Eric Kavanagh, CEO of the Bloor Group, chatted with Sri Ambati, Founder of H20.ai, in a wide-ranging conversation about the ins and outs of machine learning and its widespread adoption by the business community.
Eric Kavanagh: Ladies and gentlemen, hello and welcome back once again to Inside Analysis. My name is Eric Kavanagh I will be your host for today’s conversation with Sri Ambati. He is the founder of H2O.ai, a very interesting company that’s doing some cool stuff with machine learning. Sri, welcome to the show.
Sri Ambati: Thank you for having me.
Eric: Sure thing. I’ve been tracking this machine learning space for a while and it’s quite fascinating how this stuff works and what it does. Obviously pattern recognition is a big part of how machine learning works, but I think the exciting thing is just how applicable it is these days in all sorts of business use cases. Let’s start off with your definition of machine learning, how it works and maybe get into where its applicable.
Sri: Machine learning is the ability to recognize patterns in data. These days, businesses are producing more data than ever before – the classic big data phenomenon. But data, by itself is not interesting. You want to understand what are the underlying models and patterns behind it, so you can answer questions: What would my customer do next? Where is he coming from? What was his journey like? Can I categorize this person’s experience as good, bad, ugly? How do I make him happier?
I think all these kinds of questions can be answered through modeling the overall customer experience by looking at patterns in the data. Machine learning is truly enabling businesses today to not just ask the exact answer for a question but what are the patterns behind it, how does my customer look, what would he be doing next, as opposed to what he did last week. So predicting his next behavior. That’s where machine learning is giving value to businesses these days.
Eric: Of course, there are a couple big examples that led us to where we are today. One, you were mentioning before the call was Amazon, with their next-best offer. The other is Netflix, trying to predict which movie you’ll like next. Can you talk about how those work and how they’re different and where they’re leading us?
Sri: I mean if you look at the original provenance statement of Amazon, it was not only will we sell you books, but we also will figure out what book you will like next. The whole idea behind how can we predict what movie you will like next from Netflix. Their big prize in late 2007 or 2008 led to a tremendous interest in data science as a whole community and profession. Now at the heart of all these algorithms is something called similarity search. Trying to figure out how else is somebody like, how close the relationship is, how are big giant vectors in relation to a third vector and are they closer than each other and, if so, can we categorize them as one group? Can we segment a group of points into sub-clusters so we can then treat them differently? Some of our customers, for example, want to know how to give white glove treatment to some customers, what patterns lead to them being very successful. Figuring out those are just as important for payment systems, for example, as they are important for a credit card company. This is really trying to get behind the data and see what are the underlying patterns.
As I was saying talking before, the fundamental notion of how far apart are two points, that’s led us to build the whole scheme of maps. For example, mapping is related to how we see two different points, how far apart they are and then try to use that as a metric to build patterns within your data and weave patterns within your data.
Eric: What’s really interesting to me is if you think about how this pattern recognition technology can enable insights to be derived from data. It’s really compelling because before machine learning you needed people to come up with creative ideas and essentially superimpose their own perspective on to a data set. Whereas now with machine learning, you can, essentially, get some insights via machine learning into the different patterns that exist in the data. You’re basically facilitating that first step of a human being working with the information to try to understand what the segments are, maybe what they mean and give some direction to where that person goes and try to better analyze the data and then come up with some ideas, right?
Sri: And what variables are important. What’s happening is big data is not just the number of rows that you can collect about data; it’s also the number of dimensions.
Sri: We live in a much higher dimension space and in that high dimensional space, you’re really confused with so many signals, you’re more likely to want to go beyond and find the top ten things that matter in your data. Top ten cities that cause air travel delays. Top ten, if you will, airlines that could be causing delays and so on and so forth. At some level, machine learning is leading to going beyond traditional day-to-day. Where do I learn new patterns?
Eric: Yes, it’s really facilitating this process of exploration and discovery. Instead of forcing the person to try to think through all the possibilities, you can get a much more clear look, as you just suggested, at the different dimensions in the data. That’s often where you can get a good start when you understand what stands out at you from looking at, let’s say the segmentation of data, or just the general topology of data. You can get some really good ideas of where to go and where to begin, whereas before you would have had to have been lucky or very creative to find those proper trajectories. Does that make sense?
Sri: Yes. It’s actually lucky and also being deceived by data you have seen in the past. Human experience is good but it’s also deceived by what has happened in the past and it’s more or less a bit biased. You’re always constantly fighting between what is called bias and variance in data. Right? That’s where you head to the next topic of how would more sophisticated algorithms work, random forests versus creative boosting machines. They’re constantly in the fight between what you’ve seen in the past to what you haven’t seen. Algorithms, all learning algorithms, whether it’s human or machine, are trying to be unbiased and be able to generalize experiences. Data is nothing but capturing events, day-to-day events of life and the nature of the universe, so that the fundamental characteristics of data reflect the universe, which is very sparse, and life, which is messy. Data is sparse and messy. To some degree, algorithms are trying to cut through those two characteristics and come up with an unbiased generalized opinion about the data in forms of models.
Now, Moneyball is the other interesting aspect that has led to proliferation of this thinking that experts could actually be wrong. The idea that data can be speaking completely different from the experts, and actually win, is a concept that’s a cultural newness, a novelty if you will. Moneyball transformed how the game’s played, so now everybody wants to apply Moneyball for their businesses. That’s the heart of machine learning going into businesses is everyone wants to transform their particular industry, whether it’s a healthcare hospital trying to improve patient treatment, whether it’s an insurance company trying to predict and replace an underwriter, or whether it’s a very large cargo company trying to figure out how to manage their fleets better. All of these guys are applying Moneyball to the problem and the notion that you can apply algorithms to make predictions better.
Eric: Yes, that’s a really good point you brought up there. I can see how especially here in the United States, and anywhere that they play baseball, people can really understand and wrap their head around it, because baseball, of course, is like apple pie. It’s so American, it’s really central to the culture here in America and the fact that Moneyball came out and was able to demonstrate to people how these kinds of technologies can work, I think was really critical in moving machine learning forward.
I love this concept you brought up about bias, because the thing we call cognitive bias, which is kind of a way of saying people can be stubborn, let’s put it that way. We’ve seen cognitive all over the place. I mean, I know some very senior people from some of the largest banks in the United States who, quite frankly, went belly up or almost went belly up when the subprime crash wreaked it’s havoc on the markets. They referred to cognitive bias, they said basically, “I was wrong. I didn’t appreciate the significance of macroeconomics factors.” In that context, you can see how if these organizations had embraced machine learning and had been honest with themselves, looked at some patterns and taken this more iterative and holistic and let’s even call it slightly more humble view of what the data meant at the time, they probably could have avoided some of these huge problems, right?
Sri: It’s a topic that’s dear to our hearts. There are some companies during that time which actually used the algorithm to manage their businesses. I call them the algo-haves and the algo-have-nots. The algo-haves, like Goldman Sachs and others actually prioritized algorithms to inform the decisions and predict almost nearly accurately when the big crash was coming. I mean you can see, businesses that are hundreds of years old with amazing expertise were destroyed in the wake of the rise of the algorithm. It’s very similar to the game being played out now with Google scale machine learning there are huge – There’s a data hierarchy, but there are also people with large data acquisitions, and powerful machine learning algorithms mining that data and coming up with insights on life, on the Internet and on the regular roles.
I think what we saw as more of a motivation and inspiration is to build the same level Google-scale algorithms that were only available to companies like Amazon, Google, Goldman Sachs and Netflix, and bring them into open source so every business can now use the same power of algorithms to inform their decisions as they’re running their businesses. That’s why open source matters for us. It equalizes that significant discrepancy in the algo-haves and algo-have-nots.
Eric: Yes that’s a really good point and I’m glad you brought open source into the discussion because what you’re describing is, just as you said, a leveling of the playing field. If you think about the overall impact of open source, well we’re still on this vector, with Linux having starting the open source movement arguably. In the last five years especially, it has really taken off for all kinds of different reasons. What you’re now describing is the fact that companies like H2O.ai can essentially take some of these open source technologies and some of these algorithms, put them together and suddenly you have this capability that can be leveraged for mid-sized businesses, for small businesses, for really anyone who has the creativity and the gumption to go out there and use this stuff, right? It’s not just going to be Google and Facebook and LinkedIn using these technologies. Now anyone who takes the time and builds a team or gets a consultant can use these technologies to really get their business on a very positive trajectory, right?
Sri: Yes. I think it’s a combination. The tools and the platforms are one part of the equation. Leveling the technology, R, which is the language of predictive analytics, and open source was also created at Bell Labs, alongside C in the late 70s and never had any significant software engineering teams improve their algorithm base and improve and scale the base algorithms that are used by R. What H2O has done has actually taken the fundamental Fortran C++ implementations that were single credit written by mathematicians for themselves. We rewrote them at scale so they can actually now deploy them without much effort across a cluster of machines, much like a Google or a LinkedIn or Facebook. We made that possible for regular businesses to be able to run this at scale and not be encumbered by data growth.
That’s only one part of the equation. It’s a cliché but it’s true that technology by itself doesn’t solve the human problem. The human problems are solved by asking good questions and having the courage to follow the answers that come from that question. I think that the barrier to follow a data-driven approach for making decisions is still high. I remember hearing Michael Lewis talking about this. When the Oakland A’s found out that they had to make these hard choices, where you no longer are hiring a baseball player based on how they look on television, then you suddenly have to make this unpopular call and that takes courage. So eventually the predictive analytics machine learning database is going to inform the decision maker with far more accuracy on what to do. The doing is still going to be a courageous decision to make.
Eric: Well, people like to think that they have good ideas and I think the real key is going to be for decision makers and businesspeople from all walks and all industries to really appreciate what the data can tell them and then work from there. Because the good ideas are still critical, but the key is to make sure you listen to the data and that’s where these machine learning algorithms can really help. Especially if you’re dealing with very large amounts of data, right? Let’s take the health care industry for example, I know you guys are doing some good stuff with hospitals. There’s a ton of data, I mean really there’s an embarrassment of data out there to better understand what’s happening. One of the problems is then we’ve used what we might called old-fashioned techniques for trying to analyze that data. Whereas if you use machine learning algorithms to take a look at large data sets you can very quickly realize that there are some fairly clear decisions that can be made early in the process. For example, when someone walks into a hospital. Those decisions can really facilitate getting that person the kind of care that they need, right?
Sri: Yes. I think heterogeneity of data is definitely making an impact. Before if you could narrow down the correlation between two vectors, whether it’s disease vectors or vectors of reaction to drugs. If you could find the correlation between two events, now you can actually go beyond. With big data you can go to causation, where you can actually look for what led to what, as opposed to what happened alongside what. I think that’s kind of the big transformation moment for diagnosis.
If you look at diseases in the grand scheme of things, there are about 5,000, 6,000 diseases and 50,000 kinds and symptoms and about 50,000 lab reports so the entire doctor, patient, physician visit diagnostic problem breaks down to 50,000 categories of lab reports, 50,000 categories of signs and how do I really categorize them into the 5,000, 6,000 odd diseases that we have? It really becomes a machine learning program at a grand scale. Only if you could pull up all of the data and actually analyze it together. I think one of the problems we have in the United States of course is patient data is probably more valuable than patient life. You can’t actually share that data across different hospitals and even sometimes with the hospitals.
One of the interesting algorithms that one of our collaborating professors from Stanford, Stephen Boyd, has come up with is consensus algorithms, which can now share models but not share data between two very disparate hospitals who are both fighting, let’s say, Ebola or flu. You don’t actually have to share data but now you can potentially model two completely different sources, data sources locally and then combine the models and then kind of fight together. Whether it’s fraud, two banks fighting fraud together, or two retail stores which can no longer combine their data but can actually combine patterns in data across two different stores. I think that’s the growing body of new algorithms which can now combine and fight diseases at both hospitals or fraud at banks.
The most popular use case for us in this space is sepsis. We have several hospitals, about 160 hospitals in the United States are going live on November 1 with a sepsis prediction, which again is an algorithm that’s predicting susceptibility for a particular patient if he was admitted into a particular hospital, to a particular drug treatment, and what are the hospital‘s acquired conditions that might happen. Hospitals are now being regulated to be responsible for hospital-acquired conditions. Sepsis prediction has become a very popular used case for predictive analytics and H2O is powering a transformation where you could use much more sophisticated algorithms to find the top five reasons for a possibility and a prediction that actually makes sure you can avoid a fatal condition, a nearly fatal condition, with sepsis.
We’ve seen another popular use case where a big hospital here in Oakland trying to predict whether to admit someone to an ICU or not. As you go into a hospital, to predict whether you should be sent to an ICU or to be admitted in a normal ward. Things like that are going to be more and more real-time decisions that are made by the algorithms as you walk into a hospital. I hope people don’t walk into hospitals that often.
Eric: Right. You bring up another really good point here which is that again, with machine learning and thanks to innovation in this space, you talked about, for example, the work that you’ve done with R to harden it so that it’s enterprise ready. it could work at scale, we can talk about education, we can talk about helping people understand where this stuff works but when you get this combination or this coalescing of forces and technologies and methods you can really facilitate things like diagnoses, so when a patient walks into a hospital a handful of questions could be asked of this person to get the symptoms down. Enter that into a system, it’s not just going to be read in 20 minutes by a doctor who comes and sees you, but instead can be fed into a more intelligent system that can then spit out a handful of significant options that are available at that time.
What we’re talking about here is facilitating human decision making. Some of the fear that comes around when people hear about machine learning is, oh this is going to take my job away, but really the idea here is to facilitate human decision making and get us on the right track and expedite the process of solving problems, right?
Sri: Data science is fundamentally a team sport. What we found is that, actually even whether it’s building of H20 or the use of H2O, we found that diversity really helps. You get several cultures of the community of customers now trying to apply data science so that it is the traditional quantitative person inside the company or his business counterpart who’s asking the right questions or his engineer who’s providing the data or deployment for the data science algorithms or the application builder. We’re seeing a conversion of really diverse cultures practicing and operationalizing data science in our customer base. It’s interesting to see the conflict that arises from that, and among the conflict itself we see innovation as well. I think data science is fundamentally a team’s sport. Even data scientists, themselves, come from different walks of life. We see the most common background is a neuroscientist or a physicist or a biophysicist trying to solve life science problems or real physical scientists’ problems are now applying their brainpower to try to transform day-to-day businesses.
It’s interesting you brought up financial services of 2007, after the era of exuberance there, if you will, a lot of the physicists and financial economists and engineering quantitative analysts have shifted away from Wall Street. You see a lot more of them joining day-to-day spheres and you see a corresponding growth in data science and quantitative science in making marketing decisions, in making sales decisions, in making inventory choices and making customer experience better. We have a whole slew of quantification of the rest of the spheres of businesses and not just the financial world. I think that’s the underlying transformation that’s happening in the businesses today.
Eric: That’s great, and actually that’s a perfect segue to a question that maybe we’ll make this our final question, though I’m guessing we’re going to have more conversations with Sri as time goes by, this is really interesting stuff. It really is taking over, I mean it really is taking hold in the market place, in the data management role. A lot more people are understanding what it is and where to use it. You guys talk a lot about smart applications and how you’re basically enabling this next generation of applications to come out. People see this already even though there may not be a whole lot of machine learning under some of the apps that they use, but you can do the math in your head and realize that with all these new form factors like the iPad in a hospital or using the iPad out in the field for a UPS driver can be enabled with applications that have machine learning buried underneath them to facilitate decision making on the spot, to facilitate analysis of complex scenarios that may arise. Can you talk a little bit about smart applications and how you see them shaping up over time?
Sri: If we think about the first, earliest applications, when I say earliest of Internet applications they were static. They were just text being presented. I have been around to see them, when we were thrilled to see them. Then came the data- driven applications, the first one led to the growth of HTTP, the Apache server. The second one led to the growth of MySQL, where you’re doing a lot of LAMP stack, kind of classic data-driven applications, PHP. Then Python and R and the popularity of machine learning, there was yet another chance to reinvent experience on applications on the Internet. Reinventing Internet, itself, this is also transforming the smart devices out there. I think what’s fundamental to this transformation is the shift from being a rule-based business application to a more of an online learning based, more data presentation model-based application building.
If you think about the web logic of the world, the BEA and WebSphere and the classic Jboss and Tomcat in the mid 2000s, we had the application servers powering applications that were basically canned business logic in a way that was thought through by software engineers. Now with the onset of data science and machine learning to understand patterns and capture the entire dimensional tree, if you will, of the logic in a very data-driven method. We now have more of an intelligent application where today the application reacts to the data on the fly and learns from the patterns it has. It applies scores on the user experience and changes user experience based on the scores. It’s truly giving a net-new learned behavior; net-new concepts are being learned. While you are experiencing the application, the application is also learning from your responses.
I think that’s the two ways application are becoming smarter. They’re actually interacting back, not just reacting but interacting to the customers who are showing up at the applications. Now that leads to the applications today are deployed on the Internet and the Internet itself today is an Internet of things. You’re looking at smarter things, they’re not just sensors, they could be actors in the grand scheme of the reinvention of Internet and machine learning coming along and able to process large sums of data and pick up patterns from there and anomalies detect spikes, detect breaches, detect on the fly, find out what might be interesting to a particular customer and transforming the content delivery on the fly. That’s kind of where truly new smarter applications are coming.
Now this starts with what you call the commodification of data science to a degree, where we can actually make it look for design patterns of smarter applications. Those patterns are templates and an NLP classifier can be used to classify jobs that you post on Craigslist. It can be used to classify comments on a particular podcast. It can be used to classify a patient’s records as well. Literally the same NLP classifier can be applied in four or five different ways. That’s a pattern for a smarter application, and it becomes part of a bigger application which has a smaller subset of rules that can break, but has this learning mechanism that makes that application grow along with data.
Fundamental to some of the algorithms we have built and productized is their ability to get better with big data. Whether it’s random forest or boosting or deep learning, they’re learning as the data is coming in. They’re able to look for nonlinear patterns. The more dimensions there are, the better they get. The more rows there are, the better they get. As a result, they’re learning on the fly.
Eric: That is really amazing stuff. I mean, you guys have taken the big picture view, it seems to me, and are enabling this whole new mechanism, a whole new method for developing software that, as you suggest, isn’t so static. That’s been one of the problems with software over the years: the logic is hard coded. It tends to be fairly brittle, you can’t move it around too much, you can’t flex it otherwise it breaks and that results in constraints, it results in things not being done very efficiently. What you’re talking about now is a whole new generation, revolution it seems to me, in application design and delivery and maintenance. I think it’s fantastic stuff.
Well folks, I do want to mention the H2O World Conference. Be sure to look that up online, if you go to H2Oworld.H2O.ai you can find out more information. I wish I could be there quite frankly it’s going to be good stuff. Sri, I have to thank you for your time today and thank you and your team for really moving the ball down field in this whole space of machine learning and leveraging these algorithms, not just for specific use cases but, as you suggest, to really open up the process of software development and design.
Sri: We are really excited. It’s a phenomenal team behind this project. We got here through word of mouth and the support of our community. It’s open source, open source got us this far and open source communities have carried us into customers and they have become some of our strongest customers. H2O World certainly is where 60 speakers, the vast majority of them are community and customers on stage, talking about their experiences with machine learning and H2O and how they’re transforming their businesses by building a cohesive team of software developers, data scientists and data engineers. We would definitely love to have your audience, which is also a very grassroots movement of people to come by and experience the H2O World. Eric, we’d certainly like to have you be there; if not this one the next one. We’ll try to continue to convince you to come.
Eric: There you go. Folks it’s November 9th to 11th in Mountain View California. Also, hop online to H2O’s website, H2O.ai. I think from there you can find meetups which are all over the country. With that, Sri, once again thank you so much for your time. This is fantastic stuff, I think you guys are at the epicenter of the revolution. Folks this has been great, you’ve been listening to Inside Analysis, we’ll catch up with you next time. Take care. Bye bye.