Transcripts For CSPAN2 Michael Kearns Aaron Roth The Ethical Algorithm 20240713

Where the subject of algorithmic fairness or privacy is not frontpage news today we will speak to the two leading lights in that area they will help us understand the state of the art is now and Going Forward with that we will welcome professor kearns first to the stage. Welcome to the stage. [applause] good morning thank you for coming my name is Michael Kearns with my close friend and colleague writing a general audience book called the ethical algorithm so for roughly half an hour we want to take you had a high level to the major themes of the book then we will open to q a. So many people and certainly this audience is aware in the past decade Machine Learning has gone from relatively obscure to mainstream news to characterize the first half of this decade is the glory. When all reports were positive with all these amazing advances like deep learning and speech recognition image categorization in many other areas we all enjoy the great benefits of this technology and the advances made but the last few years have been a buzz kill many articles have been written now some popular books on the Collateral Damage we caused by this type of decisionmaking is specially powered by ai Machine Learning so weapons of mass distraction was a big seller and it did a very good job to make very real and visceral ways the way algorithm decisionmaking can result in discriminatory predictions and Racial Discrimination david and goliath is a wellknown book we are close to a commercial surveillance state and what a companies that we read these books and to be like them very much and many others like them but one of the things we found lacking as a motivation when you get to that solution section it is what we are considered fairly traditional we need better laws, regulations we need to keep an eye on this. And we agree with all of that but as researchers are working in the field we also know there is a movement to design algorithms that are better in the first place so waiting for a predicted model you could think about making the algorithm better in the first place so now there is a fairly large Scientific Community trying to do exactly that so you can think of it as a Popular Science book to explain to the reader how you can encode and embed social norms directly into algorithms themselves. Now a couple of remarks we got a review on an early draft of the book that says i think your title is a conundrum or the oxymoron. How could it be any more ethical than a ham . A hammer is a tool like human designed artifact for purposes and while it is possible to make the unethical use i could even on the head nobody makes the mistake to ascribe unethical behavior or activity to the hammer itself if i hit you on the hand you would blame me and not the hammer we know real harm had come to you because of me hitting you so basically they said i dont see why the same arguments dont apply to the algorithm. We thought about this and decided we disagreed they are different even though they are tools we think they are different for a couple of reasons because its difficult to predict outcomes and to ascribe blame because especially when powered by ai and Machine Learning there is a pipeline usually use start off with complicated data in the sense very high dimensional with a lot of variables like medical database of medical records and we may not understand where it comes from in the first place and then the usual pipeline or methodology is you take that data to turn it into an optimization problem with the objective landscape and we want a model that does well with the data in front of us it is primarily or exclusively concerned with predictive accuracy theres nothing more natural to do than to take a data set and save find the Neural Network and to decide who to give alone. So then what results is a complicated high dimensional model light clipart from the internet deep learning so theres a Neural Network with many layers so the point is the pipeline is very diffuse it may not be entirely easy to pin down blame was at the data or the optimization procedure . And even worse if this predictive model causes real harm if you are falsely denied alone because the network said you should be then we may not even be aware and because we give them so much autonomy to hit you on the hand with a hammer i have to pick it up and hit you now they run autonomously without any human intervention so we may not realize the harms that are being caused unless we explicitly know to look for them. So our book is how to make things better but to revisit the pipeline and to modify in ways that give us social norms like privacy and fairness and accountability et cetera. One of the things about this endeavor even though many scholarly communities and others have thought about the social norms before us like philosophers have been thinking of fairness people think about privacy. They never had to think in such a precise way you would actually write them into a Computer Program algorithm. Sometimes just the act to force yourself to be that precise with these concepts you would not discover any other way. So the whirlwind tour of the book is a series of discussions about different social norms and what the science looks like to give a precise definition, a mathematical definition in the algorithm and what the consequences of doing that with tradeoffs. So in general if i want an algorithm thats more fair its at the cost of less accuracy. So we have written these in increasing shades of failure to how mature the sciences in each area. When it comes to privacy this is the deal that is the most mature and what we think is the right definition and quite a bit known how to embed that definition. Fairness a little bit lighter but off to a very good start in things like accountability like the morality because we feel there are not even good technical definitions and i promise you that there is a singularity so that the rest of our time we will talk about privacy and fairness. That the twist that it takes half way through. As michael mentioned by far the most welldeveloped fields i want to spend a few minutes to give a brief history and in that process to go through a case study how we could think precisely swimming 20 or 25 years ago with a really had in mind was a data set and to release this to anonymize the records to remove the names of those unique identifiers like Social Security numbers like age or zip code that would not uniquely identify me. In 1987 the state of massachusetts decided to release data set that is useful for medical reset one researchers and the state of massachusetts corresponding to every state employee with data set there were no names or Social Security numbers there are ages and zip codes and genders. So it turns out although age and zip code and genders not enough to identify you in combination they can be. A professor at harvard figured out that you could cross reference the online data set with registration records that also had demographic information and then cross reference with anonymized set with cambridge and then with those identifiers to identify the medical record from the governor of massachusetts at the time and sent this to her desk to make a point. This is a big deal in the study of data privacy and people tried to fix this problem by using a bandaid to fix the most recent attack. So for example if it turns out combinations zip code and age so instead of reporting age exactly maybe we will only report a zip code of three digits and do this to make sure that any combination of attributes that we release doesnt just correspond to one person to know my 56 yearold neighbor who is a woman was at the Hospital University of pennsylvania with the guarantee i cannot connect those attributes so for a little while people tried to do this if you look at the data set you may already begin to realize this is what we mean by privacy because if i know my 56 yearold neighbor attended hospital at the university of pennsylvania i cannot figure out her diagnosis but i can figure out she could have colitis but it doesnt speak to the math i know that shes a patient is now at two hospitals now the other hospital has records the same way. May be a little better because now my female neighbor matches three of these. That both data sets have been released i can crossreference and there is a unique record that could possibly correspond to my neighbor and now i have her diagnosis. This is the overall problem which is the same we tried removing the name maybe attempts at privacy would work if the data set i was releasing was the only thing out there. Small amounts of idiosyncratic information is enough to identify in ways to uncover if i can crossreference the data set that has been released. Some people try patching this up for a long time data privacy was cat and mouse trying futuristic things to patch up whatever vulnerability led to the most recent attack and attackers trying new things it was a losing game. We were trying to do things we thought were private. So this is the approach over two weeks in and attempt to think what privacy might mean but think what a strong and then we will find the right answer. Lets think about what privacy should mean. If i use data sets for medical studies nobody should be able to learn anything about you as a particular individual that they could not learn about you had the study not be conducted and to make it more concrete make it as the british doctors study the first piece of evidence that smoking and lung cancer were correlated. Because every doctor was invited to participate and two thirds did they agree to have medical records included as part of a study. Very quickly there was an association. So imagine you are one of the doctors if you are a smoker you made no attempt to hide that fact youre probably smoking during this presentation but then when its published they know Something Else now they know you are at increased risk for lung cancer because now we learn new facts that it is correlated. In the us it could cause concrete harm that your Health Insurance rates may have gone u up. So if we say what privacy Means Nothing new should be learned about you as a result we have to call the british doctors study a violation of your privacy. But there are a couple of things that are wrong about that. First of all the stories play out in the same way even if you decided not to have your data included. The suppose it violation the fact that i learned you are at a higher risk of lung cancer thats not something i learned about your data in particular. I already knew you were a smoker but the violation of privacy is attributed to the facts of the world that smoking and lung cancer are correlated we know that wasnt your secret to keep because i discovered that without your data from any large sample of population. We call that a violation of privacy then we could do Data Analysis at all because there will always be correlations between what is publicly observable in what you dont want them to know and i couldnt uncover any correlation at all without a violation of this type. So this is an attempt to think what privacy should mean. And the real breakthrough came in 2006 when Computer Scientists had the idea for differential privacy. The goal is to promise something very similar with a slight twist. So think about two possible worlds not where the study is carried out or not but instead the alternative world where it still carried out but without your data everything is the same except your data was removed from the data set. The idea is we want to assert in the ideal world if it was not used at all there is no privacy violation because we didnt even look at your data. And then in the real world it was but if there is no way for me to tell substantially better for me and im guessing in the real world or idealized world then we should say your privacy is only minimal violated what differential privacy says it is the difference there is no way to tell the difference that is substantially better than random guessing compared to the world where we dont hear we can quantify to trade off accuracy with privacy so you may think this is too strong when you think about it it sounds like a satisfying definition and you may worry its too strong to allow anything useful to be done to go through this simple example 15 years of Research Shows any statistical path or analysis which includes all of Machine Learning could be done with privacy although at a cost that typically manifests itself in the need for diminished accuracy. And with the Academic Work this has moved to become something that is widely deployed if you have an iphone for example might be actively reporting statistics from differential privacy but the shock will come in just about a year the us 2020 census will release all the information under privacy. This is the fence that we talk about it is the most welldeveloped not that we know everything there is to know but a strong definition and we understand the algorithm you need while still doing useful things with data and this is real technology. I will talk similar about algorithm fairness the study of fairness is less mature than privacy in particular we already know it will be messier so we argue anyone that thinks long and hard enough will survive so differential privacy is the right definition of data privacy we know there is no monolithic definition of fairness in the past few years there have been a couple of publications of broad form that say can we all agree any good definition of fairness meets the mathematical properties cracks they say yes of course these are weak minimal properties i want them and a stronger winds also. The punchline is guess what cracks eras that there proving there is no definition of fairness that can simultaneously achieve these. So a little more concrete that this may mean in real applications that if you are trying to reduce the discriminatory behavior of your algorithm by gender that could be increased cost of discrimination you may face the moral and conceptual tradeoffs. But this is the reality the way things are so we still propose proceeding to study those alternate definitions and the consequences. So i want to show you how things can go wrong with racial or gender discrimination and how that leads to a proposal how one could address Collateral Damage. So why Machine Learning even in the past few weeks im sure youve heard of these notable instances with a Health Assessment model that is widely used in Healthcare Systems to show systematic Racial Discrimination and less scientifically a twitter storm recently over apple credit card a number of reports of married couples the husband said my wife and i file taxes jointly she has a higher Credit Rating but i got ten times the credit limit that she did we were actually with the federal regulators office that is investigating this particular issue i like the Health Assessment model we dont know if this is systematic underlying gender discrimination but this is the concern we talk about so like the two medical databases take you through how things can go wrong to build predictive models so lets imagine we were asked to help them develop a predictive model for collegiate success based only on two variables high school gpa and sat so what i show is a sample of data points each represents the high school gpa and the why value is the sat score. If this is a sample of individuals that were admitted so we know if they succeeded or not and by succeed pick any quantifiable subjective definition we can measure that success means you graduated within five years to matriculate three. Oh gpa or that you donate at least 10 million within 20 years of leaving thats what the plus and minus mean so for each point the gpa and the sat for those that succeeded in the minus means otherwise so with this cloud first of all a few counted carefully less than half succeeded slightly more minus they on plus also if i just show this and point if you build a good predictive model to predict whether applicants will succeed or not there is a line you can draw and if we predict everybody above is successful and everyone below is not successful we do a pretty good job. Its not perfect but for the most part we are doing a good job and in that simplify aid form even including deep Neural Networks to separate the positive from the negative. So lets suppose it in the steam and the same pool lets call them the orange population first of all they are a minority in the literal mathematical since fewer orange points than there were green points. The data looks different the sat scores are systematically lower but no less qualified for college. There is exactly the same number of orange plus as orange minus is not that the population is less successful in college even though they have lower sat scores. One reason you might imagine this is the case so when the green population they can afford sat preparation courses and multiple retakes of the exam and the orange population is less wealthy with fewer resources do selfstudy and take it once and take what they can get. If we had to build a predictive model for the orange population there is a good one this perfectly separates positives from negatives. The problem arises look at that combined data set what is the single model that did that population and you can see that visually. So i will pick up so many green minuses it will increase by me trying to do that. This is the optimal model on the aggregated question it is intuitively unfair to reject all the qualified orange applicants so we could call this the false rejection rate on the orange population is close to 100 percent and on the green population is close to 0 percent. Of course what we should do is notice the orange population is systematically lower sat scores and we should build a twopart model to say if you are green we apply this line and if you are orange we will apply this line. And the single model of the aggregate data it would not only make it more fair but we also make it more accurate as well. But if we think of green and orange being racebased there are many areas of law and regulation that forbid the use of race as the input. And then to decide and these laws or regulations that prevent the use or air relevant variables meant to protect the minority population. So here is a concrete example where regulations meant to protect the minority population guarantee that we will harm that minority population if we do the most sensible Machine Learning exercise. In the same way that definitions of privacy of anonymous they should all make sense we argue any definition any time you try to get fairness by forbidding inputs is fundamentally misguided. Instead you dont restrict the inputs but constrained the output behavior the way that you want. In particular one thing you can imagine doing even if forced to pick a single model i could change my objective function to say there are two criteria prick i do care about making accurate predictions. On the other hand we care about this other objective which is fairness in this particular application i could define as the false rejection rate im worried about the orange population being mistreated and that particular is false rejection. Those who have succeeded. So what is the difference on the green and orange population so instead of saying minimize that error i can say minimize it subject to the constraint the difference in false rejection rates between the two populations is that most 0 percent or 5 percent or 10 percent. If i let that go to 100 percent disparity then im not asking for fairness at all anymore. So the way it gave the nod how strong your privacy demands versus your accuracy demand the definition of fairness with zero disparity in the false rejection rates to know fairness whatsoever. And with those quantitative tradeoffs so with the different data sets actual numerical pots where the value for each point is that air of a predictive model and the why value is the unfairness of that model with false rejection rates of course smaller is better but i like to be in the corner is not happening on any one of these data sets with real Machine Learning even ignoring fairness you will not get to zero errors. But we face a numerical tradeoff at the other extreme we can ask for zero fairness and get much larger and in between we can get what is in between. And then to become quantitative enough even nontechnical people can look at the tradeoffs to understand the implications. So now to have algorithms to decide so in particular there is a big difference in medical decisionmaking versus the ads you are shown on facebook or google what you may never look at. And with that to occur and then to say hard tradeoffs right from the beginning. And then went to think conceptually about what they tried to accomplish and go through some bad definitions based on anonymity and with that satisfying definition without real data sets and algorithms. Privacy and fairness i wont talk in depth but to give a quick survey of the second half of the book. But then studying algorithms thats all we can think about without necessarily the larger context it is embedded but that context is important what it is doing affects the behavior of people and how they interact. So in the third chapter so how will it change those particular decisions in a way that could reverberate. And we talk about an example that is not consequential socially but it is clear to have an idea what we are talking about. Into what an economist would call again in the sense that the actions that i take, which will choose to drive along have negative externalities on other people in the form of traffic selfishly i would prefer that everyone else stay home and i would be the only one on the road and take a straight shot to work get there will fast but other people wouldnt agree to that solution. Different people have interests and their choices affect the wellbeing of other people, each choice i make a choice i make has a small effect on any particular other person i dont contribute too much to traffic but collectively the choices we all make a large effect on everybody. One way to view these apps like mac abbefore these apps were around i would have at best a very minimal traffic information so i would probably take the same route every day but now i can very precisely respond to what other people are doing and what game theorist would describe the app is doing is helping me compute my best response given what everybody else is doing what can i do that will selfishly and myopically optimize for me . Everyone else is doing the same thing. So the result is that what the job scheduling is they are driving global behavior what would be called the competitive equilibrium, nash equilibrium which is some state stable in the sense that everybody is myopically unselfishly optimizing for themselves. Okay. If you taken a class on game theory or just read the right books you will know that just because something is competitive equilibrium does not mean its necessarily a good social outcome and prisoners dilemma is maybe the most famous example of this. Its not at all obvious in fact he can come up with clear case studies where these apps even though they are selfishly optimizing for each individual person are making change works globally for the population at large is a sense of larger average commute times. That might not be an enormous deal we were just talking about traffic but this is just an example of a phenomenon thats much more pervasive when algorithms mediate social interactions which happens now all the time. For example, you might think the content moderation algorithms that drive things like Facebook Newsfeed and similar context like myopically facebooks interests are not so misaligned with my own in the sense that there algorithms are optimized to drive engagement what facebook wants me to do is to stay on facebook as long as possible so i can view lots of ads. The way they try to do that is by showing the content that i would like to engage with that i would like to click on and read and comment on. Myopically it seems aligned with my interests. I have a choice of what website to go to time engaging with the content that facebook is showing me might be enjoying it. But when facebook simultaneously does this for everybody, even though its myopically optimizing for each person, it might have global consequences we dont like in particular it might lead to the filter bubble phenomenon that people sort of do a lot of handwringing about our drivers globally to a society that is less deliberative. In this chapter we go through a bunch of examples trying to think about and point out ways in which algorithmic decisions can have widespread consequences on social behavior and how game theory is a useful tool in thinking about this. In the last chapter we Start Talking about another important problem which is the statistical crisis in science which some of you mightve heard about and its actually not so disconnected from the equilibrium behavior we talk about in the game theory chapter. There been a bunch of news articles showing that if you take food science or social psychology these are emblematic literatures where if you flip through the scientific journal put your finger down at random more likely than not the study you will have picked up will not replicate if you try to reproduce the results with new data or subject its not nearly as likely as it should be that you will find the same result. Heres a kcb cartoon they were nice enough to let us include in the book which is getting at exactly this phenomenon. We got our scientists here and hes got a tip someone tells him jellybeans cause acne. So he tests the hypothesis and the p value he gets is above. 05 standard level of statistical significance in the literature so he says sorry, no result. Then hes told its only a certain color of jellybean that causes acne so he starts testing them to test brown jellybeans and purple ones and pick ones and for all of his finding a p value greater than or equal to. 05 then he finds one green jellybean that appears to be statistically significant. There seems to be a correlation between green jellybeans and acne at a statistical significance levels of 95 percent which means that if you tested 20 hypotheses you would expect only one of these to incorrectly appear to be as insignificant by chance. Of course he did test 20, green jellybeans linked to acne, only 25 of coincidence. This is called multiple hypothesis testing problem. Its relatively well understood how to deal with it when its just single scientists conducting these studies and whats going on here is really just statistical asomeone has checked a bunch of hypo hypotheses but then is only publishing the most interesting one without even mentioning the others. But of course this is just as much a problem with it rather than one scientist studying 20 hypotheses we have 20 scientists each studying one hypothesis and each following proper statistical hygiene. Its just as much of a problem if only that hypothesis appears to be significant goals in the one thats published and thats what the incentives that underlie the game of scientific publishing are exactly designed to do. Because if you find that blue jellybeans do not cause acne thats not to be published. He probably wont even try to publish it its not a result that any prestigious journalist will want to present. But if you find something surprising that green jellybeans cause acne, then thats a big finding. The problem is, if youve used scientific publishing as a game even if each individual player is falling proper statistical hygiene you get the same effect thats described in this cartoon in the chapter we talk about how these phenomenon are exacerbated by the tools of Machine Learning which promote checking many different hypotheses very quickly which promote data sharing and how tools from us literature in particular, surprisingly tools from differential privacy which we talk about in the first chapter can be used to mitigate this problem. And thats it. Thank you. Thank you very much. That was great. We are going to do questions i know we have a lot of folks in the room who regularly work in this space both pakistan and beyond. We would love examples of problems that you guys have faced or questions that you have to expert by this. Maybe i will start with this, you talked a little bit about the limitations of Computer Science when it comes to answering some of these questions of fairness, having now talked about but probably for a couple months, have you found that the public wants the Computer Scientists to solve this . No. [laughter] i think in our experience, they appreciate the fact that People Like Us can the community we come from can identify the point at which this judgment involved and the sort of world decisions to be made and stakes matter and so i think there are generally appreciative of the fact that both sides need to come toward each other a little bit just like those types of the parade ogres are showing between airing on unfairness. Takes a little bit of explanation to understand what such a plot is saying but in general i think people from nonquantitative areas that are stakeholders in problems like these, people and policy think tanks and the like, they like that but i dont think they are wanting Computer Scientists per se to take a leading role in eking out a point and saying, here is your best tradeoff between airing on unfairness because it depends on the data and the question. I dont think even we think that like kims Computer Scientists should be exclusively or even in large part the ones making many of these judgments and we are careful to say in the book the scientific probably unaware of the problem and then theres the part of the problem that requires moral judgments of various sorts and those are different. For example, we do not propose that should be algorithms are necessarily Computer Scientists who define what it is we mean by fairness. Once you pick a definition we certainly dont propose its Computer Scientists who should be picking out in various circumstances how we want to trade off things like privacy and fairness and accuracy but what is important and what i think Computer Scientists have to be involved in is figuring out what those tradeoffs are and how to make them as manageable as possible. Example, the u. S. Census right now there is literally a room full of people, a committee whose job it is to look at these curves and figure out how we should trade out these very Different Things one of which is privacy which the census is legally obligated to promise to american citizens the other which statistical validity for the data its extremely useful data its used to allocate resources, school lunch programs, important things. There are different stakeholders disagree about how these things should be traded off and they are in a room hashing it out as we speak. But their work is made very much easier because we can precisely quantify what those tradeoffs are. We can manage them and thats what Computer Scientists i think have to play an Important Role in. That leads to another question i had while listening which is an ideal universe where the ethical algorithm is on every desk of every Computer Scientist and the framework you describe are actually used in action in rio and much of this is happening in industry some of it is obviously happening in government as well. What is it look like to have a community of People Living these principles . Is there a public api that we all can see . To use a rudimentary example like when we go to the Grocery Store we can look at the side of the box and we know the caloric intake we know how much fat there is we know much sugar there is. In a world where some people might comply and some people might have read the book some people might not have read the book. What is it success look like . While we dont talk a lot about this in the book, we continue to procrastinate on writing a socalled policy brief for the Brookings Institute where we are going to talk a little bit more about regulatory implications of these kinds of things. The reason i mentioned that in response to question is, once you have a precise definition of fairness or privacy, you can do what we mainly discuss in the book which is embedded in algorithms to make them better in the first place but you can also use it for auditing purposes. In particular if we are specifically worried about gender discrimination for stem jobs in google advertising, which was something that was demonstrably shown to exist a few years ago. You can run controlled studies. You can have looking api where you say we need unfettered access to make automated google queries over time so we can systematically monitor whether there are gender differences in the distributions of ads that people see, for example, so we do think that an implication of a lot maybe not all the work going on in these areas is the ability to do that kind of technological auditing and i think we believe that some of that should happen and in particular can anticipate what the objections of the Technology Companies might be they might include things like thats our intellectual property this is our secret sauce we can have automated queries which currently of course violate terms of service and our response to that is this is your regulator. They wouldnt have this access and be able to use it so for instance start a competing Search Engine in the same way the sec has all kinds of very sensitive counterparty trading data but is not allowed to use those to go start their own hedge fund for example. I think in a world where the kind of ideas we discussed in the book become widespread and embedded, a big part of it the sort of on the side of the cereal box might be things like on the side of the google cereal box heres the rates of discrimination in advertising by race, by gender, by age, by income, etc. You could really imagine having some sort of quantitative notion or scorecard if you like, of Different Technology services and how well or poorly they were doing on different social norms. I think also what we have to see as the regulations for things like privacy and fairness organ have to become a little bit more quantitative. At the moment there is this disconnect were people in the industry are not sure exactly what is expected of them. What is going to count as algorithmic unfairness for example, the issue with the apple hardware there was seeming gender discrimination it wouldve been easy to find that had only people thought to look for it then when we were chatting with the new york regulators a few weeks back one thing we heard that i thought was interesting was that sometimes companies will explicitly avoid running checks like this because if they dont check and theres plausible deniability and if they do check it subject to discovery if theres a lawsuit. This is the kind of thing that sort of flourishes when theres ambiguity but if youre precise about what exactly is going to constitute the algorithmic discrimination in the state of new york then Companies Look for it. I think our view is that even apparently strong regulatory documents like the gdp are our really ill formed document stay look really strong but they push words like privacy interpretability fairness around on the page but nowhere in those pages do they say what they mean. Its a bit of a catch 22 or chicken and egg problem. It looks like strong regulation because there demanding interpretability everywhere but nobody is committed to what it means yet and i think i do think that as is often the case even the science we discussed in the book a sort of running ahead of certainly things like laws and regulations. I think that before the kinds of changes we are discussing to take place on the revelatory side much regulatory law has to be rewritten and there needs to be cultural change at the regulators. That makes sense. Shifting gears a bit, i was struck by the fact that the differential privacy is ahead. You guys have a view of whether thats ahead because as you said there is an objective preference to answer like an answer that can be defended almost like as it theorem. Is it ahead because privacy is just more important to folks or is it ahead because there is a perception that privacy is more important than fairness . There is some choice going on like on subterranean in that it got more attention earlier and can get solved faster over no choice at all and its just a my two short comments and then aaron can chime in. There are differences in how long these things have been studied but as i said when i was talking about fairness, i really think theres a technical difference. It just so happens that privacy is lucky in the sense that there is a well grounded theory general mathematical definition of privacy thats very satisfying in which Subsequent Research has shown you can do a lot with it. Meet that definition and still do lots of things that we want to do in terms of Data Analysis and the like. Fairness just isnt like that and its not a matter of time i think its not the spirit as i mentioned that said, heres the three properties election fairness that you cant simultaneously achieve, its not like my further work the theorem will be done, its a theorem. We not talk about this much in the book but i do think that privacy is lucky in the same sense that public key cryptography is lucky. It turned out sort of a nice parallel between development of publickey photography and the development of differential privacy where there is a period there is a cat and mouse game people would invent encryption schemes that sure looked random until they suddenly didnt look random to somebody. And then public photography in the 70s suddenly put the whole field on a much firmer algorithmic and definitional level and off to the races since then. Which doesnt mean that those things do everything that you want from security are that they are perfectly immense implemented every time but i dont think we are ever going to get there with fairness and thats just life. I think its hard to projected into the future. Privacy is about 15 years ahead of fairness in terms of its academic study. And thats for good reason. We had data sets for a long time so privacy violations have been going on for a long time where as when it comes to algorithmic fairness it only becomes relevant when you start using Machine Learning algorithms to make important decisions about people and thats only the last decade or so that both weve had enough data about individual people daily interactions with the internet to be able to make those decisions and learning algorithms have become sufficiently good that we can start to automate some of those as michael said, its clear already that theres not going to be one definition of fairness but i do think that if we try to look 15 years down the road, which is what you have to look before fairness is at least chronologically a study of privacy you might still hope for a mature science that is not can have one definition but perhaps who will have isolated a small number of precise definitions that correspond to different kinds of fairness and will more precisely understand how those necessarily trade off against one another in different circumstances. Its going to look different but im optimistic that there will be a lot a one other comment i would make at it i think i didnt appreciate this until we started working in algorithmic fairness a lot which i think another difference which will persist between privacy and fairness that doesnt have to do with maturity or technical aspects that discussions about fairness always become politicized very quickly. In principle everybody agrees that privacy is a good thing and everybody should have it. As soon as you Start Talking about fairness you immediately find yourself debating with people who want to talk about affirmative action or addressing past wrongs because all these definitions requires that you identify who you are worried being harmed and constitutes harm to the group and why think that constitutes harm. Some of the things we talked about like forbidding the use of race in lending were very much in the news the past couple years was harvard at all in College Admissions and like. These definitions also are requiring you to pick groups to protect and this always become politicized i think and regardless of what definition youre talking about and i dont think this will change in 15 years. Somehow privacy and fairness are different just in the social or cultural sense as well. There expecting the algorithms to do something that society has to sort out itself. Or conversely they dont think algorithms should play any role whatsoever not only in deciding those things but in even mediating them or enforcing them or the like. We take points in the book to point out racism was not invented with the advent of algorithms and computers it was around before you can talk about it more precisely now. He could have problems of fairness that are bigger scale but you can also have solutions a bigger scale as well know that things are automated. Question in the room . Show of hands . Anyone . Theres a question in the back. Thank you so much for the talk we really enjoy reading the book. I have a question about the difference between different a ain the book it talks about google and apple being collecting user statistics the type of data they collected is actually not the type of data they used to collect. Its like a new era of data acquisition. I wonder whats your comment about this tradeoff using differential privacy as like a shield of collecting user data especially i dont know how secure differential privacy is against adversarial attacks do you see a possibility that users like under the impression of differential privacy they are willing to give out more data but only to find the data customized in the end . Thats a good question. And its a question that relates to what you have to think about algorithms not just in isolation but in the game theory context. You are right that in both the way apple in the way google used differential privacy they didnt use it to further protections to data they already have available which turns out to be a hard sell to engineers. If they already have a Data Set Available adding privacy productions correspond to taking away some of their access giving access only to a noisier version of the data. Whats a much easier sell this is why how it worked in the first two deployments is to say, heres a data set you previously had no access to at all because of the privacy concerns heres a technology that can mitigate those privacy concerns and that will give you access to it. You are right one thing that happens when you introduce technology that allows you to make use of data while mitigating the harms you make more use of data which makes sense. So youre right that one of the effects of differential privacy is actually apple and google are collecting a little bit more data. On the other hand, they are using this extremely strong model of privacy which we talk about in the book the local model and what it really means is that theyre not actually collecting in the clear, your data at all. They are collecting some random signal from the data and the randomization is added on device to apple for example is never collecting your data its collecting only results of coin flips from your data. Although more data is being collected, differential privacy in the context is really offering an extremely strong guarantee of plausible deniability. And for that reason its not subject to data breaches for example. You might worry differential privacy causes companies to collect more data and ensure maybe thats okay they are using it subject for production of differential privacy with some hacker gets into the system and the data set is released, all of a sudden things are worse off, thats not how google and apple are using differential privacy, theyre doing it in a way that doesnt collect the data at all. At the census its different. They are collecting the data theyve always collected the data but they are actually adding these protections in a way that they didnt before. They are giving researchers in 2020 access to data thats actually more privacy preserving than it was in 2010. Theres lots of tradeoffs and these are interesting things but i think these are two different use cases that show different ways in which it can play out. To follow up on that for kind of a lay audience can you explain that coin flip to use the toy example in the book, suppose i want to conduct a survey of the residents of philadelphia i could ask of you cheating on your spouse write down the answer and at the end tabulate the results calculate statistics the average the holiday. I might not get the responses that i want because people might legitimately be worried about telling me this over the phone in particular they might not trust me they might worry someones going to break into my house and steal this list. They might worry in Divorce Proceedings it might be ab heres a way to carry out the same service. I call people up and say have you cheating on your spouse . Dont tell me just yet, i want you to first flip a coin. If the coin comes up heads, dont tell me if it comes up heads but if it comes up heads, tell me the truth if you cheated on your spouse. But that is okay because what i cared about wasnt pertaining to any particular person but statistical property of the population and it turns out because of the large one law of large numbers in aggregate i can figure out very precisely the population level average because i know the process it was added for echo in this is not so different from what actually is happening on your iphone right now it is reported more complicated statistics but in the end for text completion it is so sensitive that apple would like to know what is the most likely next word given what you have texted so far and they collect data that helps them do that. But putting it down into the data that is yes or no questions and thats not so different from this. And to put it into the context you still plays seven hours of digital on your phone but if all phones have a large positive or negative number then i look at any report that says i really wont know if he laid 30 hours and 13 is subtracted so they have the same plausible deniability and then the noise averages out for aggregate or average. I dont play this by the way. [laughter]. Thank you for coming talking about differential privacy but teaching about those commercials does that mean regulators are needed quick. Definitely but for instance google uses Machine Learning at massive scale for decades to do specific things like do a prediction in the more accurate they are the better targeted ads that translates into profit. So going on not discriminating against one addin privacy the way Machine Learning is deployed will reduce those accuracy rates and reduce profits at a know how to put numbers on that but we can be sure this will happen. And just relating to what we were discussing this is why the commercial deployments we have seen so far are in experimental areas that are not in the core business of these companies. They would like to know the statistics its not a core part of their business. But they stick their toe in the water and its to their credit but i am waiting for the first Big Tech Company that says were not just going to adopt these technologies around the edges all of them , we have many Excellent Companies that do research in this area is not like any Big Tech Companies dont know about differential privacy but there is a disconnect between the researchers who study these things and people with the pml business that they see looking at profits in the middle of the pipeline. I dont know how this will play out maybe there is some organic adoption by Tech Companies and others that are consumer facing in some way to voluntarily bite the bullet to say we will take a lead that there will need to be regulatory pressure and that will take time. It with those algorithms that they need to be theory based without that black box of knowing that. No but you do have to be careful is not that Machine Learning if you train a classifier and then to predict tours of some sort then you can legitimately you can attach value to it but but when you share the data sets when you take Machine Learning one oh one with the way you get there unlike in statistics not explicitly assuming that your data like training in a random forest that you have never seen before and it looks fine on paper and can you send me your data sets . Or even i myself follow the rules of fiscal hygiene i read your paper and everything im doing is a function of response to the findings that you wrote about which is a function of the data so if anything like that happens in those go entirely out the window. It would be very hard to solve the problem and then look at the data at all when conduct the experiments i will conduct and then to rule out interesting studies because it again the algorithmic science that allows you to share and reuse data in a way that does not give up. Just to follow up im betting when you said theoretical you were referring to causality . If there is a split between medicine and economics and show him some machine learners they say lets just get the data and we practice sound statistical techniques. Certainly having strong priors to have that causal model is something i would consider to have a strong prior can help reduce the number of things that you try on the data but i dont think it is a substitute for what we discussed in the chapter again it is a matter of discipline if i think i have a causal moment but usually there are parameters to that causal model and i will Start Playing around with the strength of that causality and then you go down the same rabbit hole to test many hypotheses on the same data set and you are prone to false discovery if you are not careful its very early but the disciplined algorithmic approach like differential privacy and other statistical methods are better than human beings themselves because i have strong fryers in the form of a causal model. To make the book is insightful but it points to the challenges. But then to explain that observation that supports good policy and we see that all the time so what is your thought even taking the next step to reach any instance of policymaking quick. Thats very important Computer Science has been unwittingly thrust into policymaking. If you are a Software Engineer at facebook you treat that parameter you are affecting all sorts of things for millions of people in many ways you are in formally making policy in ways that are not thought out. Given we are already in a situation to make this more explicit and clearer how we affect policy and therefore its important to try to understand at a high level what our algorithms and how do they work and what they try to accomplish in large part of what were trying to do in this book. Its funny the word Machine Learning and with that thoughtful dissertation at the time it was a very obscure area to be studying and it is interesting sometimes a joke that through no foresight or forbearance of my own it was delivered to the doorstep of science that was thrilling for a long time because it had no downside in many ways. There were interesting new jobs in science and in many ways of how we could pay it but the other part i dont like to use this but to become public intellectuals to become much more involved in use of technology and to help solve the problems created by those technologies and their still not a lot of that yet. And a lot of it is very superficial but you start to see Computer Scientist right those oped pieces so we will need technically trained people who are willing to spend their entire career mediating between the technical part of Computer Science and those technical implications and those that work on these types of problems is a Younger Generation people that are 11 willing to make that choice and maybe People Like Us rep points in our careers if we can do this and not worry if i could have a Research Career also it is very important and is starting to happen organically very early on. I thought i would ask something completely personal working on your research for quite some time there is a lot of yourself in their. I know its important work but what is it about this subject matter the 16 yearold version of yourself that makes this what you want to spend your day doing quick. The 16 and 18 yearold version of myself i started college as a math major to think mathematically about computation and then thats cool because here you are with the feelings of how people learn or how machines might learn. Then i got to grad school with a mathematical computational lens because it was just being defined and it was an exciting time and i was thinking of the theorems that you can think about privacy and more recently you can do it with fairness but now theres a difference between writing with the expert or academic audience you try to define ideas to be concise writing this book was different it was liberating to try to write to describe these ideas which are mathematical and i hope that they succeeded and to be lucky that of not just curiosity but meaningful and important questions but that excitement to do research in the field we can take them up to the frontiers of knowledge. My origin story is a little different woods with my 16 yearold self later i would write a general audience Nonfiction Book that would make a lot more sense to me than if you told me to be a professor of Computer Science and Machine Learning because it in high school i was a very different math student i didnt try very hard i was declared english major and to realize i chose english and it would teach me how to read but at the same time i managed to hang on by my fingernails long enough then i started to take more of them. And then there is that transition where they become more interesting and it was a creative ask that. And then to program a computer to do something that it takes ten seconds and then on the purely mathematical aspects. So in some vague way it does fulfill the type of thing i wanted to do when i was very very young. I remember may be six years ago a specific moment when we were sitting in a cafe talking about algorithmic fairness. It was interesting but flawed. You like to say six years ago i realized this was important for society and we as Machine Learning researchers have a responsibility to fix the problems. I would like to say even if the research turned out to be boring. Or theres nothing you can do the solutions are straightforward and then just convince people to adopt them i would like to claim come hell or high water i say this is that we have to do as responsible citizens. Luckily i dont have to know what choice i would have made because it did turn out to be a mathematically different but it is great and the positive and negative ways where there is interesting in and Technical Work that is so simpatico it has been a great deal of fun. And then to hear your ideas and share them directly with us. Congratulations on the book. [applause] the author of 17 books this is his latest the impeachment diary what was washington dc like in the summer of 1973 quick. Almost as exciting as it is right now. I was very lucky to be both as a live it as a witness to the impeachment proceedings we didnt know how it would turn out or who would step forward whether it is eloquence or analogy but that is what we will see of two steps forward and if it is a proceeding that is moving the way that it was in 1974 or if it would be boring to make everybody annoyed and mad. In your book you talk about the summer of 73 or summer of 74 in that year give us the understanding of the timeline and what changed. There was a wonderful Senate Select Committee Led by senator irvin and extraordinary things happened and there was a remarkable testimony so a great deal of evidence was gathered that summer but the following summer was then reduced to articles of impeachment and that final debate and it will happen that way in the next two months. Do you feel the history as encapsulated with the watergate episode cracks that people would be surprised at the length of the impeachment process . Than nixon process is a template for the current situation. Man the Nixon Administration is very clear both sides could make their case. Had there not been a smoking gun back in those days it is unclear if he would be impeached. Its likely donald trump would be impeached so that the argument that was constructed back in 1974 but you want the process to be fair and to feel good about the process when its done regardless of the outcome. What was the publics perception of the nixon impeachment quick. The Public Perception less than 20 percent were for his impeachment much less removal as the process started and as it went forward with hearings at the time of the actual debate it was closer to 50 percent. And then to be voted for the judiciary committee. And with partisanship and where we are now talk about impeachment republicans are always talking about this not being a fair process. There was no nonsensical thing about a scammer which went but to argue with those charges against him that i am completely exempt to go forth for people impeachment. The senators will be faced with a moral choice which is either overwhelming evidence of corruption and it was to have the courage to step forward in nixons case. Writing 17 books and has appeared on the tv many times here is the cover of his latest book. Thank you so much