What are common threads between data science in healthcare insurance and astrophysics? CXOTalk speaks with a unique scientist with deep learning expertise in both fields.

Dr. Bülent Kiziltan is an AI executive and an accomplished scientist who uses artificial intelligence to create value in many business verticals and tackles diverse problems in disciplines ranging from the financial industry, healthcare, astrophysics, operations research, marketing, biology, engineering, hardware design, digital platforms, to art. He has worked at Harvard, NASA and MIT in close collaboration with pioneers of their respective fields. In the past 15+ years he has led data driven efforts in R&D and built multifaceted strategies for the industry. He has been a data science leader at Harvard and the Head of Deep Learning at Aetna leading and mentoring more than 200 scientists. In his current role, his data driven strategies with machine learning, analytics, engineering, marketing, and behavioral psychology components had a disruptive impact on a multi-billion dollar industry sector.

Transcript

Michael Krigsman: Part of our ongoing discussion about data science is data science in practice and so, today, we are speaking with somebody who is going to talk about data science in healthcare, health insurance and, amazingly enough, astrophysics. I'm Michael Krigsman. I'm an industry analyst and the host of CXOTalk. Before I interview our genuinely extraordinary guest, I need you now-now-now to subscribe on YouTube. That helps us a lot, so please subscribe on YouTube.

Bülent Kiziltan is an extraordinary individual. He is the former head of deep learning at Aetna. He has also been associated with Harvard contributing to their data science and astrophysics initiatives. I'm so thrilled to welcome Bülent. How are you? Thank you for being on CXOTalk.

Bülent Kiziltan: Thank you very much, Michael. It's great to be here talking to you and to your audience.

Michael Krigsman: When it comes to a healthcare company or a health insurance organization, where does data science fit in?

Bülent Kiziltan: Essentially, everywhere, eventually. It will help optimize almost every vertical, not only in healthcare but in almost every domain in the industry. Healthcare has been kind of a more conservative part of the industry, mainly because it's a regulated domain, so the access to data, how they use the data has to be very regulated. The general infrastructure of healthcare companies, especially ones that have been around for quite some time, had to be updated in order to make the best out of data science.

The obvious choices for using data science was to not only optimize the current processes, operational process, but also try to find new businesses, revenue streams that you want to create but, more importantly, why healthcare was so attractive for me was, you can use data science to improve members' health. Improving members' health is not mutually exclusive to the business objectives of companies and, in fact, they are very much aligned. We used to use a data science, analytics, predictive analytics to improve members' health, which essentially cuts costs for them and for my company. That would be the part that one would start using data science.

Michael Krigsman: That's a very intriguing and very high-level statement. Data science can improve the health of members, and members, in this case, if we're talking about a health insurance organization. The obvious question then, Bülent, is how. How can data science help improve members' health?

Bülent Kiziltan: There are many different aspects where data science can be effectively implemented. I think, just generalize, and then we can go into detail and look into examples specifically. But, we have to basically tune into each individual member, look at their profile, look at what geography they're living, what their socioeconomic background is, and what is actually preventing them to be healthier members or individuals. The lack of exercise, the lack of nutrition or malnutrition, not healthy food, not exercise implemented, not taking their drugs, medication regularly, not seeing the doctor on time, not getting vaccinated in time. We can look to the whole spectrum and try to communicate in a customized manner to each individual member and make sure that they take steps in the right direction to get healthier members.

Michael Krigsman: Is it then, ultimately, a matter of communication? How does this work?

Bülent Kiziltan: No, that's only one part of it. Data science, in this context and in almost every application in the industry, has to be taken as a holistic perspective and lives in an ecosystem where each individual part of the business process has to be effective. Unless your marketing, your communication is working really in tandem, effectively, with the data science team, the results that the data science team produces will not the create the value that it's supposed to.

When it comes to those individual members, we basically have to look at each individual

Michael Krigsman: Part of our ongoing discussion about data science is data science in practice and so, today, we are speaking with somebody who is going to talk about data science in healthcare, health insurance and, amazingly enough, astrophysics. I'm Michael Krigsman. I'm an industry analyst and the host of CXOTalk. Before I interview our genuinely extraordinary guest, I need you now-now-now to subscribe on YouTube. That helps us a lot, so please subscribe on YouTube.

Bülent Kiziltan is an extraordinary individual. He is the former head of deep learning at Aetna. He has also been associated with Harvard contributing to their data science and astrophysics initiatives. I'm so thrilled to welcome Bülent. How are you? Thank you for being on CXOTalk.

Bülent Kiziltan: Thank you very much, Michael. It's great to be here talking to you and to your audience.

Michael Krigsman: When it comes to a healthcare company or a health insurance organization, where does data science fit in?

Bülent Kiziltan: Essentially, everywhere, eventually. It will help optimize almost every vertical, not only in healthcare but in almost every domain in the industry. Healthcare has been kind of a more conservative part of the industry, mainly because it's a regulated domain, so the access to data, how they use the data has to be very regulated. The general infrastructure of healthcare companies, especially ones that have been around for quite some time, had to be updated in order to make the best out of data science.

The obvious choices for using data science was to not only optimize the current processes, operational process, but also try to find new businesses, revenue streams that you want to create but, more importantly, why healthcare was so attractive for me was, you can use data science to improve members' health. Improving members' health is not mutually exclusive to the business objectives of companies and, in fact, they are very much aligned. We used to use a data science, analytics, predictive analytics to improve members' health, which essentially cuts costs for them and for my company. That would be the part that one would start using data science.

Michael Krigsman: That's a very intriguing and very high-level statement. Data science can improve the health of members, and members, in this case, if we're talking about a health insurance organization. The obvious question then, Bülent, is how. How can data science help improve members' health?

Bülent Kiziltan: There are many different aspects where data science can be effectively implemented. I think, just generalize, and then we can go into detail and look into examples specifically. But, we have to basically tune into each individual member, look at their profile, look at what geography they're living, what their socioeconomic background is, and what is actually preventing them to be healthier members or individuals. The lack of exercise, the lack of nutrition or malnutrition, not healthy food, not exercise implemented, not taking their drugs, medication regularly, not seeing the doctor on time, not getting vaccinated in time. We can look to the whole spectrum and try to communicate in a customized manner to each individual member and make sure that they take steps in the right direction to get healthier members.

Michael Krigsman: Is it then, ultimately, a matter of communication? How does this work?

Bülent Kiziltan: No, that's only one part of it. Data science, in this context and in almost every application in the industry, has to be taken as a holistic perspective and lives in an ecosystem where each individual part of the business process has to be effective. Unless your marketing, your communication is working really in tandem, effectively, with the data science team, the results that the data science team produces will not the create the value that it's supposed to.

When it comes to those individual members, we basically have to look at each individual member. Obviously, we come up with models, then we approximate certain aspects, and we have information streaming in while we take those steps and update ourselves so that we can improve the behavior of each individual member. There is quite a bit of behavioral psychology involved as well.

For instance, when you look at the member base who are getting flu shots versus not, so we're trying to tap into their psyche in order to understand why they are not taking the flu shot and what we can do to make sure that they get vaccinated, they get their flu shots, and encourage their neighbors to get a flu shot. By taking incremental steps using behavioral psychology and cost data that we fold into our practices, we're trying to change their behavior step-by-step.

Michael Krigsman: Again, I'm trying to drill down on the multiple aspects you're describing. You're using data science to understand members' behavior in ways that would not otherwise be possible.

Bülent Kiziltan: Yeah.

Michael Krigsman: Then, you want to give those members feedback so that they can change their behavior based on the patterns that you've uncovered through your data science.

Bülent Kiziltan: That's one way of doing it. Specifically, yes, that's correct.

Michael Krigsman: What's the data that you would then be operating on in this use case?

Bülent Kiziltan: Yeah. In healthcare, there are two main streams of information streaming. One of them is from the member side, and one of them is from the provider side. From hospitals, doctors, we have information streaming in, and we have information data coming in from the member side. Sometimes they fill out forms. Sometimes they do certain applications.

Now, there is a tendency, a trend in healthcare, where healthcare companies are joining forces with pharmacies and labs, so they are trying to tap into additional information that comes from those additional streams in order to see the whole spectrum, the whole journey of a member from the pharmacy to the doctor to the lab, so we can basically come up with models that can give us some insight into the psychology of the particular person. There's also an additional trend right now where we can get some live information from wearables, like Apple watches, Fitbits, and some other means, some applications, by which we can really communicate their latest status to our infrastructure, to our recommender systems, and come up with better ideas how we can help the members be healthier. I'll be happy to dive into more technical detail if you want me to.

Michael Krigsman: I want to remind everybody that we're speaking with Bülent Kiziltan, who is a data scientist. We're talking about healthcare, and we're going to be talking about astrophysics as well. Right now, there is a tweet chat taking place, and you can ask Bülent questions using the hashtag #CXOTalk. It's really a tremendous opportunity to get your questions answered.

Bülent, you mentioned technical detail. I'd love for you to dive into that a little bit, please.

Bülent Kiziltan: Let's say we want to work on flu, in general. It's a $5 billion revenue stream for the industry. It's one of the notorious use cases, mainly because, for many decades, there has been some progress in the academic world, but it has not been successfully implemented in the industry to create value.

There are multiple components to this flu problem. One of them is, how can we predict when the flu season peaks at a certain geography or zip code? That is a whole different problem where you have to use past data, primarily from the CDC or your own, in order to predict, let's say, on November 2nd, what will be the outlook for flu in Boston as opposed to Miami?

I'll dive deeper into this because, to solve that problem or at least make some progress, I use my astrophysics or physics background, which is probably of interest to your listeners. Initially, we want to think of how disease spreads from one individual to another. They have to interact one way or the other. This is how disease is being spread.

I was thinking of individuals of, let's say, air molecules or atoms in a gas. Astronomers and physicists, for a very long time, many decades, have come up with mathematical approximations to define the dynamical interaction of individual particles. We use that all the time in astrophysics. We use that all the time in physics.

The mathematical base for this was laid down in the mid-1800s by mathematicians. This was called a Diffusion Equation. By using this equation and past data from the CDC, which is very, very useful, we can come up with models how flu is being spread in the continental U.S.

We can actually come up with models using deep learning. If we see a certain case of flu in Miami, let's say, late August, when do we expect a flu epidemic to happen in Boston, or if it's going to happen. We can use that aspect to make a prediction. That prediction plays an important role for your logistical assets, how you want to target that particular area.

If you, for instance, sent messages to your members two months ahead of time, "The flu season is coming. Get your flu shot," it might have an effect. But, there is a sweet spot how early you can reach out to your members, so you want to predict with certain confidence when a certain area will have a flu epidemic or case that is relevant.

Then, another area is what keeps your members from getting the flu shot. You want to get them immune or get the flu shot even though they are not perfect, but they are still contributing to the effort to stop the spread of flu.

There is something called herd immunity. You want to have a certain percentage of the population to have the flu shot in order for you to stop or slow down the propagation of flu. You want to get members to have the flu shot.

Then we have to look into our member base, look at the cohorts, and try to understand why they don't get the flu shot. In fact, most of the members are healthy individuals that have no problem with the science of the flu shot, but they don't feel the necessity to go to a store or their doctor to get the flu shot because they think, "I'm okay. I can get the flu and it doesn't matter."

Then we have to come up with ways to encourage them to go and get the flu shot. There are different ways of doing that. Behavioral psychology is an important one. Also, marketing strategies is an important one. I can't go into the detail of the strategy itself, but in order to look into the details of the cohort and try to track whether your approach is working in real time is a data science exercise.

Michael Krigsman: What's the accuracy of the results that you achieve using data science in these kinds of examples you're describing?

Bülent Kiziltan: Yes, so that really depends on the geography that you're applying to, and this is still an ongoing project. We have applied it in different locations in pilot areas. Typically, changing the behavior of members of individuals is very, very difficult.

If you are able to change 5% or 10% of the cohorts' behavior, I think that can be considered a success. That's great value, but it's an early assessment. I think, with the approach that we have, we are able to change a more significant fraction of the cohort that will add up significantly in terms of cutting the costs.

Michael Krigsman: Now, this is a little bit more technical, but how do you decide? There are different types of data science models and techniques. How do you decide which one to apply?

Bülent Kiziltan: Yeah. This is a question that comes up all the time, even when you talk to the most experienced practitioners in the industry. You cannot get an answer right away. They cannot tell you, "Oh, this is the problem. You can use this type of architecture. You can use deep learning as opposed to regular machine learning, or regular statistical methods," which I think are unvalued in the industry, stuff like based in statistics, for instance. It's very useful.

But then, what I do and what I've learned from my mentors is, I directly look into the data, what type of data it is. That will tell you what you can and cannot do. It's the ultimate determinant of your capabilities. If your data is no good, even if you use the most sophisticated method, you won't be able to get out something useful or of value.

Then, you have to make sure that your data is in order, it'd cleaned, and then you have to look into the data stream, whether your data is homogenous, whether it comes from one or two steams, or whether it comes from different streams. Different streams of information mean very different types of noise, different types of problems, sampling areas and biases. Then you have to basically step up your game in the sophistication or the level of sophistication that you implement or foresee in your model. It's really a problem that you have to take step-by-step.

In the end, even if you see the whole picture, it's very difficult to guess ahead of time which approach will perform better. What I do, typically, and what I tell my team to do, is to have multiple approaches and, even if you use the same approach, if it's just by two different teams, sometimes the outcome might be slightly different. It's basically a trial and error process, moving forward, and looking at the metrics of success.

Michael Krigsman: Bülent, I know that quite a number of insurance companies are using Apple watches and Fitbits, and similar devices.

Bülent Kiziltan: Yes.

Michael Krigsman: What are the data science challenges, opportunities, and advantages of using these devices in order to help? Help what, actually? What's the goal? What's the outcome of that? Let me get back to basics as well.

Bülent Kiziltan: Right. Since these are ongoing projects, I will be able to talk in general terms only. Those devices like an Apple watch or a Fitbit is basically tracking your daily routine. In some cases, actually--this is also open in the public--there is a lot of work that goes into recording your voice and, based on your voice patterns, making predictions about your health or your daily status, what you do, and your psychology, even. Even though some limited progress has been achieved in that regard.

In terms of the regular statistics, we know how much you move. We know what your heart rate profile looks like. That gives an enormous, detailed insight into your health. Once we combine that with other information that we have about you, about your, let's say, past history, about illnesses that you might have, the medications that you take, it gives us very strong indications about the status of your health.

I always say, in AI and machine learning at large, one plus one--if you do your job right--is greater than two, which means, if you look at the certain stream of information, if you get something out of it, if you combined it with another stream, the information that you produce out of that will typically be synergistic. It will increase the value of both of them and, sometimes, you can pick up information that is very sophisticatedly intertwined in what's called higher dimensions of this information space more effectively, so you will be able to tap into different patterns. You can tap into different sources of information, and you can act and execute based on that.

Michael Krigsman: In this case, if we're talking, again, in general, about insurance and using wearables, the goal is to raise awareness among the members of things that are going on with their own health, psychology, and habits that they would not otherwise see so that, therefore, they can take action, essentially, and operate themselves on that data. Is that an accurate way of expressing it?

Bülent Kiziltan: That's essentially the first order thing that healthcare companies are trying to achieve with this, especially with members where the psychological barrier is not significant. They're on the edge of making the decision and they can't. They just need an additional nudge. You have to come up with the appropriate nudge at the appropriate amount so that the member either goes to the gym, takes their medication, or gets the flu shot.

Michael Krigsman: Ultimately, I guess, that's the same. I guess the goal of any data science effort is to present the information that the person, the human can then operate on, in this case saying, "I'm going to get a flu shot."

Bülent Kiziltan: In that domain, yes, but at large, what makes AI so great is the techniques that you use are transferable directly from one domain to another. Essentially, what you do is you look at a data stream that might come from multiple streams. It might be sophisticated. It might not be sophisticated. What you're trying to essentially do is extract information and interpret that information based on the domain expertise that you have.

In the case of astrophysics, you interpret and you're trying to understand the information in the context of astrophysics. In the context of healthcare, you try to understand with that domain expertise and concepts. If you're in the financial sector, you're trying to do something else. But, essentially, the technics are very similar, if not the same.

Michael Krigsman: Let's actually talk now about astrophysics and maybe, very briefly, tell us what you're doing and what you have done with astrophysics.

Bülent Kiziltan: Right. I'm essentially trained as an astrophysicist. In that domain, what we do, and in many of the other fundamental scientists as they do, we look at information data coming from different sources. Sometimes it's observational. Sometimes there are simulations. Sometimes it's a paper and pencil; we come up with mathematical ideas in order to look for approximations. As Simon says, science is just an approximation to the ultimate truth, so we're trying to take those steps to get a better resolution on the ultimate truth in order to understand the universe.

I started as a physics major, and then I worked for the Space Telescope Science Institute, which is where the Hubble Space Telescope is being operated from. There, around 2000, I worked there as a data analyst, which is exactly what a data scientist is today because the term didn't exist then.

What we tried to do is, we tried to extract information not only from what's called
"big data" now, but also, when we have very limited data or when we have corrupt data. Astronomers have come up with really creative ways of extracting that information with classical methods.

Then, during my graduate school years, I teamed up with an applied statistician in order to implement Bayesian statistics in the work that we were doing at the time. Through that, we have been able to extract information from very limited data. I think, again, Bayesian statistics in the world of data science, I think, is still undervalued. It's very powerful. Once it's being complemented with ordinary machine learning methods, I think it's very powerful, and we have a few cases where we have extracted information more effectively.

Then during my time at Harvard, and then I spent some time at MIT, I reached out to some of the pioneers of deep learning and worked with them. I was getting together, exchanging ideas, and I was introduced to the world of machine learning through them and tried to implement the cutting-edge know-how that mostly computer scientists and neuroscientists have produced into my own domain of astrophysics. Through that, we have discovered a new type of black hole. At least, we think we have picked up the faint signature of a certain type of black hole.

Through that, I branched out into the industry because I saw an opportunity to make an immediate impact through the use of Bayesian statistics and machine learning at large in a domain such as healthcare.

Michael Krigsman: Now, what's the connection, the relationship, or the intersection between the types of techniques and data that you worked with in astrophysics and what you've done in healthcare? On the surface, to the layman like myself, it seems like you've got stars. You've got Apple watches. I don't see the connection. [Laughter]

Bülent Kiziltan: Right. We have to go back and kind of dissect the whole process. The first part of the process is getting the data. The second part is cleaning and making sure that the data is in place and is of high quality. The third part of the process is to extract information. All those three parts are more or less, well, I wouldn't say the same but very similar. Domain expertise in healthcare or astrophysics is more or less not relevant in those early processes, which take 70% of the whole data science process, I would say.

Then you have the latest stage. Once you have certain insights, you need to interpret them and turn them into actionable items. Now, this is where astrophysics and healthcare obviously have nothing to do with each other, but 70% of the process is very similar. This is why I said AI or data science is really great.

The skill set that you build, as an astronomer, as a computer scientist, is very transferable from domain-to-domain. Since this is a newly emerging field, we see a lot of data scientists in the market, but almost none of them are formally trained in data science because, from an education perspective, we still have not settled on the proper way to train data scientists. There are lots of courses that you can take, but there's no still formal training.

People from different domains are coming into the domain of data science and bringing in their really diverse backgrounds, diverse skillsets, and trying to contribute to the whole process that makes the 100%. Given domain experts play a crucial role, businesspeople who come from a business perspective play an important role in the execution in creating business value, but 70% to 80% of the data science process does require very little, if any, domain expertise. So, when it comes to the techniques, we can go into technical detail, but they are very relevant.

Astronomers, I have to kind of advertise that particular domain, mainly because astronomers and astrophysics are very unique in terms of their skillset. They use very diverse data sets. They are not limited, and they have to literally think out of the box all the time. Their creative aspect and talent is really, really important in extracting information in a domain that is just newly emerging. This is why astronomers and astrophysics make great data scientists, in my opinion.

Michael Krigsman: At the early stages when you're gathering the data and you're cleaning and preparing that data for analysis; the techniques are the same across topic domains. Then, once you have that data, now you're operating on it to solve specific problems is where you need the specific domain expertise. Is that a correct way [articulating] of what you just said?

Bülent Kiziltan: Absolutely. I think it is crucially important to partner with good domain experts. Also, on top of that, what is also sometimes missing in business operations is the complementary talent that engineers bring in, data engineers. Once a good data science team or analytics team pairs up with domain experts and engineers and, on top of that, have a business talent onboard that can execute on that, I think that is the ultimate that you want to have in any type of operation.

Michael Krigsman: We have a question from Twitter. The question is, "There are all these different techniques, whether it's deep learning or cognitive computing. And so, how should a businessperson relate to these? They're just almost meaningless words to your average businessperson."

Bülent Kiziltan: Hire data scientists and domain leaders to lead data science efforts. I think this is crucially important, mainly because of the lack of talent. It's a newly emerging field. Individuals who have deep domain expertise in analytics or data science in addition to a business experience, was not there, essentially, because the people who had the domain expertise have existed outside of the industry, in academia primarily. And, the people who have been in the industry have been disconnected from academia where cutting-edge know-how has been produced.

What we see in companies that are trying to basically get into the game of data science is sometimes they have businesspeople with very little domain expertise leading those efforts. I think, for the short term, that can be a remedy for the situation that they're in. But, in order to have and create sustainable value with AI in the long-term, I think it's essentially important to have domain experts with business acumen to run the AI operations. Also, an additional specific profile feature that the person has to have is being an educator really helps once you're in that leadership position so that you can communicate your really complex problems to the board or stakeholders within that business.

Michael Krigsman: I'm going back to the same question, really. How can a businessperson explain complex problems and models of data science to a very non-technical board of directors? How do you even approach that?

Bülent Kiziltan: Yeah. Again, being trained as an educator is important. When I was in academia, astronomers especially, they have a significant awareness of public outreach. We were giving regular talks to the public, and you have to talk in nontechnical terms and really explain very sophisticated concepts like blackhole evolution, stellar evolution to the public, sometimes grade school kids.

It's not very different in the business domain. Sometimes it's actually simpler, but you have to, basically, partner with a leader that can break things down for you and summarize things to that particular leader. This is where finding the right leader comes into play. It's crucially important to find a person that can explain complex ideas to those businesspeople.

On the other side of the table, obviously, you have businesspeople that have to educate themselves in order to grasp. There are certain CEOs that make it mandatory to almost every individual in their company to get basic literacy in machine learning, which will be very important. but, on my side, I think being an educator really helps in breaking down topics and explaining complex ideas.

Michael Krigsman: It's very interesting. Last week on this show, we had the chairman of Nokia who felt exactly as you just described that it's really important for business leaders to have a basic understanding of machine learning.

Bülent Kiziltan: Right.

Michael Krigsman: He himself went back to learn about machine learning and become a programmer again after 30 years of not doing it so that he then taught a course that now has been taken, seen, a video by many, many people, tens of thousands of people, inside Nokia, which employs 100,000 employees.

Bülent Kiziltan: It's impressive.

Michael Krigsman: It's pretty amazing, but the idea was exactly as you described. Businesspeople need to have a conversant understanding, at least, of what machine learning is.

Bülent Kiziltan: Especially in the context of what machine learning has to offer for the long-term future. I share the vision that many of the prominent members of the academic setting and pioneers of the field have on this. Machine learning will play a role in every vertical in every business where information is being stored. That's essentially every domain. In order for business stakeholders and individual members to be effective and contributing members to the operations, I think it's essential that they get a basic literacy in machine learning. That's right.

Michael Krigsman: Okay. Now, you have used the term "deep learning" quite a bit, and that's an area of specialty for you.

Bülent Kiziltan: Right.

Michael Krigsman: Can you tell us what is deep learning? How is it distinct from cognitive computing or machine learning?

Bülent Kiziltan: Right. It's a difficult topic. The reason is, it's a very specialized form of machine learning. What machine learning essentially is doing is looking at the information space and trying to dissect into pieces by using geometry, vector calculus, matrix operations, and come up with approximations by which you can extract certain patterns. Those can be called features, if you will. That whole process is called abstraction.

Machine learning, essentially, is a very powerful tool for extraction. You basically look into the information landscape and try to approximate certain types of behavior or trends within that information space and come up with actionable items. That actionable item part is called analytics.

Deep learning is, essentially, briefly, a more effective way of abstraction. But, one of the downsides of deep learning is it's very data hungry. If you are operating in a regime where your data is limited or it's not clean enough, sometimes you have to see whether regular or standard machine learning approaches are more effective, meaning you want to look into this mainly because deep learning is also computational and very expensive. If you are on a timeline, you have to basically look into how much effort will go into optimizing the process and how much value in terms of information you can get out of that.

Again, deep learning is just a subset, a very specialized form of machine learning, which can be a more effective way and, most of the time, if you have really good and a lot of data, can perform much better. The abstractions, especially in an information space where information is interlinked in a very sophisticated manner, not the first order or low dimensionality, but it's intertwined in a way, which regular machine learning cannot extract. If you can leverage that lack of abstraction by a lot of data, deep learning will give you that edge to extract that information more effectively.

Michael Krigsman: What's the takeaway for businesspeople because, in order to make the choice between models, for example, it requires the domain expertise and, of course, the understanding of the pros and cons of all of the models, how to apply them, when to apply them, and what's the right type of data in the right type of situation, which your average businessperson is never going to learn, not in this lifetime? What should they do? What should businesspeople do?

Bülent Kiziltan: The businesspeople should hire a domain expert to lead those efforts. That's the first thing. The second thing is, I want to emphasize; I see a lot of partly justified skepticism among business executives about the promise of deep learning.

I think one of the reasons why this skepticism has emerged is because there are so many people going into the field as data scientists who are not trained or sufficiently trained. Sometimes they overpromise what deep learning can deliver, and we have to be very transparent and honest about the pros and cons of deep learning.

Once you have a cycle that is achieving 80% of the value with 20% of the effort, they will stick with it. You cannot convince them to use a black box without a customized code and say, "Oh, deep learning is performing much better."

You basically have to look at this in a holistic perspective. You have to invest into your infrastructure. You cannot use the old hardware to effectively use deep learning all the time. One of the reasons why deep learning has become a hot topic very recently, even though the idea of deep learning has been around for more than two decades now, is mainly because of the advances in the hardware and the proliferation of GPUs that we can purchase on an individual basis.

Once you look at this whole picture, you have to have a transparent and honest approach to deep learning and, basically, see whether there's value in investing and whether it's sustainable in terms of the value that it creates with AI once that is in place and, also, once you have the talent that you need that can really customize your architecture, your neural network architecture. You cannot go to a black box. There are many commercial products where you just type in your Excel sheet, and you cannot just claim one of the columns is now optimized and I make predictions based on deep learning.

This is not deep learning, in its essence, and it's not what it promises in the long-term. There are lots of low hanging fruit that we can pick with this approach but, in order to be a leader in the field and bring in the cutting edge know-how with deep learning and machine learning at large, I think a diverse talent in the data science pool is very, very important.

Michael Krigsman: But, so many companies now, enterprise software companies, are promising that we use AI to solve--put a blank--every problem you can imagine. How can businesspeople see through those claims in order to make decisions based on factual benefits of these products as opposed to merely the hype?

Bülent Kiziltan: Right. Given there is a lot of hype and, because of that hype and underdelivering, there is some skepticism. But, one advantage of AI, as we kind of touched on at the beginning, is that 70% of the process doesn't require domain expertise. Back in the day, when you wanted to optimize a process and create value, optimization was just one part of it. You require domain expertise from multiple domains in order just to complete that 70%.

Right now, what you can do is, you can complete that 70% and add on top of that with one or two domain experts that are experts in their own field. What that gives the companies is they can pick up the low hanging fruit with that particular "AI." But, what I've seen over and over is that optimization or that value creation could have happened with other standard methods in the past. The reason why it didn't happen is either they didn't invest into it or they didn't have the talent in those multiple domains.

Now, AI is providing them with the means of coming up with better, more optimized business processes with data scientists plus a domain expert. But, as I said, I don't think that this type of approach will be sustainable in the long-term because of the archaic methods that were used or the lack of talent in business operations. There was a lot of inertia in midsize and bigger companies in the past. Startups, they operate in a totally different culture, so that's a different type of topic that we can discuss. But, midsize and larger companies, they create silos and, to break those silos and make those different verticals talk to each other in an effective manner, created some of the problems in the past. With AI, you can overcome that easily.

But, I say again, this is not the AI for the future. This is just picking up the low hanging fruit. If you build your strategy based on what you can achieve today, you will fail tomorrow.

Michael Krigsman: Okay. We have about three minutes left, and so let's finish off. I know you've been giving a lot of advice to businesspeople but amplify any of the points that you've made that you think are most important for businesspeople to understand in their approach to dealing with AI, machine learning, and so forth.

Bülent Kiziltan: What I would say is, stay away from generic solutions. Solutions, strategies have to be customized to your particular company, to the budget you have, to the business objectives you might have, the timeline you have, the technical skills you have onboard. You cannot just pick up YouTube or a consultant and ask them to deliver you a strategy that they have delivered to another company. Things have to be customized when you're building your strategy.

While you build that strategy, again, as rare as it may be, finding that domain expert with some business background and experience is crucially important to scale your strategy for the future if you want to sustain the value that AI creates. Also, data science analytics is a very interdisciplinary process, so I would encourage companies to diversify their talent pool and also keep on training your data scientists. If you hire a data scientist and then you overwork them 100%, if you don't give them creative space if you don't train them, they will become obsolete in six months, essentially, because things are changing on a weekly basis. I think keeping up with what's happening, being a constant learner and a student is also essential for that particular analytics team.

Michael Krigsman: Okay. Wow. This has been a very action-packed 45 minutes and really fast. Bülent Kiziltan, thank you so much for being here with us today.

Bülent Kiziltan: Wonderful to talk to you.

Michael Krigsman: I hope you'll come back and do this again another time as well.

Bülent Kiziltan: Any time.

Michael Krigsman: Everybody, you've been watching CXOTalk. It's been a great conversation about data science. Now is the time; subscribe on YouTube. That helps us out a lot. Go to CXOTalk.com. We have a huge number of videos. We'll see you next time. Have a great day, everybody. Bye-bye.