 Hello and welcome, everyone. My name is Eric Fransen of DataVersity, and we would like to thank you for joining us today for this smart data webinar, a production of DataVersity.net, with our speaker, Adrian Bowles, analyst at Storm Insights. Today, Adrian will be discussing commercial cognitive computing, how to choose and build your first cognitive computing application. Just a few quick points to get us started. Due to the large number of people that attend these sessions, you will be muted during the webinar. You've probably already noticed that. However, we will be collecting questions in the Q&A box in the bottom right-hand corner of your screen. That's different than the chat window, which is where you can reach me throughout the presentation if you have an issue you want to raise with me outside of a question for our speaker. In the Q&A box, though, we will be collecting those questions throughout, and we will reserve time at the end to get to as many of those as we possibly can. As always, we will send a follow-up email within two business days containing links to the slides, the recording of this session, and additional information that may come up during the webinar. And now a few words about our speaker, Adrian Bowles. The founder of Storm Insights, Incorporated, Adrian is an industry analyst and recovering academic, providing research and advisory services for buyers, sellers, and investors in emerging technology markets, including cognitive computing, big data analytics, and cloud computing. He has held executive positions at several consulting and analyst firms. Adrian also held academic appointments in computer science at Drexel University and SUNY Binghamton, and adjunct faculty positions in the business schools at NYU and Boston College. He began his career with research and application development roles at IBM and GTE laboratories. Adrian earned his BA in psychology and MS in computer science from SUNY Binghamton and his PhD in computer science from Northwestern University in my hometown, beautiful Levenston, Illinois. Adrian is the co-author of Cognitive Computing and Big Data Analytics, published by John Wiley and Sons in March of 2015. Welcome, Adrian Bowles. Thank you, Eric. I appreciate the introduction. It's a pleasure to be with you and the audience today. Wow. I hope you don't hear the dog. I'm in a shared space and somebody just brought in a dog. I don't know where that came from. Anyway, you never know when you're dealing with the sharing economy here. I've got a private space anyway. It's a pleasure to be with you and the audience to talk about cognitive computing. To me, cognitive computing really represents one of the biggest advances in information technology that I've seen in decades. I'd say that we as an industry, the IT industry, are finally delivering on some promises that were made when I first started out as a grad student in my first natural language processing application, which was actually on a mainframe with 64 kb, that's kilobytes. A lot of people don't remember that you could actually write programs in 64,000 bytes of space. And now we're seeing functionality that far surpasses some of the expert system-based work that I did for the Department of Defense in the 80s. I think there's still too much hype and confusion, but commercial cognitive computing applications are starting to deliver real value in a variety of industries today. And it's great to be able to share some of those advances with our audience today. As Eric said, we're going to hold questions till the end, but I will be giving you some contact information so that if you need to follow up with me afterwards, I'm happy to do that individually. So the one mark today is partitioned into three sections. First, I'm going to provide some general background and discuss the fundamentals of cognitive computing solutions. I want to look at how the critical technology elements fit together so that everybody has a common understanding tool. And then I'm going to outline the state of what I call the cognitive computing ecosystem, which includes the collection of technologies, products, and vendors. I'll have a word about investors too. Put them all together to enable the listeners today to get started building their own solutions. This section is based on research I just completed. Since the book, it's for report on machine learning vendors. And if anyone's listening and paying attention, I will tell you that at the end of the webinar, I'll provide contact information. If you send me any of that, I'll send you a copy of that report. It's got profiles on over 20 firms. And it's probably the most comprehensive source of information on machine learning vendors out there today. Finally, we'll talk about some of the critical first steps to help you get started if you're ready to build a cognitive computing application. That's a lot to cover, but I think we can do it in a lot of time. So, in the fundamental section, let's start with a definition. I like to say that cognitive computing is a problem-solving approach because if you're not solving a problem and doing some work with it, it's not really very useful. But the key here is that it uses hardware or software to approximate the form or function of natural cognitive processes. That's a lot broader than others that you may see and more inclusive. A lot of times, how people are talking about cognitive computing, they tailor the definition to their product operating, not surprising. But this one pretty much includes everything that I think I've been cognitive computing. So it's a mouthful, but the key point here is that we include both hardware-centric approaches, which are usually based on neural models of how the brains process information and software-centric approaches, which tend to focus on the functions being performed but not constraining the processing methods. So they're not actually modeling. They're not only thinking of as brain-inspired approaches. It's more like behavioral psychology. So whether you're modeling a cognitive process in a specialized hardware platform that mimics biological system or in software that simulates or emulates the functioning but not the structure of the biological system, we're going to consider that to still be a cognitive computing solution. So let's start with nature and some of the natural cognitive processes, learning, perception, and motivation. As Eric mentioned, I started out in psychology. That was my undergrad degree. And the three core courses that I remember were learning, perception, and motivation, trying to understand how this works naturally. And this breakdown into learning, perception, and motivation actually maps nicely and is relevant today for building cognitive computing solutions, not just looking at the natural ones. So in here, learning is the essential condition to kind of con on the thing without which you don't have cognitive computing. It's nice to have all these other things that we'll talk about today, but fundamentally, you have to have a machine learning solution at the heart of any cognitive computing system. So when I say it's, you know, we talk about learning in the cognitive computing sense, or specifically talking about processes by which the system can improve its own performance based on experience rather than reprogramming. So every time you get new data, that's an opportunity for the system to learn from the data and from what happens after it ingests the data and types of feedback, what types of learning in a second. But basically what makes it cognitive rather than a more traditional application that tries to improve performance by reprogramming or adding code after different experiences is that it has to be built into the system itself, into the learning engine, if you will, that it will change its behavior, it will change what it does or how it interprets data based on what it's already seen. Okay. So within learning, we're going to include reasoning, deduction, inference, and reflection. And a lot of modern cognitive computing solutions, they use one or more of them. You know, you will see some examples where you're actually using reasoning and deduction. Together, there's some inference. Basically, any one of these or any combination would qualify as learning. So of course, the system has to share what it learns and it will be useful. So things like reporting, expression visualization of the data as it's being presented to the users or to other systems, that's important. And we'll get to that in just a few moments. The three key approaches to machine learning. The previous slide was looking at natural learning and now we want to look at how it's done within the machine. So for machine learning, three things to remember. In supervised learning, it's perhaps the best known approach to machine learning. In supervised learning, the system is taught to detect or match patterns based on a set of training data. It's a case of learning by example. It should be familiar to anyone with small children or anyone who attended pre- or primary school in the U.S. or perhaps anyone who's taken into the A-class uses the case study that you're presented with a set of data and then you're going to look for things in the world as the system is operational that look like that data and try to find new patterns based on it. So it's all by example. And I'll say this is the most popular approach, if you will, to machine learning and we'll start to look at some of the examples. I'll focus on supervised learning. It's important to know that there are two other significant approaches that are important. One, reinforcement learning, which is actually a type of supervised learning in which the system learns your developed strategies based on performance feedback. So as you are, as the system is operational, it's going through with the data and it takes some action. It's like testing. Again, think of a small child that if you're not going to give them examples of everything that they have to do, you have to let them infer and test things, and then you give feedback. So this is much like in reinforcement learning with psychology. I hate to make the analogy with rats and cats and mazes in a psych class, but basically once they perform some action, you give some feedback and they learn based on whether it's positive or negative feedback, feedback being the reinforcement, whether or not that was something that is valued by the system or not. So in a cognitive computing system, when it does one reinforcement-based cognitive computing system, when it does something that we want it to do, it gets positive feedback. Reinforcement learning as a strategy is most appropriate choice when the system you're building has to perform, let's say, a series of tasks and there are just too many variables to develop a really representative training set in advance so you can't account for all the different contingencies. As an example, if you were building the very topical today, the logic cognitive system to learn for a self-driving car or an autonomous helicopter, all the different parameters you have to look at get very complex and to build a training set for that so that you can do supervised learning. Pardon the pun, but if it's a helicopter, you would never get it off the ground if you had to do that here. What we do is we let it go into operation and then provide the feedback without being specific about what it is about the scenario that you like or you don't like in terms of reinforcement. That's part of the learning that's incumbent on the system to actually provide. So finally, unsupervised learning as a technique is very different from either the two previous ones in unsupervised learning. The system is going to discover patterns based on experience. So it's looking for things that are different from what it's seen in the past, anomalies, and then bring those to the attention of the user or the user system to figure out if it's a meaningful difference. If it's something that, let's say, you're building a cognitive system for threat detection. You may have sensors throughout your system. One that we wrote about in the book was for the power grid. And they take samples through the sensors throughout the grid, voltages, amperages, et cetera. And different times of day, you may see fluctuations, but over time, the system will start to develop an understanding of what's normal and report on things that fall outside that. So it's discovering these patterns. You don't have to teach it to do that. It's teaching itself based strictly on the data. And those are interesting because if you contrast the supervised and unsupervised, supervised, you're telling it what to look for in effect in terms of the types of patterns when you're creating test data for it. You have to give examples. You have to know more about your data when you're doing a supervised learning system in advance so that you can create the appropriate training set. Unsupervised learning is all about discovery, and then recording that, and then you can have feedback. So you can possibly see here that you may have a system that incorporates both supervised and unsupervised learning for different parts or different features, that you're trying to accomplish. One of the terms that people talk about a lot these days in terms of kind of computing is maybe deep learning. And deep learning is really not that complicated. It refers, and I'm just going to use the definition that I've written out here, it's a biologically inspired approach to machine learning that leverages simple processing units. Each of the simple processing units are analogous to a neuro-synaptic element, like a set of neurons and synapses in the brain. The idea is that you have a lot of very simple processing elements that communicate with each other the way you have neural activity in the brain. And those can represent solid complex problems at different levels of abstraction. So 20 years ago we talked about this as a neural network but this is kind of the architecture for solving problems. And deep learning can be used in supervised, reinforcement or unsupervised learning systems. In general, what you're doing here is building a platform for your cognitive solution that has a very high degree of parallelism, has all these very small units, each of which is looking at a piece that may be a pixel or a region. Let's say you're building a deep learning environment to identify images in a file. Each of those is just going to look at a very little piece and then communicate with its nearest neighbors and say, I see, you know, this is a black dot or a red or a yellow whatever it is, what do you see? And communicate and then it builds from there into a larger pattern. So the key to this is that you can have deep learning structures that are supporting any of the three major types of learning. And also the deep learning could be implemented in hardware or software or combination. Okay, so we looked at learning. Take a quick look at perception. And in the world of psychology, natural cognitive processes, perception is all about how you recognize or sense that data. We'll see that there's a strong correlation between the human senses and sensors, hardware sensors, that are providing data in the artificial world. So perception is passed as an external stimuli. It relates in cognitive computing to how we acquire the data for learning. Today the focus is primarily on digital data, that's what we typically refer to as structured or unstructured and unstructured forms. For example, structure the records in the database. Unstructured might be natural language in a book or a magazine. In the future, the perception cognitive system is going to include data from sensors that will monitor real stimuli in the environment. So we're going to get closer and closer to modeling these natural processes with hardware. Within perception, if we look at the natural ones, again, the five sensors that we typically think of, see vision here and smell, taste and touch. Today what you see and hear, the way a cognitive computing system is going to do that, is through preprocessing of text or images or what I call surface structured records, things where the structure is obvious versus things that are deeper structured like natural language. We already have some pretty good sensors out there for things like temperature, tactile information and texture. And anybody who's involved with gaming is seeing controllers that have sort of haptic feedback where things vibrate, that's on the output side. Right now we're looking at it on the input side, but there's always an analogy. Anything that we have coming in is going to be analyzed through the cognitive system and then there's going to be some output. So we're going to at some point have output that mirrors all of these input types. So things like video are much harder for a machine to process, but humans don't have the same constraints. I want to look really briefly at a model that we use at Storm to help people understand this distinction between structured and unstructured. So here I've got a, it's really a five layer model of the level of structure. Zero is noise. You'll hear people say that there's unstructured data. I believe that for the most part when it's something that's really unstructured, when there's no structure, you're talking about noise and the key attribute there is that there was no pattern that was intended by the producer. So if you pick up a pattern and it's something you're imposing on, it's your representation. It's not really what was intended. But if you go from one to five in the spectrum, the independent structure would be what most people think of as unstructured. So video, free form, natural language, that type of thing. There is an inherent structure in natural language. We have syntax, we have semantics. We wouldn't be able to do programming languages if we didn't understand how to break up strings and parse and identify semantics before we executed them. It's the same thing with natural language. It's not easy. But the fact that the language has a deep structure doesn't mean that it's unstructured. And that's important because if you look at the higher level for the surface and shallow data, the things that we think of as structured, those are things where in general the intent, the source of the data, if you will, is machine. You know, you're dealing with records in a database or voltages coming out of a sensor. So those are very easy to be processed. It's once you get down in there between these mixed things like MediTank documents down into the desktop, that it takes a lot more effort. And in general, the deeper the structure, the more value you're going to get out of a cognitive computing application because cognitive computing applications today can apply the machine learning techniques to first find the structure and then find the data. The natural language is of course a perfect fit for this. I just think about it in terms of what's available to you today, things like, oh, we had applications like Dragon Naturally Speaker for a long time where you could understand the voice and translate it into the words so that you could dictate something and then have it put into a file. But it didn't understand the meaning, right? And now we're getting into those books that have seen examples like IBM's Watson where it can read text and understand what was the deeper meaning, what was intended. But it was done on Jeopardy, for example, where it gets into a DQA solution that's really looking far beyond that syntax. And these supervised and unsupervised learning techniques we talked about are ideal for that. So the final of the three natural processes that I wanted to mention is motivation. And motivation is really why we do something that in natural cognition guides the other processes, how we do or what we decide to do is based on our motivation. From an information process perspective, however, compared to learning and perception, it's a very fuzzy thing. So one day I'm sure we're going to have good sensors to understand motivation because today we already have good technologies for things like gesture recognition and emotion recognition. We're starting to have some of the output that uses gestures and where there's some output technologies today that will have a face speak to you, an avatar speak to you, and based on how you interact with it, it may be smiling, it may be firing, it may be mirroring what you're doing. So that ties in the motivation. But for now, for the rest of the webinar, we're going to treat motivation as a higher-order problem and say there's just outside our scope. So let's look at sort of the simplest version here, where you've got structured inputs or well-structured inputs and you've got simple reports that are being produced and, when I say the invisible, I'm saying that the memory itself, the knowledge, the accumulated body of knowledge about the topic is being updated. You don't see that from the outside, but the next time you ask a similar question, it's going to improve its performance based on changes in the structure of the data that's being held in memory. And here the kind of keyword is deterministic, as opposed to probabilistic. So here we're looking at something where you're building a system where there's likely one right answer or a set of right answers and you have a high degree of competence in those and you're just trying to look it up. So it's identifying the best answer from the data that you have. If we go into the next slide and look at a little more detail here, say, okay, we've got a perception or a language layer on the outside, but I was just getting to the idea of perception. Now we can open up and have much broader range input variables. So we're losing interface requirements to permit input and deeper structure with video, images, or conversational natural language instead of, let's say, a query language. And that can be used also on the output side. So with the feedback on performance we'll request more data. So we start to have a conversational system. This is showing input on the left and output on the right, but in fact the output is going to whoever provided the input. So that's where kind of the computing really starts to model perception once we get into different types of data that we have to identify the structure first. Regardless of the sophistication of the learning process or the types of sensory input, there are two basic or fundamentally different approaches to building kind of computing solutions that I want to make sure we touch on today. First, for some of you this may be a new word, neuromorphic architectures. In this, you see the highlighted hardware here. Typically neuromorphic architecture, the idea is to create a processing element based on a neural network, like the deep learning discussion we had a couple of moments ago, that learns by experience. And specifically with a neuromorphic architecture the experiences are modeled by neurons and synapses rather than programming. A lot of the neuromorphic research today is government-sponsored and university-based. The talk today is built as a commercial cognitive computing talk, and there's a good reason for including it here, which is that the commercial possibilities have attracted some serious attention from some of the large firms and some well-funded startups. So we're going to, we've actually written about some of these in our reports, and I'll be happy to send you. Today, in the interest of time, I'm just going to mention two of the large firms that are making big bets on neuromorphic hardware. IBM has made some announcements. Here I'm showing it with the Synapse project. Synapse was founded by DARPA, the Defense Events Research Projects Agency, which I'll give them a plug. It was a DOD that actually paid for my graduate student, my graduate work. And in this picture, you can see a Synapse board on the right. I actually put this on the back of an iPad just to give you a sense of scale, and this was done a couple of years ago since then that's all been convinced even further, as you might expect. IBM is now converging on building this neuro-synaptic chip with 10 billion neurons and 100 trillion Synapses. What's interesting about it, besides just the scale, is that it's only... the goal is to consume only one kilowatt of power and be contained in the system with a volume less than two liters. So in the size of a typical soda bottle, you'll have computing power of 10 billion neurons. So, well, there's a lot of academic interest and it's an interesting research problem. The commercial motivation includes developing systems that can process big data in an architecture that's not bound by Moore's law. These systems are inherently massively parallel and you can start to build on them without having to change the architecture. You just build on and add more processing elements. Qualcomm, the diagram on the left, which may be a little difficult to see on your screen, Qualcomm manufacturers, telecommunications components, they're found in most handsets in the system here in the world. They're also very active in this area in the small land. There were recently been advertising for job openings for systems engineers and cognitive devices to enable the next generation of mobile devices that we learned. And in this case, what is coming from is your behavior. So they are what we call brain-inspired chips in your phone that are from learning from things like how the thing moved and what's the environment around it as it's being produced. And that's actually pretty cool stuff. The neuromarket approach is the first major one. The second one is more of a functional equivalence where this is typically done in software. An idea is that you model the behavior of the biological system, not their structure. So you can treat all the things as if they were in a black box. And as long as you have the right set of inputs and outputs stimuli and you're getting response, you don't really care, you know, behavioral psychology sense about how things are happening on the inside. So here we've got a self-recentric approach. Typically, these systems are built on off-the-shelf hardware, although those things might be, I would say, very highly tuned to the task. And as an example, the IBM Watson system that won the Jeopardy Championship a couple of years ago, the software on that was extremely complex. There are like over a hundred some systems. There's a DQA engine in there, and it was all offline. It wasn't attached to the internet. But the hardware basically was off-the-shelf power seven at that point. Now they got it down to a much smaller footprint. But they're not building separate neuro-synaptic hardware for Watson. It's interesting that IBM happens to be building both of these things. And they are collaborating internally, but you can do one without the other, and that's the plan I wanted to make here. Okay. To close out this section, I just want to show the progression from natural processes to cognitive computing workloads, which is supported by a level of infrastructure processes here. And that's something that, as we're going into looking at how you're going to build this yourself, you need to look at the idea of data management, whether it's shallow or deep structure, unstructured or structured data, and analytics, which range from things that are descriptive to things that are predictive. All of these are supporting workloads for the actual cognitive workloads. So let's take a look at the market today. And I've partitioned this in terms of products and services. So here this hopefully follows naturally in terms of what I've said about the cognitive processes. If we think of machine learning as being at the heart of any cognitive system, to have a learning system, you need data coming in. You need input in terms of data and control, and I've segmented this into input from human sources versus sensors. And on the right-hand side, it's the complement or the mirror image. So where a person would like to interact with a cognitive system using their voice in natural language or perhaps gestures, and if you could actually see me on video right now, I just realized that I'm actually gesturing with my left hand to the screen, but it's not doing anything because it's not very cognitive. And understand things like emotions, whereas the more highly structured data that's coming in, the surface structure on the sensor's side, is mapped on the output when you produce reports. You can have things that could be machine readable. The output of a cognitive system could drive decisions for other systems. Another interesting part on the output is there's a 10-mile, besides visualization, of course, which comes along with most of the analytics tools. Narrative generation is a very cool company that I've been following called Narrative Science that is generating narrative reports based on the data. And if you've read all of some of these things, you would think that it was just a person writing it. At the bottom, we've got the infrastructure, and these, again, are the things that you need to build the system. The next slide shows about 20, I think 21 firms. These are the ones that we profiled in in our report last month. Again, if you missed the first couple of minutes of the webinar, I said to anybody who's interested in that and wants to get these profiles, just send me an email afterwards, or we can go through the data version. I'll be happy to send you a copy of the report. But this just gives you an idea that there are a lot of companies, perhaps that you've never heard of, some of these smaller companies, but also some of the bigger firms, like IBM, Microsoft, and Google, that are investing heavily in machine learning. So they've already made a lot of acquisitions. The VCs are investing like crazy in the startups. I would say right now the machine learning Gold Rush is already on. On this slide, the previous one looked at the firms that offer the actual machine learning technology. I created this one just to emphasize the fact that there are now firms that connect as an integrator or a general contractor for carrying out computing solutions. I expect this to continue, and we'll see more ecosystems like this evolve. So if you're looking at this now, and from a particular industry, and I've got the data, I want to build the application, you don't need to put it all together yourself. That's the point of this one. I'll go fairly quickly through a couple of slides now from the report that I mentioned, just so you can see there are some of these areas that are supportive of machine learning and cognitive computing that are focused on just the very specific technologies and you would need to integrate them. So for example, they voice natural language processing. There are lots of companies out there that are working on that. There are some interesting new companies that I list here that will all be companies that we interview and profile in the course of the next few months. They're doing things like looking at emotions on the input side. On the output side, there isn't much going on in narrative generation or a couple of companies that are doing a good job with that. The next slide here where I have analytics visualization and data management, really the only purpose of this was to show that there are a lot of supporting technologies that will prepare that data for input that will manage the data that you're going to need within the learning system. And one term that I'll come back to in a couple of minutes is the idea of the corpus. The corpus is a body of knowledge. The corpus formally is actually all the knowledge there is on a particular topic. So if you were interested in reading that Shakespeare, for example, the corpus would include all of the works of Shakespeare and perhaps all of the works written about it. And within a particular domain, you would need to decide how much of that information that's out there is going to come into your corpus, your learning system. And these are some of the tools that will help you with that. And the next two slides, all I wanted to say on this, really is just to show that some of these companies, even though you may not have heard of them, like Data or even Digital Reason, which has been around for a long time, these companies have attracted a lot of investment dollars out there and I'll get to another one in a second. I'm just going to mention Incutel here under Digital Reasoning. Incutel is basically the venture arm of the intelligence community. If the CIA is interested in what you're doing, Incutel will invest in your company. And it's interesting to me that Incutel has invested in a lot of companies in the machine learning space. And there are companies now, like Carious, that doesn't have a product out there. That's really 70 million dollars from all these different folks. Why I think that's kind of a signal, a market signal you need to look at, is that they're working on a neurosynaptic approach that's actually different from most others. It's based on some work that Jeff Hawkins, you mentioned, has done. And it's really getting attention of folks as a different way of representing a problem to be solved with a model of the brain that's a little controversial, but it's certainly effective in some of the early trials, I believe. So now, in the interest of time, I'm going to get into, I'm going to assume that this hasn't disturbed you and you're looking at it down the road. Yeah, that's great. Let's build one of these things. So I think the idea of application to get started is something where perhaps there's a function that's already being done by skilled professionals who can't keep up with the data. In medicine, for example, healthcare, diagnostics, a lot of the early work has been done in codifying case information. Many of the leading cancer and cardiac centers are already participating in research programs here where they're putting all their data from all their cases into a cognitive system, cognitive systems today, and can read pretty much every historical medical journal in a matter of hours or days and then keep up with all the new ones that are coming out, which no human can possibly do. So we have people who could solve the problem if they could deal with all the data, but they just can't because the field that they're in is producing too much data for any one person to process or if the people are relatively expensive and that could be something like in here retail recommendations from the site to one of the early adopters for cognitive was the North Face building the site to help guide you through your purchases. You might think, well, that's just sort of a search thing, but the difference here is between you saying, I want a backpack that weighs less than five pounds or you saying, in natural language, I'm going to Peru for three weeks, what do I need, and getting into a dialogue there. So that's where it's much less expensive ultimately to do that with the cognitive system or finally things that are high risk where we're seeing a lot of applications where network security, as I said for one example, the risks are very high if we don't act and we're seeing the uptake in cognitive computing from Homeland Security to the Department of Defense. There's one application that's being talked about in the community that students like monitoring drug traffic. I was at an event this morning talking about the $3 billion that IBM is putting into the Internet of Things and part of that is a relationship with the weather channel and now starting to look at how do you use weather data in an application. And one of the examples I gave was, they started to look at correlations and found that women want more care conditioning products if the weather was, I don't know, human. Now, you don't need a cognitive system to find that sort of correlation, but if you want to integrate your point of sale with your information, sending out information to a customer that's already exhibited this behavior, that's the kind of thing that a cognitive system can do and take the burden off the customer by recommending these solutions. So one of the questions when you're getting started is, is your industry ready? Can a cognitive computing application disrupt your industry? This is one of the diagrams that I created for our book, Looking at Smarter Cities and how we've gone from, in terms of maturity, a situation where there were a lot of silos, everything was document-centric or data-centric and a lot of manual systems, and the analytics that we had for different departments, if you think within a city, you know, you have planning, you've got transportation, you may have utilities. If you start out, a lot of things just sort of evolved independently and the departments don't talk to each other. Most of the analytics we've done as descriptive would give you what happened, but if you start to share information and build this repository, start to think of things in terms of a corpus for your constituents. When you get into predictive, we start to share information with the sensors. And one of the examples that I like to give is some of the Smarter Cities that are being constructed pretty much from nothing, particularly in Asia, are able to think about the information infrastructure before they build anything. So before they put in a power grid, think about how it's going to fit in with transportation. So if you're tying in information from sensors in the road to sensors in the rails and to sensors, weather sensors, things like that, you can start to have a much Smarter City if the systems talk to each other. And a big part of that is building this central repository. So one of the questions to ask is can you disrupt your industry? If you're not going to be making a major change, maybe right now you're just looking at a more comprehensive analytic solution rather than a cognitive one. My advice for any, even ever like this is to start with hard questions. And sometimes, you know, if this is going to be a customer-facing application, one of the things you need to look at is are your customers ready for a probabilistic or non-deterministic answer? So can they deal with uncertainty in multiple possible answers? Or is this something where they're going to be better served by something that gives you a fast response that says, here, take this. There's an old saying, Hobson's Choice. There's one guy who used to rent horses in the stable and Hobson's Choice meant that when you showed up, you would get whatever horse Hobson was going to give you. In some circumstances, in some environments, maybe that's what you want. But if you can deal with a situation or if the complexity of your data is such that you want to make a number of choices supported by probabilities and be able to explain the evidence. The evidence is a critical thing. How important is it to be able to explain how the system is about an answer? So if you're building a system and you want to recommend a course of action, a medical course of action, a diagnosis, then it's very important for you to be able to say, well, these are the five facts that led me to this conclusion, versus you're recommending a sweater. Well, we don't need that. So the issue here is understanding which of your applications is going to help you differentiate yourself. And a lot of that is based on, is the data there? Can you create the data using a machine learning environment? And can you identify with your customer base which of these features and which of these technologies you should invest in first? Because in those places, in those situations, you're not going to build everything all at once. Okay. And so we're going to close here. This is the diagram that I use to help people understand the mapping between the natural processes and the actual features, if you will, on the cognitive side. So foundation things are not in and of themselves cognitive. Anything that's in that first level, the learning level, whether it's experience-based learning, hypothesis generation testing, that's where we're getting into it. That's sort of building in that hierarchy. And then you add the other features, depending on customer need. As I said, it's possible to deliver applications with a subset of this functionality and still really change the user experience and change your industry. So I believe we'll assume this is version zero of our framework and we expect it to change. I should point out this is not a maturity model. You don't need to build from the bottom up. You could start with just the natural language processing, for example, and then add other features. Excuse me. Okay, so if you think you're ready, the key steps, and these are outlined in some of the other things that we've written, and I'm happy to treat this as the beginning of the dialogue, you have to identify the domain. Just what part of your business do you want to go cognitive, if you will? Choose the primary learning, machine learning model. This is going to be something where you have the data, you understand the data, you understand the patterns, and you can provide training data. So maybe you're going to do a generalized supervised learning algorithm. You can identify this, you have to reach the sources or if not, if it's very complex, like we said with autonomous vehicles, then you have to be able to identify the states or the events that are going to be reinforced. You need to be able to do that as part of your requirements. And finally, if you're unsupervised learning, identify the discovery parameters. What is it that you're going to want to know about? So once you've done that, you're going to be able to identify the discovery parameters, what is it that you're going to want to know about? So once you've done that, it's initially identifying the actual data sources. And I've worked with some companies on an assessment to help them understand whether or not they're ready. One of the key questions is, you may have a great idea that the data isn't there yet. You may have a great idea, but you need to integrate data from a variety of sources. Some of those are going to be internal, some of those are external. Some of them are folks that are not familiar with stuff with some folks at an IoT event in Boston a couple of months ago. I was doing a talk on analytics and IoT. There's so many opportunities out there right now with new devices that are producing data that you can get access to that you could fundamentally change the way your clients, your customers, actually see a problem. The example I gave there just quickly was on a website called thinkinful.net you could drill down and look at the bicycle rack in the hotel where I was giving the talk and see how many bikes were there. You can combine that with data from the weather company to see predictively what would be there and then start to build ancillary services. There's just a lot of opportunities to integrate but we think that has big data, predictive analytics, the whole idea of cognitive to make it smarter, but you can start to guide that to the availability of new information from the Internet of Things. Everything changes. My final slide here today, just an overview of a virtuous cycle that we want to build and this is sort of just a typical supervised learning scenario, but if you're planning to build your first cognitive app, you need to be sure that each of these steps account for before you start building that taxonomy, the representation of your data, understanding the sources and you're going to build this virtuous cycle. The reality is there are tools for each of these out there now. There are people that can help you. I think it's probably the most exciting time in IT since I got started many years ago, excuse me, but the key is identify the right app to begin with because it needs to be something that's going to be significant enough that when people have success with it, they're ready. I'm going to close by just saying that we'll open it up to questions, but if you would like more information, there are three sources, one, obviously you can buy the book, but we have a free LinkedIn group that I've started, and if you're interested in joining that, just send me an e-mail or connect with me on LinkedIn, and of course I'll give a plug for our host at the university. There's more data conference coming up in August. I'll be doing a panel session there and overview with more details on getting started. So with that, let me turn it back to Eric. Adrian, thank you so much. A very rich presentation, and that's just not only my personal opinion, but we have a lot of audience questions, so let's go ahead and jump right in. The first one here comes from someone asking about MLP. Do you happen to know what the current priority is for the big vendors to support non-English languages? That's a great question. I know there's some work being done in Spanish, and actually one of the vendors that I mentioned, Altilia, is an Italian company. There's some work there. I've seen some in Chinese, but in terms of general availability, I don't know that people are talking right now in terms of what they're going to deliver. At the simple level of natural language recognition, the parsing, there's already a number of languages and domains like medicine that are supported by vendors like Nuance, who's one of the early entrants there. Great. So I'm going to try to parse a question out of one that was typed in here. You said that Watson software is extremely complicated. The questioner appears to be asking about the nature of that complexity. Is it that what Watson is doing is a purely brand new achievement, a new idea, or just an extremely difficult success in engineering? Well, I would say when I say that it's complex, the version of Watson that has had the most exposure was the one that played jeopardy, and they have publicly acknowledged a number of the subsystems they've written scientific papers on. That can be sensitive to this. They are actually a client of mine. So I don't want to overstep to what I said. But the big issue there for complexity is what they were trying to accomplish with that first version had never been done before in terms of the deep QA, the deep question answering. So if you're familiar with the jeopardy game, it's not just reading a question and providing an answer. It's reading an answer trying to figure out what the question was. There's multiple levels there and it also required them to ingest a lot of data beforehand because due to the rules of the game, they couldn't be online to look at anything afterwards. So everything had to be partitioned in a way that they could meet the specs of the game and provide their best answer or best ranked answers in under three seconds, I believe. So that version of the system was actually run on a system that had 2,880 cores. So there's a lot of parallelism in there. Okay. Quick question here. What was the name of the organization that you mentioned with the controversial alternative brain model? Oh. NuMenta. NuMenta is a company that's founded by Jeff Hawkins and Donna Dubinsky. They were the founders of Palm and Handspring. So if you've ever had a Palm pilot, that was something that they did. And Jeff has been working on a model of the brain that's focused on the neocortex. That's very interesting. I've heard him speak a couple of times. It's one of the things where he's sponsored a research institute that's affiliated with one of the universities in California. And so NuMenta is his company for turning that into pieces of software that can be used by others. They have an open source product. And then some of the teams split off. And the vicarious company, one that cut the 70 million inventory funding from a lot of different folks, is actually also commercializing something based on their same model. Okay. Someone has asked for you to give input on the difference between machine learning and the data mining. That's a great question. The problem we have with terms is they tend to get fuzzy. And you could actually use machine learning algorithms in data mining. So strictly speaking, when I'm looking at machine learning, I'm looking at one of those approaches. It's a question of whether or not something has to be trained, how do you reinforce it, and whether or not there's additional programs to be done. If the system can either be trained or you can get it to improve feedback by, sorry, get it to improve performance by providing feedback, either as reinforcement learning, or after an unsupervised learning system gives you a result. If you provide feedback into the system about what you do with it, then all of that would be machine learning. You can certainly apply those algorithms to data mining. They're not mutually exclusive. Okay. We have time for one more question. So here we go. How do you differentiate cognitive computing from good old fashioned AI, the extra focus on machine learning, perhaps, but as the question or points out, machine learning has a long history in AI as well. Yeah, you know, at the beginning, you kindly use my own headline when I described myself as a recovering academic, actually. I studied a lot of AI, and I taught AI in the university studying it in the computer science setting. So when we're looking at cognitive computing today, you'll find people who say, well, you really just, you know, sort of putting a new coat of paint on an old idea. I think the idea, the distinction that I would make is that it's almost like AI has now matured. When we talk about AI, there are several distinct areas just like within psychology. You know, you've got behavioral psychology, you've got different approaches. Within AI, the study of vision and pattern detection. AI uses a lot of different techniques that have been brought into this cognitive world. I would say the simplest way to look at it is cognitive computing, some of that is an application of a number of AI techniques that matured now to the point where they are commercially viable given the ability we have to process big data effectively and manage it cost-effectively. I don't mean that in any way, as a slight to AI. No worries. Adrian Bolz, thank you so much for this great presentation and the Q&A. I'm afraid that is all the time we have today, but just to remind everyone, we will be posting the recorded webinar and slides to dataversed.net within two business days, and those of you attending today's live session will receive a follow-up e-mail to let you know how to access that material. Thank you to all of you. Please mark the dates for that smart data conference, and we hope to see you all online for the next webinar in this series. That will be next month on the second Thursday of that month. Thank you very much. Have a great day. Great. Thank you.