future there well uh i'll try to keep this shut me down if i get too carried away because right now this is actually my one of my biggest passions i think that the coming decade is the decade when ai can finally make a difference especially in online education um i've been working with company called ck-12 which is a non-profit company that offers online uh educational materials for kindergarten through 12th grade are actually quite a few people in india quite a few students in india using this software ck12 and they've had millions of students come and use their software and so they have uh here are the kinds of things that i think we can do um the low-hanging fruit by which i mean maybe for the first five years um is uh because we have these trajectories of students going through the curriculum at the level of keystroke recordings so we know things like this student came through and they tried practice question three but they got it wrong they answered b the correct answer was c then they watched this video but they didn't go all the way through they shut down the video at two minutes and 30 seconds in then they read this paragraph then they answered this other question correctly so we know sort of in the timings on all those that's the kind of data we have so from that there are two um two things to try to do one is let's learn to predict the knowledge state of the student that's not directly observable to us we just get these signals from the things they've done but if you think of there being a latent knowledge state of the student we can try to estimate that from this data and the way we represent the latent knowledge state is by a vector of 20 000 questions and the probability for each of those twenty thousand questions the probability that student will answer that question correctly if we ask them it right now so that's how we represent the latent state and of course it's not fully observed but we do observe them answering different questions and we see what led up to then we see so that's one thing is trying to estimate the latent knowledge state of the student the second is a second low-hanging fruit is what you might call the reinforcement learning problem at any given point in the curriculum there are several teaching actions we could take for this student we could show them a video give them practice question 43 show them a paragraph while they're answering the question we could give them a hint hint number five or no hint whatever we can give them motivating statements like good answer you got 10 in a row right now we can do all those things those are teaching actions and so the reinforcement learning question is why can't we learn the right policy for choosing the optimal action at each point in time and what do we mean by optimal we mean we want to maximize the quiz score at the end of the session so those are kind of i think the the kinds of things that we can do and the reason i'm so jazzed about this is that uh up until again this is the first decade where we've had enough data to do that up until recently we never had uh systems that have been used by millions of students but now that we do there's just this opportunity to learn to teach better that we've never had and by the way a human teacher teaching for a hundred years full time would never see as many students as these systems are seeing so this you know data wise this has the potential to turning into superhuman is certainly a superhuman experience base whether we can turn that into superhuman teaching competence remains to be seen but those are the low-hanging fruit that then beyond that you know think about conversational systems um uh the ck12 system has a little chatbot on the side that you can ask questions to and it can sometimes uh jump up and down and offer you hints and things like that all the technology for conversational systems that is accelerating and advancing now at a rapid pace that becomes available for these kind of systems what about things that the student does off screen one of the centers that i mentioned these 20 million dollar centers the us is funding one of them at the university of colorado is focusing explicitly on observing a table full of kids working together to to solve problems using cameras and microphones and trying to and doing eye gaze detection and trying to understand um what kind of interventions with those kids what kind of assistance support with those kids could a computer partner an ai partner that's their phrase who's sitting at the same table provide so i think that the openings are huge the potential is huge and i'm glad to see there's this investment of funding in the u.s i think worldwide this is actually let me make a separate point i think really important way to think about um ai's impact on the world is that there are fundamentally two kinds of applications from a nationalistic perspective say from the indian perspective are applications where it's zero-sum where if one country does well um the other country does less well military applications of ai are the prototype example event if you use ai to improve your military that makes the other countries relative to you fall zero sum it's probably rational for countries to decide not to collaborate on zero sum applications but there are other applications that are the opposite of zero sum they're win-win and education is one of them if we can figure out technology to make teaching work better in india and in belgium and in the us and in china the whole world wins there's no nobody's harmed by that and therefore in these win-win applications education healthcare climate really the rational thing is to have much much more international collaboration than we have now and the people i speak with in government don't seem to get it that there are these two very distinct kinds of applications and we need to do some education of our politicians around that point yeah sorry different issues i think it's an important issue yeah yeah i'm going to switch to manik who has a question uh thank you amit uh tom i wanted to go back to your comments on uh large language models and i wanted to discuss the large uh in these large language models now i'm not an nlp expert so i wanted to get your thoughts on is there something magical that happens or is there a phase transition that happens that when these models cross a certain size then they start achieving superhuman accuracy say a billion parameters or a trillion parameters or whatever and so it does size really matter and so is there a sense that you know if we were able to train let's say a quadrillion parameter model one day on the entire corpus that the world has to offer we'd get superhuman accuracy in a range of tasks and then if the goal is to go larger and larger what are the different ways in which people have explored this right is it just to increase the number of parameters or the number of layers in the network or or what could be done over here yeah i think that's a great question and i think the jury is out but i think the current consensus um in the papers that describe these larg

2022-06-02 03:13

Other news