Emerge 2019 Program Keynote: Dana Deasy, CIO, DoD

What I thought I'd do today is take you through an update. It's not the first time I've talked about the Department of Defense digital modernization program, but maybe refresh it for some of you. And for some of you who have not had a chance to hear it, I think you'll clearly see that the art of the possible for us is going to be how do you deliver in a concurrent manner these four pillars across them. Clearly the largest IT organization on the planet, as you know, we have an excess of over 200,000 professionals between contracting, civilian, and in service that touch the technology world, and an excess of 47 billion dollars. So very large estate, a lot of moving pieces, and we have a digital modernization program that wants to change the way we think about delivering cloud, how do we bring AI into the Department of Defense, how do we modernize our command and control communications, and how do we think about what does next-generation cyber look like, offensive, defensive, and the resiliency standpoint of it. So big agenda. The art of the possible is being able to do this in such a way that it integrates seamlessly into how we're already delivering IT. So what I'm going to do today is walk you through that, but I'm going to spend probably 70% of these four pillars just on the AI part, as we have a lot of really exciting things now going on in that space. One thing I always tell people, I really think it's important because people look at this agenda and say, 'Well, where are your priorities amongst these four?' And I'll say there isn't a sense of one is more important than the others, but there is definitely a deliberate strategy around how are we delivering and sequencing these in terms of getting this done.

So first part of this strategy is, I think Sue's that Kent was up here earlier, as you know, the government has a smart cloud strategy very much aligned to what is occurring across all of the federal agencies. We are also participating in DoD. I always like to say though, people get captivated about the cloud strategy from basic infrastructure. I have and I will continue to say it is not the cloud itself that is exciting, but it's the art of the possible of what it represents and what we will be able to do with it. All the way from, let's face it, the Department of Defense is very episodic in nature, and because of that, if you think about the tactical edge of that deployment, of a fleet deployment, of troops deployment, of a forward base, you know, how do we move from a world where traditionally compute power has always been centralized to how do you decentralize it, to how do you truly take it out to the tactical edge. That for us is one of the most exciting aspects of cloud. And of course, cloud is very well suited to handle episodic events. It's, as you can see on this chart, it's what we choose to do with it. And for us, that is going to be all about a whole different way of delivering applications, building applications on top of it. I think anybody in the room who's kind of studied the history of clouds know that this idea of being able to move to a world where you've got to serve both very unique purposes, which we call fit-for-purpose clouds, and general-purpose clouds, and the portability of all that is, I think, something that's actually going to come to pass. I was just out recently in Silicon Valley, had a chance to learn and hear about where cloud is going in the future. And so for me, this idea of interoperability and the idea that we will need to be a multi-vendor, multi-cloud environment still stands true. I've always said from the beginning that is where the Department of Defense needs to get to. And I have always said that as we go to build out, probably considered to be one of the largest general-purpose clouds, we will need to find a partner to start us on that journey and take us through a lot of very important learnings. Cloud itself, interesting, important, but it's what we're going to choose to put on top of that that becomes really important. And since we've declared in a very big way that we're all in from a machine learning and AI standpoint, for those of you who work in this world or understand the nuances of what it means when you say that you want to really embrace machine learning and AI, the day you say that is the day you said you've also solved for a very large infrastructure problem. Because to do AI and machine learning at the scale that we're proposing to do it, it assumes in that strategy that we have a very large-scale cloud available that can handle supporting AI solutions all the way out to the tactical edge. This is why cloud is so foundational. It's what we're going to put on top of that, which leads to a discussion about AI that's really important.

Most of you know that earlier this year the executive order for artificial intelligence was released. Concurrent with that, we released our artificial intelligence strategy that talks about everything from how do we establish it, what are the conditions for, what are the right projects to do in AI, how is it we'll go about doing this in a joint-up manner across the Department of Defense. And then a big portion of that strategy document, for those who have had a chance to glance through it, it talks about this idea of establishing a Joint Artificial Intelligence Center. What I want to do is I kind of want to walk you through where we're at on that, but more importantly, continue to remind folks when we talk about AI, what is it we're really talking about at the end of the day. And the center of this chart really simplifies it, and I'm not trying to dumb it down, but I am trying to bring the light. There is a pattern that you go through when you develop AI solutions, and the pattern always starts with you got to get the right problem solved for. And once you do that, you then can start to target on where is the data, what is the data, what's the formats of that data, how do you bring that data in properly, condition that data that it's then in a readable format for you then to start to apply algorithms against. You then start a training process, you then test the results of that, then you have this iterative do-loop to make the algorithm smarter, and then at some point you have a version of it ready to release in production. So given that's the process you go through, the question then becomes is can you create some sort of a center that can start to refine, put policy, governance, tools behind that process. To do that, we've said that the Joint Artificial Intelligence Center, for something of a scale the Department of Defense is never going to be able to handle the full suite of every conceived AI idea across the services, but what we can do is start to look at those activities that cut across all the services where they do have a common problem to solve for. And we are referring to those in our AI strategy as national mission initiatives. And I'm actually going to take you through a couple of them today just to really bring it to life. And then there will be initiatives that we'll work on with a specific component, we'll want them to own that, but will want them to use the center's tools and capabilities, and referring to those as CMIs. At the end of the day, we look at this as a human-centered activity, meaning that everything we do, we always think about the human that's still in the middle of these decision-taking, no matter what the algorithm looks like. And we have to keep thinking about that. Many of these solutions are literally going to have to get out to the tactical edge, and I'll talk to you later on about some of the challenges we think we're going to face and actually how do you get an AI solution out to the tactical edge. So you've got this process of what it is you're trying to do through the center, but we recognize there's one more really important part, and that is the way we're going to go about establishing the AI center is we're being very deliberate. This is going to be a very large partnership engagement model across Department of Defense. As a matter of fact, of the first two initiatives I will share with you, one of them is based on a partnership we found through a small partner, private sector, and another one is based on the academic world. So already in our first two initiatives, this evolving partnership becomes a very important part. We also recognize this is an all-of-government effort, so we're spending a lot of time starting to work with the AI Select Committee. We're working with the recent presidential order on establishing a committee to look at AI, everything from ethics to how do you do this at all-of-government scale. The bottom part of this chart simply points out that the real heart and soul of the JAIC is that ability to create a form of reuse. If you think about that process you go through from selecting data, conditioning data, to training, to testing, there are tools that we will develop for different solutions. It's making those tools available in some form of a library where various services can come to and not have to reinvent the wheel from scratch. It's what we're looking to do. We think there's three fundamental things, that foundational tools. It's the data itself and how do we smartly think about how we're going to handle unstructured, structured, and semi-structured data. And then what are the actual agile development processes we will use, what sort of orchestration layers will we use. And then finally, there is a set of actual technical tools that we'll develop inside of the JAIC. So that's the model for the JAIC.

We recognize when we started this, I was asked the question one day last year, what will be the things that will hold the JAIC back, what will be the things that will cause us to be highly challenged. As I said, one will be the problem defined correctly of what we're trying to solve for when we go to start an AI initiative. And then of course, the almighty data is always going to be the second one. It's not a case that we're short on data. It's across the Department of Defense, all the agencies have ample amounts of data. It's getting access to that data in a format that's usable to teach the algorithms against. So I'm going to bring the first one to life by saying when we started the JAIC center, we said, okay, what's the problem that we can solve for that's common across all of the Department of Defense. And the first thing we looked at was maintenance cost. And you can clearly see here, there's a real opportunity for spending 78 billion dollars on maintenance. We said that's too large of a problem set. We said, well, let's take a particular area of maintenance and we said let's look at aviation platforms. Well, that cut it in half down to 33 billion, still too large of a problem. So we kept whittling this down until we got to a common asset that we thought would be a first good use case for AI and said the Department of Defense. And that led us to the H-60, the Black Hawk. This is just simply showing the Black Hawk under maintenance that still, I would argue, is still too large of a problem to solve for, as there's a lot of different key components that allow that asset to operate. So what we did was we sat down with the folks from SOCOM and we actually started going through the operating environments, the conditions of which the Black Hawks are used. And what we concluded was a very significant problem. If you look over here to the left, was when you fly these Black Hawks in conditions where there's a lot of dust and a lot of sand, the sand builds up inside the engines. That sand turns to glass, and then that causes fatigue and eventual wear and maintenance on the engines. We said, now that's a particular problem that we can use AI to work on. So we have actually already delivered approximately three weeks ago, I believe, version 1.0 of this algorithm to SOCOM that they're now using on the Black Hawks. And you're going to see it's then going to be delivered to the Army, the Air Force, and the Navy. So this is a really good example of what do I mean by developing a common toolset that can then be applied to a common problem across the Department of Defense. In this case, I will use a combination of the private sector, small companies through an industry day that we identified, as well as universities that are helping us to solve this problem.

Next, humanitarian assistance and disaster relief. Now, this raised a lot of eyebrows when we first chose this one, and I have to tell you, it's absolutely a perfect use case to bring inside of a Joint Artificial Intelligence Center. So at the heart of one of the things that DoD needs to do is saving lives, and we do that through humanitarian disaster relief efforts. It's core to the mission set of the DoD. It's an opportunity to partner and learn. In this case, what we believe to be, remember the top of that JAIC chart, I said that we're going to need partnerships, we're going to need private sector, we're going to need the academic world, we're going to be international, and we're going to need to use other domestic agencies. So this next one I'm showing you is actually using all four of those types of partnerships to deliver. And the other thing that's really quite interesting in picking this is it allowed us to provide us a laboratory for applying AI in a chaotic environment. And I'm going to explain to you in a minute here what we mean by a chaotic environment. So this was one of the reasons why we were very drawn to this space because of those attributes. So what is the chaotic environment? So this last year, many of you are very aware of the wildfires that took place in the state of California and the devastation that they left there. So we looked at this problem set and we started working with the agencies that are involved in disasters and we started working with the firefighters. And we said, what is it that is most challenging in trying to solve for a fire of that scale on the left, so it does not lead to the devastation on the right. And what we found was, if you looked at the bottom left of this chart, it's all about the fire line. It's being able to map the fire line, understand the intensity, the strength, and the direction of how that fire line moves. And today, the way that is done is they use these very large boards, they'll take them out, they'll map the fire line by visually looking at it, they'll bring that back, they'll then pass that information on to the firefighters. There's obviously a lag of time there that takes place in this scenario. So what we looked at was could we use AI to actually map the fire line, teach the machines how to interpret what a fire line looks like, the intensity, the direction, and in doing that, can we then actually send directly to handheld devices, which you see up there on the far right, of a firefighter, that they then get a real-time response of the nature of what's going on with the fire line. And that's exactly what we are working on right now is the version 1.0 of delivering this to firefighters. Here again, what's really interesting about this one was a combination of the university, private sector, and providing the tools. And this is a great example of working with other federal agencies to solve for this particular problem set. The other one was looking at hurricanes and the aftermath of the flooding that occurs. And as you know, this last year, the Carolinas truly were so severely damaged by the flooding that occurred in the aftermath of these hurricanes. So what is here again, getting to this thing I said earlier, what is the right problem you're trying to solve for here. And the case of this, if you look to the left, this represents the current state of technology of taking imagery and looking down and what we call a WAMI, a very wide area, and being able to see what is occurring as far as the floods. The problem with the left when you look at it from the human naked eye is being able to distinguish, you'll see here this kind of darkened area is where flooding has occurred. However, your depth, the height of the flooding, and where there is humans or other assets that are in the way of that flooding is very hard to distinguish. So what if we could use AI, taking the sensor off a wide area imaging, be able to teach it how to distinguish land from water and other sorts of assets, which is what we're trying to demonstrate there on the right. And you can see here the reports, the results can be quite dramatic in terms of what it does for you. So once again, this is another example of where we're working with multiple agencies and we're working with both the private sector and university to come up with the state of the algorithms to deliver this one. And once again, this one is getting ready to deliver version 1.0 in the very near term.

I want to give you a sense of some of the things we're thinking about next. So for that first one and that second one, remember the model I showed you of ingesting data and conditioning the data, train it, put a production algorithm out. So we'll go through a set of iterative process now as we continue to get more data. For example, off the H-60, and surprise, surprise, not all H-60 data is the same. Different generations, different sensors. So this is a really good example of what I mean by even though we may be able to find a common asset to work on, the idea of all data being common off the asset is just not a reality. And we're going to have to learn to this iterative approach of how to bring more data in. This leads to a really interesting discussion in the future that we're having conversations already with the RAND organization and others about, we think about image optics, sensors, they've always been designed for human interaction. And the world where you now have the ability for the machine to be a partner of yours, how do you think about next-generation assets that we will build in the sensors of how it can be read directly by a machine. So it leads to some interesting conversations about future acquisitions and future development of assets across the DoD. Another one that we are currently just in the early days of starting to scope a problem definition around is in the cyberspace. And as you imagine, there is a lot of problem. If I asked 50 of you in the room to describe what you believe would be a really good problem that AI could solve for in cyber, I'm sure I could get 50 great ideas. Well, we're going through that iterative process right now. We believe it's going to be something around identity and recommendations of being able to find anomalous behavior on a network. We have here again a huge amount of data that exists across all the services that we've collected for some time. This starts to be able to identify baselines, and here again we can start to use thinking about what are the right algorithms for the right data sets to get after this anomalous behavior. For this one, we are actually right now working with the U.S. Cyber Command and we're also working with the NSA folks to start to be able to say, okay, if it's anomalous behavior that we're looking for, once again, where did the data sets come from and can we get the problem specific enough that would allow us to get a version 1.0 out there. And where and how would you put that in the hands of people and how it is that they would actually operationalize that. One of the things we're really learning about the JAIC is, and this is really no different than any other technology I would say anybody has ever delivered, it is not only the use of the technology but then how do you get that into the hands and operationalize that so it can be used and supported in a continuous basis. This one will be really important that we get that operational delivery model right because think about the distributed nature of how we support cyber across the DoD. This is a really good example where nailing the operational aspects will become really important.

So what are the art of the future possibilities, the art of the possible? You know, right now I would say that one thing that I know has gotten a lot of press recently has been this whole discussion on suicide mitigation. If you think about the amount of data that the VA has collected in this space, this is a really interesting problem set. You know, how can you start to use all that data, all those triggers, all those particular elements to start to map out, could you help solve for a suicide mitigation. Operational planning, obviously one thing the DoD does a lot of is operational planning for different sorts of operational planning sets of problems. And then finally, predictive medicine, which could be possibly a variant of the first one, but this obviously for the nature of what the DoD does, there will probably be lots of medical applications that we'll be able to look at and think through in terms of how it will help. So right now we have our two initial deliveries going on. As we get ready to enter fiscal 2020, we will do probably iteration two, three, and four of those first two, will start to iterate and actually start to deliver on the cyber sensing one, and we will start to bring to life probably a couple of these as we get further into 2020. The real thing we have to also solve here is how do we get more parallel different problem sets going. The JAIC is not only a physical place, so it's both inside the Pentagon, it's over in Crystal City where we have it set up right now, but it's also going to be physical hub locations next and close to certain universities in the United States. But our biggest challenge continues to still be people. That probably wouldn't surprise you. There's a variety of different types of people you need. You need architects and you need engineers, you need data scientists. Today, the JAIC is staffed by a combination of civilians, contractors, and detailees from the various services. And it's probably today a little bit of a third, a third, and a third. By the end of this year, we hope to have the JAIC staffed up to approximately 70 people, considering that when we kicked off the JAIC back in November, December, we had three. So we've gone from three to almost 70, from no deliveries to our first two successful AI deliveries already occurring as I discussed earlier.

Okay, the last two things I want to bring up is so we've discussed cloud and I've kind of built this case around why cloud is so foundationally important. And there's no doubt that the more that we actually start to deliver and practice working with a general-purpose cloud, one of the things that that starts to do is it starts to solve for the pockets of data that sits in unique places. Cloud doesn't necessarily put it all in the same format, but it does allow it to be more accessible. And that's one of the things we're really excited about as we go to solve that problem set and artificial intelligence called access to data. So let's say we're highly successful in getting a cloud going and let's say that the JAIC can start delivering more and more solutions, and we're going to use that compute capability from the cloud to get it out to the tactical warfighter. Now, it's that last part, getting it out to the tactical warfighter, is where the challenge sits. And I have made this comment more than once, we could have the best cloud in the world, we could deliver awesome AI solutions, but if I can't get that algorithm and the right data set out to the tactical edge, it's all for naught. So to that end, that brings up the third leg of our strategy, and that is what does next-generation command and control of communication need to look like for the DoD. And I'd say that there is two problem sets that are distinctively different but have to be done smartly here in this space. One is, you know, if somebody said to me, well, what is the problems that we're trying to solve for, I always tell the story of, you know, I started my career back in the very early 1980s, the PC hadn't been invented, the World Wide Web hadn't been invented, cloud was not even on anybody's diagram, and malware had not even been a term, and cyber was just starting to come on light right next door to where I work in the aerospace industry and the Space Shuttle program. The initial Navstar GPS constellation was being built. If you think about the era when that was built in, that was built all in a pre-cyber era. And I would say many of the assets from position, navigation, and timing, tactical comms, SATCOMs, much of this is still legacy that we're carrying. So problem number one set is we just have to solve for a legacy that we've kind of grown up with. That legacy today is obviously it's congested, it can be detected, it can be displayed, it can be denied, and could be disrupted. All problem sets we have to solve for, and we know we have to solve that for space, undersea, air, land, and we have U.S. Cyber Command themselves. So that's one half the equation. The other half the equation is to be able to fight using space, undersea, air, land, and maritime assets with our ally partners really brings us other big problem called data operability into life. And so to that end, one of the things we've just kicked off recently, I've just signed a memo that went across Department of Defense announcing how we're going to work with the Joint Staff to get after this data operability problem while at the same time also addressing how do we move to next-generation command and control and communications. So now that I kind of laid out we have a foundational cloud, we've got to put AI on top of that, and I've talked about the JAIC, and then finally we have to be able to solve for a world where many of our common approaches are congested, they are detected, they can be exploited, and if you're going to fight in those multi-domains with multiple ally partners, data operability becomes obviously increasingly important. And then finally is the cyber landscape. Now when I have the cyber conversation, I always say to folks, you have to stop and say what conversation are you in. When you're in a cyber conversation, are you an offensive conversation, are you in a red conversation, are you in a blue conversation, or are you in a resiliency conversation. So General Nakasone and I have become very close partners in this. Him and I started in our respective jobs the same week, so we had a very good opportunity to become partners very early on. I look to him to obviously work on the red and the blue, and I've taken on the responsibility to say what does resiliency and remediation look like across Department of Defense. And to do that, we had to paint this picture called, you know, what were the big asset problems that we need to solve for. We talked about remediation of the cyber landscape across Department of Defense, and this is all the way from the endpoint that we all use on our desktop every day all the way out to the weapon system at the other end of that. But as we know, adversaries will look for your most vulnerable point anywhere along this chain, which is whether it's your endpoint, whether how you access your applications, the applications itself, the data, how you encrypt that, how you communicate that, and how do you actually enable that in a warfighting asset. And so we have a program of work now underway to address all this. And oh, by the way, there's two other key elements that often get forgotten in this conversation, and that is we use a very large industrial base, many of you are in this room today, thank you for everything you do for the Department of Defense. And then finally, we have the critical infrastructure of the U.S. itself and how do we think about those critical industries that we depend on their availability. So I will end by sharing with you that hopefully I've been able to paint a picture for you of how the four things fit together. On that last one called cyber, here again, multi-dimensional problem, how do you fix the past whilst at the same time as you're developing a future with cloud and AI and next-generation embedded cyber in a way that gives you more resiliency from the start. We know cloud is our foundation. We know AI is going to be a lot of our enabling the warfighter of the future. It will rely on a strong foundational cloud. We know command and control communications and getting the results of cloud and AI out to the warfighter, the tactical edge, is a very large problem set. We know fighting in a multi-domain environment has become a very large problem set to solve for. And then finally, building in resiliency to those first three and also solving for kind of the problem set of our current asset base is the last part of the strategy. So the art of our possible is how do we do all that, how do we integrate that, how do we time that, program manage that, sequence that, budget for that, find technical solutions for all that. And all of that is going to require a huge, huge partner base to pull this off. So that is the art of our possibility across Department of Defense. Listen, thank you very much. It's been a real pleasure having a chance to talk today, and hopefully this has given you some better insight to what we're doing at the DoD. Thank you, everybody.