Dr. Cassie Kozyrkov, Chief Decision Scientist, Google, "AI IS DECISION MAKING AT SCALE"

[Music] I'm really excited about our first keynote and it's my pleasure to introduce dr. Cassie kozakov so Cassie is an extremely unique person she has four degrees economics psychology neuroscience and statistics she is the chief decision scientist at Google when she advises leadership on decision process AI strategy and building data-driven organizations and what's really interesting is that she straddles these multiple worlds and perhaps due to this unique background she has created a really wonderful opening in the field of AI by coining the term decision intelligence and which is really a way of broadening the lens and creating a space for human decision-making in the world of machine learning in the world of AI so it's created a broad opening where decision science can make a contribution to AI so this is a very big deal and I want to also say on a personal note that when Cassie gives talks she usually addresses 5,000 people 10,000 people at a time so for her to come to our conference she's doing this entirely because she feels she belongs here she feels like this is her community and we're gonna make a dent in this topic together so we're so delighted and extremely grateful that you're giving us our opening keynote thank you welcome Cassie [Applause] so I am delighted and honored to be him in fact when my participation was announced for this conference there was an online comment that really resonated with me and it was welcome home Cassy and that is exactly how I feel about this community now to me I started with data when I was maybe eight years old and I discovered that Excel was awesome because you could like put numbers in there and then summarize them and that was fun so while the old other eight-year-olds were all playing outside I was fiddling with spreadsheets and so to me playing with data also feels like a fun almost childhood activity and then growing up a little bit I started to think about why is this data worth playing with is the data important for its own sake and my conclusion was no it isn't because what we think we know about the world doesn't really matter until we act it is through our actions that we impact reality and so while data are really pretty it is decisions that are important but to make good decisions you should use a useful information that's available and so to understand the decision-making you have to bring in the data stuff which is why I trained myself both in the decision sciences and in the data sciences but now I come back to you to tell you that I have been a spy for you all along because I my most recent graduate degree is in mathematical statistics and after that when I joined Google I joined as a data scientist and in fact funnily enough they advised me during my interview so I actually applied and was offered two different positions at Google one of them was in user experience research more on the psychology side of things and the other was in data science and before my data science interview someone suggested to me Cassie don't tell them about your other stuff with psychology and decisions just go straight data science just sell yourself as a statistician I said alright fine I'm gonna trade do that when I got that job I was joining as a research statistician and I had intended to for a while maybe leave or put on hold the whole decision stuff then I started looking around and seeing how data science was actually conducted without the decision component and I found myself slightly shocked things needed some fixing and needed some help which is what inspired me eventually to build our decision intelligence training program which I saw as an improvement of data science by bringing in the decision sciences so an augmentation with a decision perspective and now we've trained over 25,000 people at Google just at Google in this and I'm also beginning to blog and make videos out there for the rest of the world to to start joining us so that's a new initiative for me but when I explained decision intelligence to Googlers the way I explain it is this I say you can think of it as machine learning plus plus but really it is about reducing type 3 error now what is type 3 error those who have taken a stats class let's have a quick refresher of what you learned there type 1 error is when we incorrectly reject a null hypothesis type 2 error incorrectly not rejecting type three error incurred sorry correctly rejecting the wrong null hypothesis or if you like the Bayesian statement and a lot of you like the Bayesian stuff basal statement on the same is simply all the right now to solve entirely the wrong problem and we see this a lot in data science we see folks making really rigorous effortful calculations for no good reason an engineers building systems for you know a year two years and when pressed why did you do this what was it for they can't give you a good answer now this type three error thing tends to be mentioned in stat 101 and there are titters haha hehe solving the wrong problem and it's presented as an emblem of hopefully working on the right things helping take better actions with data and then it's promptly forgotten so there's some expectation that we can mention it once in class and then people go through a PhD and say statistics of computer science and then they go to work and then all of a sudden they transcend out of their math then start to look around and think about decisions and usefulness what is the plan there for people to know what to do and how to do it and to even want to transcend as we like to say in google's site reliability engineering division this is in fact their motto hope is not a strategy just hoping that people transcend out of their mathematical training and start to think about actions decisions usefulness that's not much of a plan there and that's what inspired me to want to start to train people to actually do this right now what is data science before we get any further a lot of people have a lot of different ways of describing data science some people joke that it's what statistics on a Mac or statistics for people who can write code I prefer to think of it this way it is the discipline of making data useful and if we see it that way then it would be an umbrella discipline for statistics machine learning /ai and of course analytics but how do we separate all these things and what's the difference between statistics and machine learning what's the difference between statistics and analytics let me try a few texts on Amazon but the first one if you are a practicing data scientist I hope you find this as awful as I do some people try to split these disciplines based on the tools absolutely horrible but let's try a different common one based on the math used what you should know is that if you are a sufficiently bloody-minded data scientist with a point to prove you can use any of the algorithms and equations to any of these purposes so this is also not a good division so instead the division that I propose and that I teach when I teach data science is this one none many few not many few what any guesses see in this audience exactly in this audience you get it not in many few decisions but I remember once giving this talk to a data science community and someone blurted out data points to which I have to say when you have no data descriptive analytics is a little bit tricky to do so yes it is indeed about the decisions that we want to make under uncertainty analytics would be about getting inspired we don't even know what decisions we want to make so we want to gather inspiration and if that is the way that we're thinking about things then the excellence of an analyst is speed how quickly do we run around in all these data that we know nothing about to produce inspiration but also as this community knows because all the way that humans are really good at seeing Elvis's face in a piece of toast and other good pattern finding and because we're really good at confirmation bias and seeing what we want to see analytics cannot be taken as a serious and rigorous source for data-driven decisions because they're the decision is not specified in advance now these other two it's about how many decisions you want to make if it's very few or just one important decision under uncertainty that is statistical inference and again even statisticians sometimes don't realize that what our field is about it is about changing your mind if you are Bayesian it is about looking at a formally stated opinion and changing your mind about that opinion with data if you are a frequentist you need to have the notion of a default action and actually you are committed to that forms your null hypothesis and then you ask do the data make me feel ridiculous about my null hypothesis and therefore switch me to the alternative action and all this business about learning things about the universe it doesn't quite work nicely in the frequentist framework but the action selection does so again decision-making doesn't matter this machine learning at everyone it is about a recipe for making repeated decisions and of course with uncertainty because if there were no uncertainty we could just look at the answer up in a database so this is my breakdown so of course this is very decision oriented and what you also realize here since you can use the same tool for any of the three and you can use the same equation for any of the three the way that you should go about working with your data depends very much on your decision framework so the person who kicks off the analysis the decision maker the decision professional is the factor that determines everything else about your data science project so what if we don't have any training decision-makers then you have disasters I get asked this question a lot a lot lot and I also and I'm when I'm talking to leaders of large companies whose names will remain unmentioned but that you would recognize they say there's lovely sentence to me then it is our data scientists are useless well why are they useless because what you do is you hire a bunch of PhD nerds and you put them in a room and you tell them to go do stuff and what they're gonna do is what they're most comfortable with so I propose a cheeky hypothesis for you perhaps people who stay in academia for a really long time to get say a PhD are really susceptible to the rewards that you get in academia and what are those rewards especially for the mathematically oriented you get a pat on the head a gold star every time you use the most complicated thing in your arsenal not do the most useful thing not solve the most important problem but where you show off your methodological complexity so you take a bunch of people who are very excited very reinforced by that action to the word cycle you bring them into industry and then you just expect them to what contribute fix your business because you think that data are magic and you pronounce the data with a capital D oh dear so instead what I tell people is this let everyone in your organization engage in analytics and treat that as mere inspiration look at data in a in a shallow and quick way don't take it too seriously don't be overly methodological complex and then when it is a decision-maker who looks at that inspiration and says I actually might want to make a decision based on soar I might want to automate something based on this then you go and you do it rigorously and you engage in statistical inference with machine learning and AI so the principle here and what I'm teaching all these folks and when I help these fortune 500 companies with their strategy but I'm also teaching them in analytics until a decision maker says otherwise until a trained decision professional steps in and actually specifies the framing of the decision that's going to require AI or statistics now some pushback that I get at that point is that's all good and well Cassie where do we find train decision professionals ok that's also my questions my best attempt at that is helping to train folks and to democratize what we know about as far as I can and really cheerlead for decision intelligence and decision making decision science but what I think is really going to be important is for all of us to step up and understand what role we and our students and our colleagues should play in this a a future so let me take you through then the my view of machine learning the stuff that you learn in grad school in a stats PhD or a artificial intelligence PhD is like this little red book in the middle of an end-to-end process if you're going to apply data science and so you learn this piece here on the front are things like framing the decision in the first place thinking through usability thinking through populations of interests all that kind of stuff you don't actually necessarily learn metric design even the a a ph.d program and then on this side the engineering to get it to work in production and then things around safety nets and reliability big into in thing and yet what is taught in most of the data science programs it's just this little piece so I think that we really need to take an end-to-end view and if there is one community that is most qualified to think about this stuff from that big-picture perspective I think it as I say this wrong so machine learning here's my view of that I'm cheeky I go on stage in front of as you said 5,000 people I'll say this buzzword that you're also excited about what is it actually it's the thing labeling it's not about building humanoid robot type entities it is about automating repeated decisions decisions like is this photo a cat yes or no or how much should we bid on the next auction and of course it can be used for analytics same algorithm right different purposes where it merely surfaces some inspiration a lot of the unsupervised learning applications are exactly this kind of thing put like with like and see if we're inspired clustering or anomaly detection something is unusual in the data are we inspired but that's just analytics sometimes quite often it is supposed to automate a business process and it does result in the irrevocable allocation of resources for example if I want to serve a search result I am fighting for limited user attention once I've put that search result in place that's what the user saw that will also influence whether the user will come back to my search engine in the future so the the most interesting and the most careful applications and the most common applications of machine learning and AI are ones that can be framed as decision making now a quick note about machine learning versus traditional programming what's the difference well we need to get from information to an answer via a recipe or model model is dis fancy word for recipe and the way that it's done with traditional programming is the programmer thinks really hard about the problem communicates with the universe and then handcrafts that recipe of that model to take incoming data and output answers decisions in a machine learning them we don't handcraft those recipes anymore now the programmers effort shifts to putting examples into the system and it is the machine learning algorithm that does this to check am i sometimes it is thought of as a superset to machine learning but the way the term is really used today is as a synonym for deep learning so when people talk about AI this AI that they're really referring to deep learning and deep learning is for solving tasks like this one cat not cat you just took in some really complicated information through your senses and you got the label as if by magic how do you know that that's a cat can you tell me exactly what you did with each pixel - I'll put that later no our brains have had eons of evolution we can just do this we don't even know how we do and so if we think about creating a stitching and crafting that recipe or model and what's gonna be in that recipe look for oval type things maybe triangles well will it handle that situation unlikely and yet for you no sweat you can see that that's a cat instead of trying to handcraft where you pretty much have no hope of succeeding why not instead use examples and an algorithm that can pull out a pattern of what to do with those pixels and automate this task for you so AI people are really excited about today is about solving those complicated tasks that we cannot depend croute solutions for but actually it turns out that the approach is the same for whether you're doing an AI task or a machine learning task so you don't even need to worry about the distinction which is why I lumped them together and I talked about machine the essence of it is about explaining with examples instead of constructions that is what machine learning boils down to and we're really excited about it because we can now solve completely new classes of tasks because we have a second way to talk to computers we used to only be able to talk with explicit instructions now we can speak with examples and so that's the reason to be excited some tasks we can't come up with those instructions they are ineffable and AI is about automating the inner football it lets us do more than we could before but hang on we're talking about really complicated tasks with really complicated recipes are we expecting that we can just open up those recipes and read what's going on in them and know how it's working not really it'll be thousands hundreds of thousands of lines of gibberish that our human minds can't wrap themselves around so how might we trust a system like this well I propose that testing might be a better basis for trust and I suggest to engineers to think about this if you are planning a trip in a spaceship and you have the choice between a spaceship where you have an exact stack of physics descriptions about how this thing purportedly flies but no one has ever flown in it or another spaceship no one knows how the thing flies but it has completed a lot of successful flights like the one that you're going on never has had an accident which spaceship will you choose I don't know maybe it's the statistician in me but I strongly prefer the one that has been tested so I suggest testing as a basis for trust testing on new data of course overfitting what they they call overfitting in machine learning and AI simply refers to memorizing the exam problems before taking the exam how do you make sure that your student is not memorizing their way to a perfect score test them on things they haven't seen before test them on new data and this is a statistical test and this is what keeps you safe it makes sure that your system works so that's the this setup that AI folks understand up to here let's take a more complete look if you over focus only on the decision outputs those little decisions that the system makes which lend themselves quite a lot to decision analysis also you might miss some bigger decisions going on so I'd like us to start thinking not only about AI for making decisions but also think decisions for making AI this picture gets more involved there are two phases in an AI project the prototype phase where we are supposed to train and test a toy thing and then the production phase where we're going to modify that code so that we can deploy this thing safely and reliably in production and we continue to monitor it and check it and maintain it to deal with it changing the world analytics has the to play in both of these analytics is all about iterating wisely in the prototype phase we have a lot of potential inputs we can't try them all which ones do we start with which models seem to be promising how do we do now - for a meter that kind of stuff lends itself really well to analytics on the production side it's about monitoring does this thing still work are there any anomalies surprises things happening that we might need to react to now unfortunately a lot of companies fail horribly at analytics because they miss a really important point if we don't know what we're looking for in advance then we cannot tell the analysts to go find a specific thing and make their mandate they're looking on something looking for something specific that makes sense right so what you have to do is give analysts an allotment of time for exploration it is a time investment now how do you think about the decision to invest some amount of time in exploration well turns out that most business leaders have no idea how to think about that you know who does people in this room what about statistics statistics in both cases boils down to does it work cuz it's all good and well that you tried to build some machine learning system but you know if you want it to work you're going to have to test that it actually does work both in prototype and that it continues to work when it is out there in production so we've said testing keeps you safe and to some extent to the data science community gets this and there's some data science groups it's still in their infancy that confuse all three areas of data science and they think that analytics is actually statistics and all kinds of other horrible things and they don't split their data delight but it's an extent that the data science community is mature and I gotta get this thing you have a pristine allocated test dataset you check that your system passes that exam before you release it on to the world now here's the part that a lot of them haven't thought about who sets the exam the only guarantee once the system passes testing on new data is that it works according to whatever the exam was who sets the exam who decides on what data on what textbook problems the student must pass who decides at what quality that exam must be taken how many questions must the student ace who decides the metric even if it's cat not cats we've got lots of different options from precision to recall to accuracy to s measures to all kinds of things who picks that performance metric who's qualified to do that you know what happens out there today like that just they make it up after the fact in fact even worse show you an order of operations thing in both of these phases we should be asking what it means to work that's for a decision professionals to do and therefore decision points major ones on lots of other little should we attempt it in the first place and how do we think about scoping machine learning product should we production eyes it when it passes testing in that first look it's not set to run live yet and it has not been tried on live data and it might influence later on the world that it exists in so the decision to actually invest engineering resources to building it is a separate decision from should we launch it should we release it at a hundred percent to all of our users and if the world changes when the data changes out from under us should we keep it how do we think about these four decisions and what I'm not seeing anybody talking about except insofar as this community and my own work how do we be formal about these questions I want to also tell you what what goes on there in my experience in applied data science in industry what order do you think people do this in the order of their comfort zone again if you have a PhD in algorithms design that formal AI research PhD you think algorithms first so you're like neural networks they're cool where can I find some data to apply to these neural networks and then as an afterthought we're in the business mind I shove this system you can see there might be some dangers there if you're an analyst and you are first and foremost in love with data you think first about the inputs what data do we have then what algorithms might be cool and then again as an afterthought what do we actually serving but you you know how to do it better you know that where we start is with the outputs don't worry about how the system works in its guts what are we serving what are the right sort of outputs and if mistakes happen which mistakes are how much worse than which other mistakes and how do we score the whole thing and how do we set the phone's criteria that's this community and I'm not seeing a lot of this out there in the world and this frankly scares me now some stuff that we know here we need to set performance criteria in advance out there quite often you built the thing first and then afterwards you see whether you might feel like launching it statistically significantly now what we over here know is that sunk cost fallacy is a thing right and humans after pouring three years of love into a system will begin to bargain with themselves and say oh but the performance is not so bad I mean before you know it you have some poisonous rubbish running around out there released so you have to do this in advance setting performance criteria upfront that also means that you the decision professional a trained decision professional upfront early on in the project to kick this whole thing off and when you create your metrics you need to think about creating metrics which if they were passed in testing would mean that you actually want this system to exist so that it is the your metric means that both the spirit and the letter of your wish are fulfilled by a system this is one that again the data science community is not very good at but I think this one knows a lot about and then I'm going to leave you with this thought in the last two minutes which of these is better a reliable worker or an unreliable worker any thoughts if you're saying reliable worker definitely hundred percent I'd say careful it really depends and it depends on the quality of the decision-maker because a reliable worker will scale up an intelligent decision makers wishes but if the decision maker is unskilled and foolish then it will be like those coffee shop science Kaffee and making you do stupid things faster if we have a bad decision maker we don't want reliable workers we want diversely unreliable workers in humans who will go and do their own thing all kinds of different ways and they will mitigate the effects of this untrained decision-maker now here's the scary thing and here's pretty much why I do what I do I realized this and then I was terrified people are scared of AI for all the wrong reasons the reason you should be scared is that technology is a lever for whatever for the decision-maker the human who is deploying that technology AI is the ultimate reliable worker if it passes testing it will do exactly what you asked it to do and it will do it at scale we are amplifying with big data technologies we're amplifying the human decision maker we're allowing them to have massive outsized impact on the world and sometimes that decision maker doesn't even realize that they are wielding a lever this big they think they've got a team of magical data scientists and everything will be okay whereas actually this is one human amplified hugely able to impact everything really quickly in the past that decisions were self-limiting it was only Kings or generals who could have that much impact on the world where is the rest whether you made the decisions well or poorly you would be okay but not today not with these tools that impact is brewing far too much and yet we are not training decision makers which is why I want to appeal to this community to realize just how you are as the starting point step 1 of all of this stuff because the world is headed towards Big Data technologies we will have an AI future an ICA I as the proliferation if you will of magic lamps it's never the genie that is dangerous it is an unskilled wisher the wisher who does not think through what they are actually asking for and there are a lot of myths in humanity's stories around silly wishes that then have dangerous consequences bad decision-making and the fight to come back to bite you this is the future that we are moving in we need to make sure that we have trained a new breed of worker the decision professional who can work at the first stage and throughout the entire project so I hope that you are excited to feel like you should come into AI and make that community recognize all the wisdom that you have been gathering over the past 25 years because hope is not a strategy just hoping that humans are going to learn how to make decisions wisely work with these big technologies that's not a good plan so let's be more thoughtful about this let's let's all work towards a brighter future good decisions amplified by these technologies and if you know of any decision makers working in AI starting projects I have a blog post that is an 18 minute guide for things to think about step-by-step as those people are starting out so that they don't hurt themselves so that can be found at this link bitly quest ito which is my twitter handle underscore DM or decision maker guide this stuff is really important and I'm really excited to be home with this community I think about tomorrow's AI will be built on what we we hear dude thank you in this community we rely a lot on uncertainty and uncertainty is not to be found in any database how do you reconcile the work of AI which deals with data which is past and present with thoughts about the future which is the realm of uncertainty in which a lot of us spend significant chunks of our time interesting question statement the idea that uncertainty is not found in any database when you really don't have any uncertainty you do not need AI because what do you do you look the answer up that's what you do when you for example need to automate something like output the correct social security number that goes with this person's employee ID what do you need pattern finding for just look up the answer the requirements to find patterns and make recipes out of them so that we can deal correctly with future data is fundamentally about uncertainty we only need AI if we have uncertainty but now here's the thing we've got all the data that we are using in the prototype phase is essentially a sample B population is the data set in which we're going to need to serve this system so we are in some hardcore statistical thinking here we as the decision-maker have to think about designing that exam in terms of what is the population what is that data set the full one that I expect to serve my system in how well does my sample the data that I have access to in my database actually cover where I need to serve my sister what assumptions am I willing to make against it's not magical magic it's about taking assumptions plus data and coming out to a reasonable course of action so what assumptions am I going to be making and if I'm wrong about these assumptions it's some sensitivity analysis stuff how bad could it go fundamentally uncertainty ism and what we think about in this community in terms of uncertainty is definitely also coming in and I'll just mention that Cassie wanted to actually talk to people one-on-one so please feel free to ask her questions in the break I will just ask you the top water question here in the room how do you reconcile ethics and artificial intelligence and where do you inject it there do you inject it inject ethics so what's that nice way that Ron Howard puts it decisions are immoral not immoral the epochs must come from us and the when I say us I am referring to the decision-maker for the project just because you can build something doesn't mean you should build something and in fact if you don't think about what you're building and why you're building it you are most likely to end up building something that you shouldn't but when you are deciding whether this system should exist and why it should exist that is the place where you are injecting the effects and when you are thinking about which populations and you should think about three types of populations by the way the data population the intended user population and the entire potential user population the people who are affected but may not be direct users of your system when you are thinking about who those populations are and how you're willing to trade off performance across various aspects of those populations you are injecting ethics in there when you say something like oh well our system right now we only have data about the US our system is going to launch originally only in the US so we have a decent sample for that population but later we might expand globally you're also making the statement that you are quite happy to then relatively fail on the global scene because you have preferred information you've limited yourself a biased your system with respect to a global population whereas you might be unbiased with respect to the local population all of that is down to the decision-maker all these little assumptions and settings they are themselves subjective and are the injection of ethics thank you so much Cassie um I know it's not just me you should have guessed this by now she's keeping tabs on our community she knows a lot about what Ron Howard has said and and and not only that she has she's gonna bleed to panel discussions tomorrow in the high-tech track and she has she's co-chairing the track so she's not just here as a keynote speaker she is a part of our story so please talk to her in the break with your questions and also feel free to use the cards to talk to each other about data science so with that we will reconvene at 10:05 yeah so 10:05 we'll see you back [Applause]

Channel: Society of Decision Professionals

Views: 13,230

Keywords:

Id: bCjMhZZYlP4

Length: 43min 1sec (2581 seconds)

Published: Tue Mar 12 2019