 Hi, I'm Peter Burris, and welcome once again to Wikibon's weekly research meeting from theCUBE Studios in Palo Alto, California. This week we're going to build upon a conversation we had last week about the idea of different data shapes or data tiers. For those of you who watched last week's meeting, we discussed the idea that data across very complex distributed systems, featuring significant amounts of work associated with the edge, are going to fall into three classifications or tiers. At the primary tier, this is where the sensor data that's providing direct and specific experience about the things that the sensors are indicating. That data then will signal work or expectations or decisions to a secondary tier that aggregates it. So what is the sensor saying? And the gateways will provide both a modeling capacity, a decision-making capacity, but also single-dutertiary tiers that increasingly look across a system-wide perspective on how the overall aggregate system's performing. So very, very local to the edge. Gateway at the level of multiple edge devices inside a single business event, and then up to a system-wide perspective on how all those business events aggregate and come together. Now what we want to do this week is we want to translate that into what means for some of the new technologies, new analytics technologies, that are going to provide much of the intelligence against each of this data. As you can imagine, the characteristics of the data is going to have an impact in the characteristics of the machine intelligence that we can expect to employ. So that's what we want to talk about this week. So Jim, could be honest, with that as a backdrop, why don't you start us off? What are we actually thinking about when we think about machine intelligence at the edge? Yeah, Peter, we have the edge device, the device in the primary tier that acquires fresh environmental data through its sensors. What happens at the edge? In the extreme model, when we think about autonomous engines, let me just go there very briefly, we think basically a number of workloads that take place at the edge, so the data workloads, it is so fully adjusted, it may be persisted locally, and that data then drives local inferences that might be using deep learning and machine learning chipset that are embedded in that device. It might also trigger various local actuations, things, actions are taken at the edge, if it's a self-driving vehicle, for example, and actually it might be to steer the car, brake the car, turn on the air conditioning or whatever it might be. And then last but not least, there might be some degree of adaptive learning or training of those algorithms at the edge. Or the training might be handled more often but the secondary or the tertiary tier, the tier, the tertiary tier at the cloud level, which has visibility usually across a broad range of edge devices and is ingesting data that is originated from all of those many different edge devices and is the focus of modeling, of training on the whole DevOps process where teams of skilled professionals make sure that the models are trained to a point where they are highly effective for their intended purposes. And those models are sent right back down to the secondary and the primary tiers where influences are made 24 by seven based on those latest and greatest models. That's the broad framework in terms of the workloads that are going to take place in this fabric. So, Neil, let me turn to you because we want to make sure that we don't confuse the nature of the data and the nature of the devices which may be driven by economics or physics or even preferences inside a business. There is a distinction that we have to always keep track of that some of this may go up to the cloud, some of it may stay local. What are some of the elements that are going to indicate what types of actual physical architectures or physical infrastructures will be built out as we start to find ways to take advantage of this very worthwhile and valuable data that's going to be created across all of these different tiers. Well, first of all, we have a long way to go with sensor technology and capability. So when we talk about sensors, we really have to divine classes of sensors and what they do. However, I really believe they will begin to think in a way that approximates human intelligence. About the same time as airplanes start to flap their wings. Now, so I think, you know, let's have our expectations and our models reflect that so that they're useful instead of being hypothetical. That's a great point, Neil. And in fact, I'm glad you said that because I strongly agree with you. But having said that, the sensors are going to go a long ways but there is a distinction that needs to be made. I mean, it may be that at some point in time, a lot of data moves up to a gateway or a lot of data moves up to the cloud. It may be that a given application demands it. It may be that the data that's being generated at the edge may have a lot of other useful applications we have anticipated. So we don't want to presume that there's going to be some hard wiring of infrastructure today. We do want to presume that we better understand the characteristics of the data that's being created and operated on today. Does that make sense to you? Well, there's a lot of data and we're just going to have to find a way to not touch it or handle it any more times than we have to. We can't be shifting it around from place to place because it's too much. But I think the market is going to define a lot of that for us. So George, if we think about the natural place where the data may reside, the process may reside, give us a sense of what kinds of machine learning technologies or machine intelligence technologies are likely to be especially attractive at the edge, dealing with this primary information. Okay, I think that's actually a softball, which is we've talked before about bandwidth and latency limitations, meaning we're going to have to do automated decisioning at the edge because it's got to be fast, low latency. We can't move all the data up to the cloud for bandwidth limitations. But by contrast, so that's data intensive and it's fast. But up in the cloud where we enhance our models, either continual learning of the existing ones or rethinking them entirely, that's actually augmented decisions. And augmented means it's augmenting a human in the process where most likely the human is adding additional contextual data, performing simulations and optimizing the model for different outcomes or enriching the model. It may in fact be a crucial element or a crucial feature of the training by in fact validating that the action taken by the system was appropriate. Yes, and I would add to that actually that you use an analogy. People are going from two extremes where they say, some people say, okay, so all the analytics has to be done in the cloud. Wikibon and David Fleuer and Jim Cavillis have been pioneering the notion that we have to do a lot more at the client. But you might look back at client server computing where the client was focused on presentation, the server was focused on data integrity. Similarly, here the edge or client is going to be focused on fast inferencing and the server is going to do many of the things that were associated with a DBMS and data integrity in terms of reproducibility of decisions in the model for auditing, security, versioning, orchestration in terms of distributing updated models. So we're going to see the roles of the edge and the cloud rhyme with what we saw in client and server. Neither one goes away, they augment each other. So Jim Cavillis, one of the key issues there is going to be the gateway and the role that the gateway plays specifically here we're talking about the nature of, again, the machine intelligence that's going to be operating more on the gateway. What are some of the characteristics of the work that's going to be performed at the gateway that kind of has oversight of groupings or collections of sensor and actuator devices? Right, good question. So, for example, everybody's familiar with now about a gateway in this environment, a smart home hub. A smart home hub, just for the sake of discussion, has visibility across two or more edge devices, could be smart speaker, could be the HVAC system and sensor equipment and so forth. What it does, the only way to perform the smart hub any sort is that it acquires data from the edge devices. The edge devices might report all of their data directed to the hub or the sensor devices might also do inferences and then pass on the results of the inferences as data to the hub, regardless. What the hub does is, A, it aggregates the data across those different edge devices over which it has visibility and control. B, it may perform its own inference of data models that look across an entire home in terms of patterns of activity. And then it might take the hub, various actions on top of it by itself without consulting an end user or with the, you know, I mean, you know, might take actions in terms of beef up the security, adjust the HVAC, just, you know, adjust the laying in the house or whatever it might be. Based on all that information streaming in real time, probably if algorithms are aligned in the term, one of that data shows an anomalous condition that deviates from structural patterns. Those kinds of determination, whether it's anomalous or a usual pattern, are often taken at the hub level because it maintains sort of a homeostatic environment as it were within its own domain. And that hub might also communicate upstream to a tertiary tier that has oversight, let's say it's a smart city environment where everybody in that city or whatever might have a connection into some broader system that say regulates utility usage across the entire region to avoid drama and that kind of thing. So that gives you an idea of what the role of a hub is in this kind of environment. It's really a controller. So, Neil, if we think about the, if we think about some of the issues people really have to consider as they start to architect what some of these systems are going to look like, we need to factor both what is the data doing now, but also ensure that we build into the entire system enough of a buffer so that we can anticipate and take advantage of future ways of using that data. Where do we draw that fine line between we only need this data for this purpose now and geez, let's ensure that we keep our options open so that we can use as much data as we want at some point in time in the future. Well, that's a hard question, Peter, but I would say that if it turns out that this detailed data coming from sensors that the historic aspect of it isn't really that important, if the things you might be using that data for are more current, then you probably don't need to capture all that. On the other hand, there have been many, many cases historically where data has been used other than its original purpose. And my favorite example was scanners in grocery stores where it was meant to improve the checkout process, not have to put my stickers on everything, manage inventory and so forth. It turned out that some smart people like IRI and some other companies said, we'll buy that data from you and we're going to sell it to advertisers and all sorts of things. We don't know the value of this data yet, it's too new. So I would err on the side of being conservative and capturing and saving as much as I could. So what we need to do is we need to marry or we need to do an optimization of some form about how much is it going to cost to transmit the data versus what kind of future value or what kinds of options of future value might there be on that data? That is, as you said, a hard problem, but we can start to conceive of an approach to characterizing that ratio, can't we? Why hope so? I know that personally, when I download 10 gigabytes of data, I pay for 10 gigabytes of data and it doesn't matter if it came from a mile away or 10,000 miles away. So there has to be evidence for that. There's also ways of compressing data because this sensor data, I'm sure is going to be fairly sparse, can be compressed, is redundant. You can do things like RLL encoding which takes all the zeros out and that sort of thing. There are going to be a million practices that we'll figure out. So as we imagine ourselves in this schemata of edge, hub, tertiary or primary, secondary and tertiary data and we start to envision the role that data's going to play and how we conduct or how we build these architectures and these infrastructures. It does raise an interesting question and that is from an economic standpoint, what do we anticipate is going to be the classes of devices that are going to exploit this data? David Floyer, who's not here today, I hope you're feeling better, David, has argued pretty forcibly that over the next few years we see a lot of advances made in micropostor technology. Jim, I know you've been thinking about this to fair amount. What types of function might we actually see being embedded in some of these chips that software developers are going to utilize to actually build some of these more complex and interesting systems? Yeah, first of all, one of the trends we're seeing in the chips that market for deep learning just stay there for a moment, is that deep learning chipsets, traditional, and I say traditional, the last several years the market has been dominated by GPUs, graphic processing units. NVIDIA, of course, is the primary provider of those. What, of course, NVIDIA has been a long run for a long time as a gaming solution provider. Now, what's happening with GPU technology and the latest generation of NVIDIA's architecture shows where it's going. Does it embed more deep learning and optimize capabilities at the very chipset level? It's called tensor processing. I don't want to bore you with all the technical details. The whole notion- Oh no, Jim, do bore us. What is it? Deep learning is based on doing a high speed of fast matrix math. That's essentially it. So fundamentally tensor cores do high velocity, fast matrix math. And the industry as a whole is moving toward embedding more tensor cores directly into the chipset, higher density of tensor cores. I mean, NVIDIA and its latest generation of chips has done that. They haven't totally taken up the gaming worry of the GPU capabilities, but there are competitors and they have a growing list, more than a dozen competitors on the chipset side now, are all going down a road of embedding, are more tentacle processing units into every chip. I mean, Google is well-known for something called GPU processor processing units. They're chip architecture, but they're one of many vendors that are going down that road. The bottom line is so that the chipset itself is becoming authenticated, it's becoming optimized for the core functions that CPUs and really GPU technology and even ASIC FPGAs are not traditionally geared to do, which is just deep learning at a high speed, many cores to do things like base recognition and video and voice recognition, previously fast. And really that's where the market is going in terms of the enabling underlying chipset technology. However, what we're seeing is that what's likely to happen in the chipsets of the year 2020 and beyond, there'll be predominantly tentacle processing units, but there'll be systems on a chip that, and I'm just talking about the future, I'm just saying it here now, systems on a chip that includes a CPU to manage a real-time OS, like a real-time Linux or whatnot, and with highly dense tentacle processing units. And in this capability, these will be low-power chips and low-cost commodity chips that will be embedded in everything, from your smart phone, your smart appliances in your home to your smart cars and so forth, everything will have these commodity chips. And finally, every edge device, everything will be an edge device, and we'll be able to provide more than augmentation, automation, all these things we've been talking about in ways that are not necessarily autonomous, but it can operate with a greater degree of autonomy to help us human beings to live our lives in an environmentally contextual way at all points in time. All right, Jim, let me cut you off there, and because you said something interesting, a lot more autonomy. George, what does it mean that we're going to dramatically expand the number of devices that we're using, but not expand the number of people that are going to be in place to manage those devices? When we think about applying software technologies to these different classes of data, we also have to think about how we're going to manage those devices and that data. What are we looking at from an overall IT operations management approach to handling a geometrically greater increase in the number of devices and the amount of data that's being generated? Well, there's a lot of... Well, hold on, George. There's a couple of dimensions to that. Let me start at the modeling side, which is we need to make data scientists more productive or we need to push out to a greater... We need to democratize the ability to build models. And again, going back to the notion of simulation, there's this merging of machine learning and simulation where machine learning tells you correlations in factors that influence an answer. Whereas the simulation actually lets you play around with those correlations to find the causations. And by merging them, we make it much, much more productive to find models that are both accurate and then to optimize them for different outcomes. So that's the modeling issue. When you think about actually, which is great. When you think about some of the data management elements, what are we looking at from a data management standpoint? Well, and this is something Jim has talked about, but we had DevOps for joining the... Essentially, merging somewhat the skills of the developers with the operations folks so that there's joint responsibility of keeping stuff live and... But what about things like digital twins, automated processes? We talked a little bit about breadth versus depth, I, Tom, what do you think? Are we going to build out or are all these devices going to reveal themselves, or are we going to have to put in place a capacity for handling all of these things in some consistent, coherent way? Oh, okay, in terms of managing. In terms of managing. Okay, so, all right. So digital twins were interesting because they pioneered a, or they made well-known a concept called essentially a semantic network or a knowledge graph, which is just a way of abstracting what is a whole bunch of different data models and machine learning models that represents the structure and behavior of a device. In IIOT terminology was like an industrial device like a jet engine. But that same construct, the knowledge graph and the digital twin can be used to describe the application software and the infrastructure, both middleware and hardware, that makes up this increasingly sophisticated network of learning and inferencing applications. And the reason this is important, it sounds arcane, the reason it's important is we're building now vastly more sophisticated applications over great distances. And the only way we can manage them is to make the administrators far more productive. The state of the art today is alerts on the performance of the applications and alerts on the essentially the resource intensity of the infrastructure by combining that type of monitoring with the digital twin, we can get essentially a much higher fidelity reading on when something goes wrong, we don't get false positives. In other words, you don't have, if something goes wrong, it's like the fairy tale of the P underneath the mattress, all the way up 10 mattresses, you know that it's uncomfortable. Here, it'll pinpoint exactly what gets wrong rather than cascading all sorts of alerts. And that is the key to productivity in managing this new infrastructure. All right guys, let's go into the action item around here. What I'd like to do now is ask each of you for the action item that you think users are going to have to apply or employ to actually get some value and start down this path of utilizing machine intelligence across these different tiers of data to build more complex manageable application infrastructures. So Jim, I'd like to start with you. What's your action item? My action item is related to what George has said. Model centrally, deploy in a decentralized fashion, machine learning, and use digital twin technology to do your modeling against device classes in a more coherent way. Now, one model will fit all of the devices. Use digital twin technology to structure the modeling process to be able to tune a model to each class and device out there. George, action item. Okay, recognize that there's a big difference between Edge and Cloud. As Jim said, but I wouldn't elaborate. Edge is automated low latency decision making, extremely data intensive. Recognize that the Cloud is not just where you trickle up a little bit of data. This is where you're going to use simulations with a human in the loop to augment Systemwide. Systemwide, with a human in the loop to augment how you evaluate new models. Excellent. Neil, action item. Action item. Well, I would have people start on the right side of the diagram and start to think about what their strategy is and where they fit into these technologies. Be realistic about what they think they can accomplish and do their homework. All right, great. Let me summarize our meeting this week. This week we talked about the role that the three tiers of data that we've described and will play in the use of machine intelligence technologies as we build increasingly complex and sophisticated applications. We've talked about the difference between primary secondary and tertiary data. Primary data being the immediate experience of sensors. Analog being translated into digital about a particular thing or set of things. Secondary being the data that is then aggregated off of those sensors for business event purposes so that we can make a business decision often automatically down at an edge scenario as a consequence of what the signals that we're getting for multiple sensors. And then finally tertiary data that looks at a range of gateways and a range of systems and is considering things at a system-wide level for modeling, simulation, and integration purposes. Now, what's important about this is that it's not just better understanding the data and not just understanding the classes of technologies that will be used, that will remain important. For example, we'll see increasingly powerful, low-cost, device-specific, arm-like processors pushed into the edge. And a lot of competition at the gateway or at the secondary data tier. It's also important, however, to think about the nature of the applications and where the work is going to be performed across those different classifications, especially as we think about machine learning, machine intelligence, and deep learning. Our expectation is that we will see machine learning being used on all three levels or machine intelligence being used and against all forms of data to perform a variety of different work, but that the work that will be performed will be naturally associated and related to the characteristics of the data that's being aggregated at that point. In other words, we won't see simulations, which are characteristics of tertiary data, George, at the edge itself. We will, however, see edge devices often reduce significant amounts of data from, perhaps, a video camera or something else to make relatively simple decisions that may involve complex technologies to allow a person into a building, for example. So our expectation is that over the next five years, we're going to see significant new approaches to applying increasingly complex machine intelligence technologies across all different classes of data, but we're going to see them applied in ways that fit the patterns associated with that data because it's the patterns that drive the applications. So our overall action item, it's absolutely essential that businesses start considering and conceptualizing what machine intelligence can do, but be careful about drawing huge generalizations about what the future of machine intelligence is. The first step is to parse out the characteristics of the data driven by the devices that are going to generate it and the applications are going to use it and understand the relationship between the characteristics of that data and the types of machine intelligence work that can be performed. What is likely is that an impedance mismatch between data and expectations of machine intelligence will generate a significant number of failures that often will put businesses back years in taking full advantage of some of these rich technologies. So once again, we want to thank you this week for joining us here on the Wikibon Weekly Research Meeting. I want to thank George Goldberg, who's here in our CUBE studio in Palo Alto and Jim Cabela's and Neil Raiden, who are both in the phone. And we want to thank you very much for joining us here today. And we look forward to talking to you again in the future. So this is Peter Burris from the CUBE's Palo Alto studio. Thanks again for watching Wikibon's Weekly Research Meeting.