Why we need AI for the future of IoT: connecting trillions of smart devices in ultra-dense networking environments

city wifi AI ultradense

Can AI help us connect trillions of smart devices? There are currently perhaps 20 billion devices connected to the internet: things like laptops, phones, smartwatches, TVs, smart speakers, smart home devices …

In a decade, that could be 50 billion … and a lot of it is enterprise IoT.

In this edition of the The AI Show — crossposted into the TechFirst with John Koetsier podcast —  we chat with Intel and the National Science Foundation, which has funded $30M+ into projects to use AI to figure out how we’ll manage ultra-dense wireless networks … how we’ll keep it secure, and how we’ll keep everything connected.

Joining me in this episode:

  • Vida Ilderem, VP, Intel Labs
  • Thyaga Nandagopal, National Science Foundation
  • Pu Wang, University of North Carolina at Charlotte

Listen: Using AI to connect trillions of IoT devices

Don’t forget to subscribe to TechFirst with John Koetsier wherever podcasts are published:

Watch: Using AI to connect trillions of IoT devices

Subscribe to my YouTube channel so you’ll get notified when I go live with future guests, or see the videos later.

Read: Using AI to connect trillions of IoT devices

John Koetsier: Can AI help us connect trillions of smart devices?

Welcome to The AI Show with John Koetsier.  Right now, there are maybe 20 billion devices connected to the internet. It sounds like a lot, right? Laptops, phones, watches like that, smart devices that we wear, smart devices in our homes, smart TVs, smart speakers, and of course, industrial IOT. I learned a little bit about this myself just in the smart home space about a month or two ago when we changed our WiFi, and for about 30 days after that I was changing devices everywhere that had to get updated.

In a decade, however, those 20 billion devices could be 50 billion, and a lot of that is enterprise IOT. How will we manage that? How will we keep that secure? How will we keep everything connected? 

Vida Ilderem, VP at Intel Labs

Vida Ilderem, VP at Intel Labs

So Intel and the National Science Foundation have just funded research into this, it’s Machine Learning for Wireless Networking Systems.

And to learn more, I’m going to invite three people who are directly involved in the project. We’ve got Vida Ilderem who’s the VP at Intel Labs, and she’s also director there. We’ve got Thyaga Nandagopal from the National Science Foundation. And we’ve got Pu Wang from the University of Northern Carolina at Charlotte, I believe that’s North Carolina at Charlotte.

Welcome to The AI Show everybody! I think this is a bit of a record, this is the most people we’ve ever had at any one time on The AI Show, so it’s a record. Welcome!

Thyaga, let’s start with you. We’re talking about billions of devices here and you’re doing things around ultra-dense wireless systems. What does that mean? Can you clarify that? What is an ultra-dense wireless system? How many devices are we talking about, what’s the density? 

Thyaga Nandagopal: Right. So just for context, your traditional 4G networks that your current mobile devices rely on typically can support a region that has about 300 to 2,000 devices in their coverage area. But at any given time, maybe 20 to 40 devices are active at any given time. So that’s the current density that we play with right now.

Which is why sometimes when you go to a stadium your network doesn’t work because suddenly everybody’s using their device at the same time, and even if it’s only hundreds of people you’re having trouble. We are talking about future networks where we are thinking about device densities — think of what a stadium, a covered sports stadium, you are going to have tens of thousands in a small region. That’s the kind of densities we are going to start from and go all the way up to millions of devices in a coverage area of a single cell site in a wireless network. 

John Koetsier: Millions of devices and one single cell site. Wow! Talk about, I mean, that obviously seems challenging just from a bandwidth perspective, but I’m sure there’s other issues as well. What are some of the core challenges of networking in an environment like that? 

Thyaga Nandagopal, National Science Foundation

Thyaga Nandagopal, National Science Foundation

Thyaga Nandagopal: So think of us, all of us, the four of us sitting here. So if any one of us have talked to you, for example, John, I just raised my hand and you can look at me and say, ‘Okay Thyaga, go ahead and talk first.’ Imagine if you are in a crowded room of 10,000 people, everybody’s raising their hand ‘Me, me, me, me, me!’ Just you, John, trying to figure out who should I point to and who should I give an option to speak? You’d be struggling to figure that out, right? And you have to be fair, you have to make sure that everybody gets a chance to speak. Just doing that can stress you so much. And that’s the challenge, the biggest networking challenge that we have in the future generation of networks. 

John Koetsier: Wow.

Thyaga Nandagopal: Just for the base station to decide who should I give an opportunity to talk and make sure that everybody gets a chance to speak in a given amount of time so they can get their data across is one of the biggest networking challenges.

John Koetsier: And that’s just the first step. That’s just the first, I mean, smart devices will often ping a network and say, ‘Yeah, I’m still here, yeah, I’m still here.’ Right? And sometimes we see noisy devices that are doing that often, right, like hourly or even more frequently, and so we get upset about those a little bit. But Vida, what do we need to learn about networking in those environments?

Vida Ilderem: So one of the challenges is you and I, as humans, when we are on a call or something we want to have continuous sessions. We don’t want it to be dropped, or we don’t want to see that sun going on our buffering, buffering, buffering, right? So when we talk about very dense areas, as Thyaga indicated, we have to serve not only people but also objects. So one of the fifth generation 5G is doing is to bring in a communication, addressing the things-to-things communication.

So you’re going to have very heterogeneous traffic in your network, you have human-to-human traffic which could be video, is the most application which is being used today, workload. And then you have to have the capacity, your network has to have capacity to address this need. And so managing the traffic for the different loads — a load from a think device is very different than a load from, for instance, what you’re doing right now streaming video. And the other thing is we need the thruput, we need high data rates, which means we get our information faster. Also, the stadium is a perfect example, when you have so many wireless devices there and there’s so many air waves going around you get lots of interference, right? So how do you manage that interference? So the heterogeneity of the network and the number of services you have to address is becoming very complex for our next generation wireless networks. That’s what we are facing. 

Pu Wang, University of North Carolina at Charlotte

Pu Wang, University of North Carolina at Charlotte

John Koetsier: And there’s different challenges. You talked a little bit about that. There’s different challenges for different things, right? We’re using a high bandwidth scenario right here where we need continuous contact and there’s video going both ways and audio, so we need that to stay good. Whether it’s these things that ping in occasionally, or a little bit here, or somebody loads a mobile webpage, they need a burst of bandwidth and then it shuts off. I mean, this is a really, really complicated thing to solve it sounds like. 

Thyaga Nandagopal: Yes, indeed. 

Vida Ilderem: It’s true, and that’s why this collaboration with NSF is so important in this research area, because you have the discipline of communication, computation, and AI coming together to address these challenges actually.

John Koetsier: Let’s talk about that a little bit. Let’s talk about the projects that you awarded, maybe give a little bit of context on the awards and what you’re trying to accomplish. And then maybe give us a bit of a rundown on some of the key projects that you’ve awarded. 

Thyaga Nandagopal: Right. So, we traditionally have been funding a lot of work on networking, much of the work that you see in 4G today have been realized through NSF research that has been funded in the past and we are far more than that. So we have been continuing along the path and we have been investigating a lot of 5G for a very long time. As part of that, one of the, what we just heard about the complexity of the network and the challenges of managing the network the decision-making process in real time, has essentially spurred us to think about — and the research community that is working on these problems to think about — can we use artificial intelligence techniques, AI and machine learning techniques to help take the human out of the loop and make intelligent decisions on the fly?

So sometimes additions have to be made using data that we have never seen before, that a human just cannot learn in time to make the right decisions. So that’s one of the challenges that the community has been wrestling with for a while. And we have been seeing a steady stream of projects that have been seeking to, you know, maybe how come we can’t use AI for these problems? 

So at the same time we have, Intel and NSF, have been working together for many years now, almost five years now. Intel came to us and said, ‘Look, there’s no VC, this is a pressing challenge that has immediate impact in the 5G sphere. Would you be interested in working with us and we would love to do a joint program on this.’ Again, just want to point out that Intel and NSF has five or six programs right now.

I brought before this particular effort and we said, this is perfect, there’s a continuation of this partnership, let’s continue, let’s do this. So we created this targeted effort as a special call, which is on top of everything that we normally fund. But specifically addressing machine learning and its applications for optimizing network systems, and network systems could be wired and wireless networks, and the edge networks where both of them kind of come together.

And so this is the name of the program, Machine Learning for Wireless Network Systems. So ML-WNS, you know, it’s a pun, right? It’s a well-designed pun, I would like to say. And one of the interesting things that we anticipated was we received an overwhelming amount of interest from the research community which kind of, in some sense, was a very good surprise for us. We are glad to hear that there are so many people who are willing to challenge so many of these problems.

We’ve got so many good projects and I think our Intel partners and us are extremely thrilled to see the slate of ours that are coming out of this, which I think we can talk more, but they cut across multiple teams,right? So we are talking about using machine learning to optimize the design of the wired network, the core network, that is the wired part of it that you don’t normally see. The talking about managing the spectrum that is used by radios to communicate with the base stations and other devices, and then the devices itself, right, the data that the devices send themselves.

So if, when multiple devices are sending data and let’s say these are sensors, how can you learn on the fly without having to send all the data back to a central server? Can you learn efficiently, there’s quite a cost of distributed learning. So it cuts across multiple teams and I think there are much more to explore. We wish we could continue doing more or not put more money into this, and we will continue to do more in the future, but this was an extremely promising start for bringing AI and automated, you know, autonomous management of networks into the network management and wireless networks sphere.

John Koetsier: Excellent. And we’re going to get into one of those projects pretty soon. We’re going to bring Pu Wang into this conversation. I’ll just ask you, you know, one of the challenges that’s probably topical and top of mind for you is okay, we’re in this dense environment, there’s a lot of signals flying around, now there’s an emergency and the police need to communicate, or an ambulance, or whatever, fire services, that sort of thing. Are you thinking of ways of prioritizing traffic like that? 

Thyaga Nandagopal: So network prioritization is one of the key topics in this thing, right? So how do you determine, automatically prioritize, and isolate traffic to provide service guarantees is an essential piece of any network optimization problem. And we believe that there are some projects that actually are trying to address that. Again, you know, it’s not always necessarily first responders because that’s an easy example to grasp, right? 

John Koetsier: Yes. 

Thyaga Nandagopal: But more often than not, it’s you want to send a tweet that needs to go out, is probably more important, or if you want to have a video conversation that you want to have, you want to have that traffic prioritized more than somebody who’s trying to watch a YouTube video. It’s more of a one-way communication which is not necessarily as critical. So these decisions, right? The first responder scenario amplifies it, makes it very crisp, but you know, we, the networks have to make decisions on the fly all the time. 

John Koetsier: Yes. 

Thyaga Nandagopal: And that prioritization is somewhere where artificial intelligence can excel because it kind of can sense the traffic patterns and it can see traffic flow shifting. So for example, one can argue that if there’s an emergency, yes, you have to prioritize the first responders, but you may also want to prioritize people who are tweeting about it because that’s information for the first responders to learn from. 

John Koetsier: Interesting, interesting.

Thyaga Nandagopal: Right? Human probably would not make the decision.

John Koetsier: Yes. 

Thyaga Nandagopal: How our AI can sense the fact that there’s lots of events of interest, and therefore we need to prioritize traffic around the area of interest and automatically make the cells, you know, capacity from other cells focus them, focus the energy from the surrounding cells into that region. So the structure of the network itself changes dynamically really fast to prioritize more traffic in that area so the tweets can get out, the videos can get out, and the first responders can also try and communicate with people on the ground. So that’s something AI can do really well. And we are excited to enable that kind of solutions with this program, which we cannot do today at current network management solutions.

John Koetsier: Interesting, interesting. And Pu Wang, obviously we need that intelligence dispersed, right? We need that intelligence close to where things are happening, right? We need that intelligence in our wifi nodes and networks, and our cell towers, and other things like that. What’s your project? It sort of looks like mesh networking.

Pu Wang: Yes. Somehow it is related to mesh networking. So overall, my project is to develop smart wireless multi-hub networks, such as mesh networks, to enhance the AI performance of actual devices, such as cell phones, like robots, auto-driving cars. So that’s the overall goal of my project.

John Koetsier: Interesting. And talk about what we need AI to do there and what you’re using it for. 

Pu Wang: So basically there are two types of AI technology we are using for this project. One is called the Fidelity Learning, another one called the Reinforced Learning. So the Fidelity Learning is, basically it’s emerging distributed machine learning paradigm to improve data privacy of compare with a traditional centralized machine learning. So when we talk about AI, basically we need lots of data to train the AI models. The data include personalized data. For example, like when you are using cell phones, like we have this smart keyboard from Android or iPhone which can give you the next word prediction, or sometimes also kind of help you to finish the whole sentence, right?

So actually behind that technology is the AI model. So in order to train the model, generally we need to collect data from all the users and put all the data in essential server. So that’s increased or raised a lot of privacy concern because people do not want to share the personal data. Definitely you do not want the server to see oh what you are typing in the past, right? So how to solve this problem, which basically it’s one of the key challenge faced by AI community. So now the final learning is one of the solutions, which kind of drives the privacy issues. So the key concept is that we let many, many edge device like your cell phones, to tune a global shared model by using their local …[section not audible]… so that we can keep the data local and private, but we still can worry and create a high performance global model. So that’s how we can make the edge devices smarter and smarter. However, we can keep the data private and secure. 

So another technology we [are] trying to do is using Reinforced Learning. So that’s actually, it’s regarding how we can develop some wired mesh network to actually to make the edge devices to be coordinated in a much more efficient and timely manner. So wired mesh network basically is like, is a low-cost communication infrastructure or paradigm which use lots of large routers that are connected to each other using wireless links, instead of wiring such as fibre optical cables. Also, wireless mesh network actually can provide much more low-cost and efficient networking services for a large population of people, including those people living in the low-cost community, or low-income community, or underdeveloped regions, or rural outliers. So the reason, one other example is the SpaceX Starlink satellite network, right? So SpaceX actually launched thousands of small satellites which can connect with each other to form a wireless mesh network which can provide internet service to everyone on earth. So now we actually try to utilize the mesh network to coordinate the training process of the edge devices so that we can actually democratize AI by making AI accessible to everyone in the low-cost manner, including the people, as I said, living in the low-cost or low-income regions.

John Koetsier: Interesting.

Pu Wang: So then the reinforced learning actually can teach the network to make smart and optimized decisions so that we can improve or optimize the performance of the AI models by using the minimal training time. 

John Koetsier: Okay, okay. Very, very cool, good stuff. Vida let’s turn to you. We’re getting more and more smart devices in our homes, offices, factories are getting them as well. And currently, you know, they may have local smarts, but they’re reaching out to a network, they want to send some data to the cloud or other stuff like that. They just kind of reach out blindly, expect that there’s a network that they can attach to, are we going to make those smarter as well? How will AI change the picture there? 

Vida Ilderem: A very good question, John. So the way you look at it, the first instantiation for AI is basically efficiency, right? So if I have a device which is very power-constrained, do I really need to keep pinging the tower when I’m not near a tower? Right? So we can use that intelligence to preserve your battery power, right, your energy. That’s at the device end and on the network end, you know, the same thing. How can you boost the efficiency by the backend office to make it more productive? And you can have your return on investment on it. The other side, as Thyaga mentioned, is this whole work-load balance because the type of services that are coming online. So we talk about very diverse and divergent requirements, you know, you have a Twitter feed, or you have a video feed, or you have a parking meter waking up and sending a kilobit of data and going back to sleep, right? So how do you manage that so your network has the best performance to offer you the best quality of service and experience for the application you need?

The third important part is the management of spectrum. Spectrum is a very scarce resource. It’s a very expensive resource. Why this is, you know, you pay sometimes billions of dollars for this spectrum, right? So you have licensed spectrum which is a control, you have unlicensed spectrum like wifi. So your cellular base, your LTE, 5G are licensed. Your wifi, Bluetooth are unlicensed. And then you have shared spectrum. So what AI can help is if the spectrum is not used, maybe intelligently, it can access that spectrum for your managing the workload and the traffic in the pipe, if you may. So it brings the intelligence in there and you have, you know, every network has multiple spectrum states right? You have your LTE/4G. Now with 5G, you have the new bands opening. So there are many, many spectrum waves going on. So the idea is also for AI to manage interference between these different spectrum accesses and usages. So there are many, many ways AI can step in and improve the next generation services, we dynamically adjust the spectrum for optimal performance, and of course improve the efficiency of devices and the network on both ends. 

John Koetsier: Interesting. And maybe spectrum hop to find a place where you can get communication and maybe say, “Hey, wifi is busy, I’ll instantiate some Bluetooth mesh networking or something like that dynamically …

Vida Ilderem: That’s right. 

John Koetsier: … and go for it, right? Does that require AI-specific chips? I mean, what are you doing there for on-device intelligence? 

Vida Ilderem: So, depends, that’s the question. So if you are on battery operated devices, which are like set down, forget in the field or/and you want to process the data at the source, either you collect the data so you need a very low power, you know, efficient. So maybe you need a specific chip for AI to do that intelligence at that node, but your node could be a remote roadside unit which you can use. Higher use requires a higher performance, like, you can use your high-performance CPUs or GPUs, general purpose GPUs, and in that case, you can say, hey, I don’t need a very specific, what we have currently can be applied towards AI. So it depends, there’s a range of this. So if you’re constrained for power or form factor, that type of thing, you probably move more towards  ASIC or a specialized solution. If you’re looking at bigger workloads towards cloud and edge, then I think some of the existing solutions may be sufficient. They have to be optimized for that workload. 

John Koetsier: Very, very interesting. Maybe I’ll turn this question first with Thyaga, but others are welcome to jump in on this one. We’ve seen that Smart Home sometimes isn’t so smart. And sometimes smart devices as well, inside companies, have been vectors for attack. I believe a couple years ago there was an automated system in a fish tank that was a vector for attack in a company and it was like a smart home device, smart office device, and that was just the way that the hackers got in. We’ve seen a lot of that in Smart Home with cameras and other things like that, bad security. What are the particular security challenges of very dense ecosystems if we move to 50 billion, a hundred billion devices globally? I mean, right now I might have, I don’t [know], it might be 40 or 50 smart devices in my home, and in the future it could be everything — every window is smart, every surface is smart, every wall knows where it is, you know, the roof and the gutters know when there’s water coming down, other things like that. You essentially have a smart building envelope and interior. It’s hard to count discrete elements of intelligence in there but it could be hundreds easily. What are the challenges of security in that scenario? 

Thyaga Nandagopal: So, do you want me to scare your audience completely? 

John Koetsier: Absolutely! 

Thyaga Nandagopal: So, I think there are numerous problems. I think part of this has to do with scale. Part of this has to do with the fact that a lot of our devices are increasingly drawn by software and we all know software is inherently buggy, right? I mean whether we like it or not, that’s the reality of the situation.

And also it comes from the fact that they’re hydrogenous devices. It’s no longer, you know, we no longer use devices provided by a single vendor who controls everything. There are a multitude of devices manufactured and operated by many different people and it’s very hard to kind of know everything that’s happening in the network. And of course the last and most important one is the human factor. Right? You know, ultimately security comes from the fact that you can trust an individual to do the right thing, and when you no longer can trust the individual there’s nothing really you can do. So, those are like the macro, right? Those are the macro issues. But if you look at the real challenges, number one, when you have millions of devices or even thousands of devices that are operating, traditional hacking and attacks have come from a single malicious device.

Now, what is really worrisome to security experts is what we call as low persistent threats, right? So these are things where a small number of many device — I mean, sorry, a large number of devices that are doing small things here and there that are not noticeable by a human being, right? Now I send a packet here, you send a packet somewhere else at different times, and together you sneak out, you sneak and manage a big attack on the system. I mean, a simple analogy is a distributed denial of service, right? You know, we started out denial of service, which is a single device pumping a lot of traffic. We know how to handle it. Now we have millions of devices pumping traffic all tied to the single server, you cannot stop any one of them. It’s much harder to solve, right? Taken to the extreme, you could have thousands of IOT devices, launching different kinds of attacks. Now they could all be asking the base station, ‘Hey, I want to talk, I want to talk, I want to talk,’ drowning out anybody else who really has a real purpose to talk. As simple as that, right?

They could all be sending, saying that, ‘Oh, I am done, my network is bad in this area,’ you know, ‘I can’t get your signal’ and immediately the base station thinks that something is wrong with the signal propagation. So straight, I just automatically cause a starvation somewhere else. So let’s take a bunch of what you call this, in the aggregate when multiple of devices work in concert in a malicious manner. That can overwhelm anything that we know how to defend against it, right? This is what I call as a “collective behavior.”

Now, of course, different vendors will have different IOT devices. Each of them may have their own software vulnerabilities and any one software creating a vector of attack suddenly leads to you having to figure out where did this attack come from? The network operator has no idea, right? Why is this device which was working fine until yesterday, suddenly behaving? Is it behaving badly because it has misconfigured the owner, the person who owns this device accidentally misconfigured the settings to keep sending too much data? Or is it behaving maliciously? How do I as operator decide, or do I decide to shut this off and cost the company and the customer to get upset? Or do I kind of figure out a way to call them up and say, ‘Hey, what’s going on?’ How do we alert them? So the operator has to make these decisions. And again, if you look at it, the scale becomes much harder. If it’s 10 devices, 20 devices, it’s easier to pinpoint and kind of isolate it. You’re talking about millions of devices, now this becomes a problem. 

John Koetsier: Yes. I can totally see that. 

Thyaga Nandagopal: Exactly. And then suddenly you also have to worry about legacy devices. Some of these devices have to be supported for a long time in the network. Let’s say I’m putting in an IOT sensor that is going to be embedded on my road or in the bridge that is supposed to be transporting the condition of the bridge for 20 years, right? Fast forward 10 years and suddenly the sensor is either behaving as designed, or is behaving or stopped working, or it’s suddenly sending some random stuff. Now, how do I go figure this out? There’s nobody who even understands what software is running on on that sensor anymore, right?

John Koetsier: Yes.

Thyaga Nandagopal: Now, so that’s other problems. So a lot of it has to do with secure software development, number one. Of course the standards themselves have to be secure, right? Because now increasingly we are moving into more of a software-driven ecosystem with the 5G and whatever comes after 5G. 4G and previous generations used to be much more hardware-driven, now we are going software-driven. So the inherent challenges with all the software bugs that we all face every day move into the network domain very fast.  And of course, then there’s a human element of it because how do we as humans, you know, handle security challenges when we can’t even communicate with the device.

So traditionally until now, our communication has been human centric. We talk to other humans, or the humans are at least on one end of the communication, right? You are either consuming a video or you are sending something. But now what Vida pointed out is different machines are going to be sending traffic across each other. Who are you going to talk to, to figure this out? Right? So are you going to say, ‘Hey machine? I don’t, what are you doing?’ There’s no one to tell you, the machine doesn’t know how to talk to you. So we need to have a language to figure out how to communicate and troubleshoot a machine — which we don’t have yet. So these are some of the challenges we would have to solve to make these networks robust and secure. And again, AI plays a very important role in this. 

John Koetsier: I can totally see that, and especially in security and understanding big patterns that are invisible to a human about what might be going on. Do we have a reliable estimate today of what percentage of network utilization is machine-to-machine traffic and what that might look like in 10, 20 years? 

Thyaga Nandagopal: I’m going to defer to Vida because … 

Vida Ilderem: I was going to say, I don’t have that number, sorry.

John Koetsier: That’s okay.

Thyaga Nandagopal: Yeah. Any total numbers have suggested that the number is rising, right? I remember a number from some a couple of industry reports that talked about the number being around 10 to 15% in 2018, but it was growing at an annual rate. I mean, they’re projecting that to be 50% or so by 2025.

John Koetsier: Wow. 

Thyaga Nandagopal: So that’s a huge number, right? I mean, it’s growing pretty fast. Again, a lot of these are projections, of course, right? 

John Koetsier: Yes.

Thyaga Nandagopal: It all depends on how much we see IOT become prevalent and pervasive and how much of it really relates on a cellular network and equity. A lot of these devices still use wifi and for that part, we all know that your Roku devices and your home devices talk a lot in the background, even when you’re not watching a video or … 

John Koetsier: Very talkative.

Thyaga Nandagopal:  They are talkative, right? Now we don’t know what’s going to happen with the cellular network, will they continue to do so or not, but some of the projections are essentially showing rapid growth in that model. 

John Koetsier: Very, very interesting. Well, thank you all for participating in this. I much appreciate it. 

Vida Ilderem: Thank you so much. 

John Koetsier: Excellent. For everybody else, thank you for joining us on The AI Show, whatever platform you’re on please like, subscribe, share, comment. If you’re on the podcast later on, hey, rate it, review it. Thank you! Until next time, this is John Koetsier with The AI Show.