 It's really fun for me to follow a talk about, you know, essentially blockchain technologies and distributed ledgers because I'm talking about something entirely different. So we'll have a bit of intellectual whiplash this morning. So just so you know, my perspective on the world is that I am a computer scientist. I've worked in machine learning and data science for a very long time. And most recently have been working in the context of the enterprise, but not only in financial services. So financial services is the largest cluster of companies that we do work with. Certainly there are many interesting use cases there, and I'll touch on several of those as I go through the talk, but I like to maintain a pretty broad perspective on the world and look at, you know, use cases for machine learning and AI technology that cross everything from sort of consumer applications through retail, lots of pharmaceutical healthcare, insurance, financial services, wherever else we may go. And so I wanted to say good morning. I hope that you will walk out of my talk with an understanding of the current state of what is actually practical with AI and machine learning and a few new ideas and also a process for thinking through those ideas and doing your own technical discovery and diligence in the space. And so that's my goal for the next 20 minutes. I have a lot of material, and so I'm going to talk through it really fast, but if you have any questions, you can find me later. Or you can find me on Twitter at hmesin or hillary at cloudera.com with 1L. We are living in the future. It is just unevenly distributed. If you want to see hilarious stock photos, look for stock photos of people using VR headsets because you'll see them like, you know, in the pool with their VR headsets on. You'll see things like this. When we look at the media coverage of this technology, we see a lot of stuff that's fairly nonsensical. So things like, you know, okay, the future of corporate growth is fueled by AI. That's pretty recent. Maybe that makes sense. Teenager teaches an AI to rap like Kanye West, okay? That's pretty cool. But what we see happening is an outstanding amount of hype and this anthropomorphization of what the technology can actually provide for us. So if you take one thing away from my discussion here, it's really that we're talking about computer programs that are built on top of data that improve with the introduction of more data into those systems and feedback loops. We are not talking about some actual recreation of human intelligence or some kind of, you know, science fiction type thing, though I do love science fiction. Now, I like to put it in a framework because it feels like we have AI coming out of nowhere over the last year, but this is actually not the case. AI has existed as its own academic field of research for, you know, about, since the mid-1950s was really the first wave of AI research. But if we were here having this conference 10 years ago, we would be talking about this part here. We'd be talking about big data. And what that meant at the time was that we now had open source and available infrastructure that allowed us to get all of our data in one place and to query that data without building our own unique proprietary platform and without, you know, doing a lot of custom work. That was a revolution, but it was a revolution in cost reduction, which is not nearly as exciting to talk about as smart robots. But it was one where now all of a sudden this technology that people have particularly in finance been applying for 30 years to only very high-value problems could now be applied to trivial problems. For example, I was the chief scientist at Bitly, which really is those little short URLs on social media, and we were able to study questions like, do people click on more photos of cats or dogs on social media, involving things like 10 billion documents, which was something 10 years ago that was really nuts. You never would have invested that kind of effort in understanding cats versus dogs. But that was 10 years ago, and once you can count things in your data, you can do analytics that is counting things for some business purpose. Right? Okay. So once you can do that, you can count things cleverly, and that gave us data science, which is not just counting things in historical data, but building predictive models, building ideas and representations off that data that can teach us something that's perhaps non-obvious or be used to predict even the future or infer missing information. If you've ever used the Zillow.com, what is this property worth? That's what we're talking about, right? And now we have machine learning where we're counting cleverly, and then we have these feedback loops, such that our systems continue to improve over time with the introduction of more data. And so you see how these things are related. Machine learning provides a set of techniques that you can put under that broad umbrella of data science. And so today we have AI, which has returned from a terminology point of view because of the rise of deep learning, which is a subset of techniques used in machine learning that are based around neural networks that have given us not just more efficient and cost-worthy capabilities, but have actually given us the ability to do some things we couldn't do at all five years ago. And I'll get into more of that later. But we're talking about things like solving image object recognition problems, where you can take a photo and it'll say, oh, there are 20 human beings in this photo, one camera, and one clock, right? So all of this creates a technical foundation on which we can start to imagine the future. That's the tomorrow part of this talk. But before we really get into that, I want to talk about what this looks like when it's successful. And my favorite way to frame AI and machine learning today is to say the work we had to do five years ago was to make it exciting, to get anyone to care so that when you go to a cocktail party or networking event and you say, I work in machine learning, people don't just turn their back and walk away and say, oh, maybe the blockchain guy will be more interesting to talk to. This is about making it boring. And what I mean by that is that we have to let the technology fade into the background. So this is my favorite machine learning application. I don't work for Google, I never have. This is the Google Maps traffic view. The most remarkable thing about this is that it is really boring in that you can glance at it, you make a decision, you put it away, you don't think about it. You do not have to know that there is an incredible amount of data analysis machine learning that goes on to produce this visualization for you. They're getting real-time data from people's phones. They're integrating it with public data sets. They're making predictions off of historical data. They're visualizing it. They're streaming it to you over a mobile network onto a device so you can be in your car. And by the way, I live here in New York, so I don't drive, and I think this is nuts that people are driving cars, looking at their phones and making a decision about where they're going to go. And you do that, and you can do it so well with this application that even if you know how to get where you're going, you're still gonna load it up because you want that real-time contextual information to make a better decision. You can do that with no understanding of the technology. This is one of my favorite AI applications in the world. However, this is really hard, and one of the things we do in our practice of machine learning is try to understand not just the possibilities, but the boundaries of how to apply techniques. And so I'm going to share with you a personal example of being a machine learning edge case. So this story begins with my name, which is a little weird. So it's Hillary with one L, Mason, right? It's not a very uncommon name. It's not a very common name. I happen to share this name with a British character actress. It's quite a bit older than me. I was raised away in about 2005 at a very old age. She had a very successful career. Now why this matters is that because in her later career, she played roles like witch, ugly hag. And this was a search engine from 2009 where they had the innovation of combining photos with text results. And of course, here is my bio at the time. I was a professor. That is her in the role of ugly hag. And the implication here is obvious, right? And this named entity disambiguation problem is still a problem for us in machine learning in every domain. And so you're going to say, Hillary, you're making fun of some startup from 2009 and nobody cares. So let me make fun of somebody a little bit more expert at this, which is Microsoft Bing search engine. The Bing folks, by the way, do not get the credit they deserve for building a brilliant search engine. But once again, named entity disambiguation, that is my photo. I think you can see the resemblance here. That is not my date of birth. And that is not my date of death. And it's also not my middle name. But this is a hard problem, right? This entity disambiguation problem, the idea that there could be two not at all famous but well-known enough, Hillary's in the world is just blowing Microsoft's mind unless we just make fun of Microsoft. My friends at Google, if you searched, not today, because I checked, but if you search for this movie, Robot Jocks, which is an awesome 1991 movie about giant robots punching each other. So a fine piece of cinema. There I was. This was a few years ago. And I made fun of this on Twitter and Google Plus at the time, which was the best way to get in touch with Google employees or friends who work at Google. And then, of course, they fixed it right away. And it was just funny. But then about a year ago, I thought, hi, I wonder what's going on there. And there I was back again in Robot Jocks. Yeah. I don't even know where that photo came from. And so I share this example because not to pick on anyone here, especially who works at Google or Microsoft, I think you're all wonderful and you build amazing things. But rather to say that this technology sends this amount of potential to make our lives more efficient to build new products. But it also has limitations. And when we have conferences like this, we tend to talk about the potential, but not about the limitations and not about where things tend to go a bit wrong. And I have a bit of a twisted sense of humor. So that's usually what I want to hear about. Like, where are we using this that we need to be a little bit more thoughtful? And so we're still waiting today for the way we... And maybe this problem will never be solved. But we are waiting for the way that we broadly think about the use of the technology to catch up to what's real. And in order to accomplish that, we have to make it boring. We have to say, AI is not something that we're excited about. AI is just one tool. It's just as exciting as your C compiler. I also want to mention here in this audience that the biggest opportunities will be surprising to you. So a lot of people think this innovation, specifically in machine learning, happens only in academia or perhaps in startups. They certainly do good work, but I can tell you from experience that academics are generally focused not specifically on work that will help you build production systems that solve your problems, but rather on ideas that are novel, on meeting fairly arbitrary benchmarks that will get their papers published. You should fund the heck out of them, but don't expect them to solve your problem in production in a scalable and repeatable way. Startups, likewise, are highly resource constrained. So they don't have people. Sometimes they don't even have domain expertise in the parts of the industry they want to impact. They don't have data, and they haven't been doing this for a very long time. And so we have the enterprise. We have large companies operating complex businesses, huge amounts of human and technical expertise on where the ROI would be in that domain. Huge amounts of data generally created as a side effect of operating those businesses for some time. And these are the people who are looking to those academics and startups to solve their problems. And so I want to point out for everyone in this room, you can do this. You are in fact the best positioned people to do this. And in financial services, we see a ton of interesting use cases. So I have a rule of thumb to work with companies which is generally we have to find some clear ROI on a cost savings or process improvement basis using machine learning. Lots of people in FinTech especially want to start in security, anti-money laundering and fraud detection. These are really fruitful areas because a small percentage improvement is very high impact. There are other areas as well, things like understanding your customers, turn analysis, some of the marketing techniques, which are pretty easy to get started in. But if you only think about those ROI, the ROI in the terms of cost reduction, you put a boundary on the amount of potential your use of AI will have. Think also about new revenue opportunities, new growth opportunities that can come out of the same technologies. That's where the real potential is. So I have to go fast, so I'm going to do this. I want to give you our framework for looking ahead where the technology is going and then share some specific things we've come up with. But I really encourage you to take this home and give it a try and see what you come up with. So this is how we look six months to two years ahead in machine learning for the enterprise. And we do have a research product around this but we publish a lot in the open and also publish a fair bit of open source. First thing is to drink coffee, have ideas. There are a lot of companies, when I see their list of projects, they're always good ideas. I get very worried because you are missing out on a huge amount of opportunity that would likely look like bad ideas on the surface. That's why it's really big. Validate against robust criteria. So step one is create a very broad sweep, get as many ideas, potential projects as possible and then go through and validate capabilities. So is there research activity if they're working in one domain you can transfer to another domain? Has somebody done something in another industry that you can use or in an academic context that you can use? Is there a meaningful change in economics that would enable you to actually implement something at scale? Meaning that we knew how to use deep learning effectively, for example, for a few years before it became very common. Even now it's not very common. But we couldn't afford it because GPUs were too expensive and so I could draw a graph like this for pretty much any of the constraining financial factors on machine learning implementations, cost of storage, cost of compute, cost of bandwidth, all of that. And it looks like this. This is my favorite example, though. This is the same size chip from 2005 and 2014 cost the same price, 128 megs, 128 gigs. So if you can't afford it today, just wait. Is it becoming commoditized in open source? That means that you have robust software and infrastructure you can build on without having to own and create it yourself, being the classic example in big data, but I don't need to tell you all about that in this room. And then is data available, either data inside of your company or that's proprietary, data you can purchase, data you can partner for, or data you can get from the big, wide open world of data, either from sensors or from Wikipedia, which is the dirty secret of any every natural language startup. There are in fact errors in Wikipedia that you can actually track through to the train models that these startups are selling. It's pretty fun. And then progressively explore the risky capabilities. That means, you know, have a phased investment plan. In machine learning, we do this in three phases. So we do validation and exploration. Does the data exist? Can I build a very simple model that's better than random in a week? Then I do algorithmic excellence. What is the best possible way to solve this? Then we do operationalization and scale. At each one, you have a cost gate to make sure you're not investing in things that aren't ready and to make sure that your people are, you know, happy making progress and not going down little rabbit holes that are technically interesting, but ultimately not tied to the application. You can tell I've managed technical people for a long time. Predicting the future is really hard. So this is a set of postcards from 1900 in France and they're trying to predict what the year 2000 will be like. And I love them because they actually sort of got the problem statements, but the actual implementations were a little weird. So like, these are firefighters with wings. And this is a school where they're sort of wiring things into people's brains because they hadn't thought about wireless. And this is a farm where they've got a robot. And this one actually kind of exists, though the control panels don't really look like that, of course. But predicting the future is really hard. So we write these reports on different technologies. They are designed to be six months to two years ahead of what you would put in production, which means practically we're the ones reading 30 papers and then writing some code that'll actually work in the enterprise context and also is aware of all of the security governance and other fund requirements that many of us are subject to. I'm selected a few to tell you about because these are the things I'm most excited about. So one is natural language generation or the ability to take structured data and output narrative. If you hear about this in the news, it's about reporters freaking out that their jobs are getting replaced. We have a prototype here that writes real estate advertisements, so that one's pretty fun. The real impact here is not in replacing reporters. It is in making structured data accessible to a wide audience without requiring sophisticated parsing of things like graphs or raw data. We have several use cases that have come out of this work, everything from a celebrity fashion magazine that is able to take the structured data of what Kim Kardashian wears and automatically generate the reporting around that, all the way to a very large financial services company that is supporting their customers with portfolio report emails that are customized to them, increasing engagement that they also tend to be retirees, so there's a lot of value there in language for that audience. The third use case that I get excited about is actually one of our bank customers that is now writing about 80% of their compliance reports using some of this technology. Believe me, those people were really happy not to have to write the same thing again and again. I have to talk about deep learning in any talk that has AI in the title. This is fundamentally a new capability, and it is based on something that is inspired by the way we thought the brain worked 60 to 70 years ago. It is not itself a brain. This is the fundamental element of a neural network. It's a neuron. You have inputs that are ones and zeros. You have weights, sums, and biases, which are really in the simplest formulation, just adding it all up. And then you have an activation function, which in this formulation, again, is just a rule. And you have layers and networks of these. And if you want to know why it's deep learning, it's because you have more than one layer. Deep is not itself a technical term. In fact, there's some deep learning that's just one layer anyway. But this is the fundamental unit that we're talking about here. What's exciting is that these networks do not require you to do manual feature engineering. You do manual network design, and then the network does the feature engineering, which allows us to apply it to problems where humans were really bad at the feature design, like recognizing objects and images. This is another one where I'm going to show you something that went wrong. So we created this thing that recognizes objects and Instagram photos, and you can add your Instagram photos to it. Bleak or burger, because I love cheeseburgers. You can see that there's one photo of a flag and then everything else is a photo of burgers. So it nailed hamburgers. Dishes is where it thinks it's a food, but it's not entirely sure what food. Those are mostly cut in half. It thinks some are hot dogs or meatloaf. Lots of the American culinary stuff. Crabs is worth looking at, though. Because as a human, you look at this and you slowly realize, okay, this neural network has learned that anything with French fries near water or a dock is a crab. This is kind of a problem. Again, I'm sharing this here because this is funny, but this same edge case has caused major issues for companies like Google, and they've misclassified people in photos as animals and in autonomous driving context in medical contexts. So it's important, again, to understand what we're doing here. We have incredibly powerful techniques. We have a medical device company changing the way they build surgical robots around this math. We have people who are able to do very fast, scalable analysis of satellite photos to get information in front of people making investments, right? But you have to think about it. You can apply some of these same techniques to text. So autosummarization is something I've been following and pretty passionate about for years, is take an article, simplify the article, and then in this system, extract the set of sentences from the article that contain the same information contained in the full article. Now, this relies on a technique called word embeddings, where you're actually able to create a vector that in some sense represents the meaning, and this is metaphorical of what is in the sentence. This is pretty cool. It's mildly useful. It gets more useful when you have a corpus of, say, 60,000 documents, and then you say, oh, there are 10 points of view in this cluster of documents. Here's a summary of each point of view. We've seen this particular technique useful, again, in an investment context, in medical patient records. The math is the same. Many applications here. We also just saw a really great paper from Facebook that used these same techniques to do automatic translation between languages where there's no pair corpus available. So you can build a model of language, build a second model of a language, and translate from that first one to that second one, which is incredibly powerful. I'm almost done. So interpretability is solving the black box problem. That is a set of techniques you put on top of black box algorithms that permeate inputs, look at outputs, and figure out how they work. Here we've applied it to a black box churn analysis. You can see all the features, and you can play with the features and see how it changes the classification probabilities. This is important for those of you in this room. If you are legally obligated to be able to tell someone how a system you build works, you need to know how this works. Because otherwise, you just have to use purely interpretable techniques, which is a tiny subset of the full spectrum of machine learning capabilities. I should point out as well that a black box might be a black box for mathematical reasons. It's a compression of a feature space, or it might be a black box because the person who built it just doesn't want to tell you how it works. This works on either case. Lots of fun stuff there. As we progress in the development of these technologies, there's still a few things we need to keep in mind. One of the biggest topics in our field right now is how we incorporate ethics, how we comply with expectations of privacy in the practice of data science. I'm going to take this opportunity to push a short e-book that I wrote, along with DJ Patil, who worked for President Obama as his chief data scientist. It's pretty short and it's very free. Our goal with this is really to try and get folks who are practicing out in the world of machine learning and data science to think about their tools for them to sort of practice ethics in the context of their work. I'm going to end with a lot of optimism. I think we are in fact building this AI-first enterprise. This technology will find its way into many fundamental processes of the businesses that we all run. When I say, let's make it boring, I actually think that's what makes it more exciting. I will end with that extremely confusing way of thinking about the world. Thank you very much and thanks for having me.