Episode 91: A Conversation with Mazin Gilbert

Guest

Dr. Mazin Gilbert is the Vice President of Advanced Technology and Systems at AT&T Labs. He leads AT&T’s Research and Advanced Development of its network and access transformations. In this role, Mazin oversees advancements in artificial intelligence, software-defined networking and access, digital transformation, cloud technologies, open source software platforms and big data.

Mazin holds 176 U.S. patents in communication and multimedia processing and has published over 100 technical papers in human-machine communication. He is the author of the book titled, “Artificial Neural Networks for Speech Analysis/Synthesis,” 1992, and an editor of a recent book on “Artificial Intelligence for Autonomous Networks,” 2018.

With more than three decades of experience under his belt, Mazin’s previous work includes Bell Labs, BBC and British Telecom. He’s also worked in academia at Rutgers University, Princeton University and Liverpool University. He became an IEEE Fellow in 2012. Mazin earned a bachelor’s and a doctoral degree, with first-class honors, in electrical engineering from the University of Liverpool. He also earned an MBA for Executives from the Wharton Business School of the University of Pennsylvania.

Outside of his technology career, Mazin is an entrepreneur owning five limited liability companies specializing in commercial and residential real estate and the dental industry. He also serves as a Chair of the Linux Foundation Deep Learning Foundation board, and a board member at the International Computer Science Institute (Berkeley). Mazin loves to spend time with his daughters and is an avid runner.

Transcript

Byron Reese: This is Voices in AI brought to you by GigaOm. I'm Byron Reese. Today my guest is Mazin Gilbert. He's a VP of AT&T Labs and their advanced technologies. He holds a PhD in electrical engineering from Liverpool John Moores University, and if that weren't enough, an MBA from Wharton as well. Welcome to the show, Mazin. 

Mazin Gilbert: Thank you for the invitation, Byron.

I always just like to kind of start out talking about what intelligence is and maybe little different [question], like, why do we have such a hard time defining what intelligence is? Yeah, that's where I'll start.

We always think of intelligence, certainly machine intelligence... we always compare machine intelligence to human intelligence, and we sometimes have a challenge in equating machines to humans. The intelligence of machines are radically different than humans. The intelligence is basically the ability to perform functions that may, one, be superior to any basic system to do; or two, require some form of context, some form of interpretation, some form of prediction that is not straightforward to do.

In machine intelligence, we really use that for anywhere from its basic form that could be as simple as moving data from one place to the other, all the way to its most advanced form: to be able to process petabytes of data to tell us how to best optimize traffic in our network. Both of those forms of intelligence, the most basic form and the most intelligent form, [are] absolutely essential to running a communication network.

But I mean, why do you think AI is so hard, because we have a lot of people working on it a whole lot of time, we've got a bunch of money in it, and yet it seems that we still don't have machines able to do just the simplest, most rudimentary, common-sense things. I haven't ever found an AI bot that can answer the question: ‘what's bigger—a nickel or the sun?’ Why is that so hard?

I think we segregate AI into two classes. One class of AI are sort of rule based systems. So these are expert systems that we've been using as a society for decades. These are rudimentary bots. We actually have over 1500 of those deployed in AT&T. They do the rudimentary tasks, think of ‘if-then’ type of statements. They are very basic, but they do some amazing jobs in automating functions that otherwise humans would have to do at a scale, and there's not enough humans to do those jobs in some cases.

Where it gets harder to understand is this sort of new wave of AI, of machine learning, deep learning based AI. Those are harder to understand because people equate those to some robot having the intelligence of a human, thinking like a human, making decisions like a human, and those don't really exist today. And even what exists today are still in their rudimentary early forms. The machine learning type of AI that exists today, even in deployments (and we have a bunch of those already), the reason they’re hard is because they are very data driven. That's the basic concept of an AI machine learning system today, data driven.

We deployed our first commercial AI system for customer care in about in 2000 called ‘How May I Help You,’ and then we had to go collect large amounts of data from our call centers to do the most basic thing. And as a result, there's only a few of these systems you can build that if you have to go and collect large amount of data and have this data checked, evaluated, labeled by humans, which could take weeks, months, years, so the assistant can learn and do a function, that makes it really hard. So even when you think about for the most largest and commercial deployments today of AI, the Siri and Alexa and others, there are hundreds, if not probably thousands of people behind that...

But that just kind of kicks the can down the street a little, doesn't it? I guess, then I would say, “Why is building an unsupervised learner so hard?” Why haven't we been able to just make something that you could point at the internet, it can crawl around and it can sort it out? Why do you think that's so hard?

So the concept of generalized artificial intelligence, which means that you build intelligence in a system and that system can do anything you want, it can classify internet traffic, it can recognize what you say, it can tell you what kind of an image – this is a cat or a dog, those systems do not exist, not in research, not in any commercial arena, they don't exist.

What exists today are systems that have been developed, trained by humans to do one narrow function, and those systems are not easy to develop, because of the concept of: not only you need to collect large amounts of data, you need to teach the system what is the truth and what is the right action to take. I think of them as babies. You don't train a baby in two hours or overnight. You don't. It takes years to train a baby with a lot of feedback and it also provides feedback and sometimes supervised feedback on what is right, what is wrong, what is a picture, what is not a picture, what's a word, what's not a word, how to pronounce something.

That's sort of what we need is that these systems require years of data collection with a lot of supervision and knowing the truth (just like any baby) for them to even get close to understanding and operating a simple function.

Right, but to use the baby analogy, maybe a baby takes years, but if you just measure the number of cycles, it might be thousands or tens of thousands, what should take a computer a minute to do, right?

Yeah. That is correct.

Why isn’t it that we’ve had a just maybe super fast baby learner by now? I guess, it's kind of like you keep telling me what the problems are, but what I'm trying to get at is why is intelligence... like the thing that we are most aware – we are intelligent and we can think about our own intelligence, and yet we don't even have a consensus definition of what it is. We don't know how to build it. And so from that standpoint, it makes me feel like maybe this brute force...

So machine learning is a really simple idea. It says ‘let's take information about the past and study it and make projections into the future, and maybe that only works on a very narrow set of use cases. Maybe that really isn't intelligence at all, and it's just faux intelligence, it's just we've managed to mimic intelligence, but we haven't built anything that has any actual intelligence whatsoever, is that possible?

We like to think that the intelligence is the reason why it may be hard to describe it here, because in the machine learning AI space, intelligence is embedded in the data. So if I show you two pictures, a cat and a dog, you look at them and say, "Well, this is obviously a cat, this is a dog." Maybe a baby would not know that until they're trained to do that, but the question is, ’How did you know one is a cat and one was the dog?’ There's clearly... you acquired some data, visual data, you processed this data, and somehow in your brain, in your cells, in your neurons, you decided to do some firings that determined one is a cat, one is a dog.

So your intelligence is embedded in some neural systems and this is what we're trying to replicate. We may not be able to write intelligence in a formula—that's not how we do it in machine learning, we don't know how to write it in a formula. But we are able to do: think of neural nets or deep learning, which got a tremendous amount of attention, which a lot of people say has some resemblance to biological neurons (may not be exactly true), but at the end of the day, there are a bunch of neurons, millions and tens of millions of these neurons and they learn intelligence or inferences based on data.

So you throw it a lot of data, a lot of pictures of the cats and the dogs, and you say, "Well, this is a cat and this is a dog." And then these systems are then able to take arbitrary pictures of new cats, new dogs that it's never seen before and are able to classify them very accurately. So somehow these systems learned that form of hidden intelligence that was hidden in the data.

What these systems are not be able to do however, is to tell you how I've learned what a cat and what a dog is. I can't tell you, these systems cannot tell you what a pig looks like, they cannot classify a pig from a horse, whereas humans can do that. That generalization—humans are very good at that. That sophisticated generalization called generalized AI does not exist today.

Well, so that's an example of transfer learning where humans are really good at taking something you learn in one sphere and apply it in another one, and maybe...

That's absolutely correct.

And maybe that's what we really are lacking because we don't do that well. I mean, there are examples in machine learning of doing it between language translation. I can translate between this and this language and therefore it makes it a little easier to do this dialect of that language. But we really don't know how to do that, and you wonder if it's something that can be done with digital computers as we kind of think of them now.

That's right. And I'll tell you an experiment we actually have done and we use this today... this is real, is that we have taken 30 sophisticated large systems that recognize images of objects. These systems actually exist, these models exist out in the community. They're based on 20 layers of perceptrons, neural net perceptrons, multi-layer perceptrons, very sophisticated, took people a long time to build those. And what we did, we did apply transfer learning, because what we want the system to learn is not the object or the images. We want them to learn...We're in the business of communication, [so] we want them to learn in a cell site, in a radio site when you go up a tower and look at the tower, what is each piece, what object is it looking at? This is basically a wire, and this may be a rod, and this may be the radio unit, and we want this system to learn not only what the object is, what's wrong with the object.

So what we did, we actually applied transfer learning using these systems that were built on cats and dogs and others and we changed one layer, only the last layer in the neural net, to learn something that's very specific to AT&T and we were amazingly successful in doing so. So yes, these systems don't know how to generalize to any form of intelligence, they don't. But they are able to keep some learning, some intelligence they've learned from a different task, and with some minimal effort and training, you can get it to do a new task.

So tell me about AT&T Labs. I can imagine your charter must be very big because AT&T touches so many aspects of people's lives. So tell me about AT&T Labs.

So AT&T Labs was created decades ago. Its root goes back to AT&T Bell Labs. This is where UNIX and C++ and the transistors were invented. In fact, the early work on artificial intelligence did come from AT&T Bell Labs by Claude Shannon in the ‘50s. And AI has been something very dear to us because it's all about: ‘How do you use data to do smart and intelligent automation at scale?’

We are a large business, we serve a 100 plus million wireless customers. We serve businesses, we serve our broadband customers with entertainment and advertising. And so we have a very broad business and as part of this business, for us to succeed, it's about scale. So we do a lot of AI work to build the network of AT&T to build... there's a lot in the news today about 5G.

We are using AI to build 5G. 5G is all about many, many small cells and a lot of computes sitting in your backyard following you. The design, the production, the automation of that is we're applying these AI capabilities to enable us to scale, because at the scale we're looking at to do in the United States, it's just prohibitive to have enough humans to be able to design the network to be able to optimize the traffic where it's required every second of the day, and to be able to place these cells wherever it's needed, and the compute wherever it's needed around the clock.

You mentioned Claude Shannon. He was a fascinating guy. He wrote a paper in '49 or '50 about how a computer could play chess, and it was kind of a really watershed event, because up until that moment, computers were really good calculators and they manipulated numbers. And then one day Claude Shannon says, here's how you could teach one how to play chess, and that kind of mental leap gave us the world we have today where it's like, oh all these other things can be reduced to math.

So you would like to say that Claude Shannon did something that wasn't necessarily – you would almost say that was pure science in a way. I mean AT&T Bell Labs wasn't working on a chess-playing computer or anything. So I'm curious, is everything you do applied science? Do you only take on projects that you can say, “ah, this is why this would be useful to us” or are you up doing stuff out there that you don't even know—if it will ever touch your business?

It's a really good question. I think that when Claude Shannon was working on the chess game, he was also working on information theory. This is what he's famous for, which by the way, gave birth to the digital age. We would not be in existence today with what digital communication, ones and zeros and bytes going around our network, without him. He invented that. He's the one who invented the electric mouse that's actually the core of how AI started of a mouse learning how to navigate through a maze.

What people don't understand is that Claude Shannon did those – we have actually a museum in New Jersey with a lot of his inventions—he did them for a purpose. It wasn't just he was interested; to him, it was a hobby; but he actually had a purpose for that. If he solved this, he had an interest of where this will go. Same thing with us at AT&T Labs is that we have very broad interests.

Our business is communication and entertainment and, frankly, there are an infinite number of problems we work on today that have implication on our business. Some of them may be a little far out. It may be three to five to seven years away; some of them are within the next 12 to 24 months. But we do have to do that, we have to look at things not only for today, we are actually inventing some things that may only be in use five, seven years from now. So that's the beauty of the AT&T Labs: it's very business focused on today, but it also has the capacity to [be] looking at three, five, seven years from now to ensure that this space of communication and entertainment continues to thrive in the world.

And I would be remiss if I didn't plug my favorite book about Claude Shannon. If anybody's interested in that is called A Mind at Play. A lot of people my age (I'm 50), a lot of people my age were really inspired by the You Will series of commercials you guys did back in the ‘90s. (And to those who don't remember them or weren't around then, AT&T ran all of these commercials, there would probably 10 or 12, but there's a Wikipedia entry on them and you can see them online. And they're like, “Have you ever sent a fax from the beach? You will. Have you ever bought tickets to a concert with your phone? You will.”) And it foresaw computer dating and being able to call in and check on your house through like remote cameras and all these things, none of which existed. Like you could see it would maybe be there someday.

But they were remarkably prescient. They got them all right, except maybe sending a fax from the beach. But all the rest of them were really on. Does AT&T Labs do similar – like, what would have been the reason to do that? Do you still do that kind of aspirational... because those were very inspirational and you still publish things with the sole aim of getting people excited about the future and AT&T's role in it presumably?

That is absolutely correct. So number one, I should mention, two-three months ago, it was the anniversary of the “You Will” campaign, 25 year anniversary of the “You Will” campaign. And actually we gathered the brightest minds at AT&T and a number of futurists and we brought them together that we work with very closely, and we created the next 25-year campaign. We actually – it's available if you go to YouTube or others, you'll be able to see what we think the next 25 years looks like or what we think would look like. It's very fascinating.

Clearly, what happened in the past 25 years, you're absolutely right, we were – it's great to see that a lot of these capabilities have come to light. I think the next 25 years are going to [be] even a lot more exciting because now we're living in the era of the digital age, and with 5G there's going to be an explosion of data not between humans and machines but between machines and machines and machines to everything.

We're moving from the Internet of Things to the Internet of Everything. So if you watch the You Will, the next You Will that we put out in the past two-three months ago, you'll see sort of the thinking of what we are trying to do in the next 25 years. Some of those we're not – well, we never built the campaign to say that AT&T is building all of this. What we said is that AT&T is enabling those through this high connected mobile network and wireless network, and that is what we are providing.

For some of these applications, AT&T is actually leading and we are building those; but some of those, we will not be the one doing it, we will do it through partners. In fact, we just opened up Foundries at AT&T with our 5G and edge capability to allow small customers, medium businesses, any type of business to come and innovate together with us. So towards the 25-year vision campaign, it is not our intention to invent it all, it's our intention to work with a larger community, open the network, open the access and really have others come and invent and innovate with us.

It's funny because when you think about infrastructure projects in the United States over the last 200 years, you had the interstate highways which were government initiative; and the internet, for that matter, which was largely governmental; you had the transcontinental railroad which was not governmental, I mean, it was subsidized.

Anyway, I'm curious to what extent you think that forces of free markets are able to build out the next generation of infrastructure, or is there a role for government to play in that and to potentially take a leadership role in it? Or is that one of those things that competition is going to build us something better than we could even conceive of right now?

I cannot comment on the role of governments here, but let me tell you what we're doing today in building the network of the future in 5G because that's real, that's capital intensive. We are changing the ballgame here than how infrastructure has traditionally being deployed, certainly for a communication company like us. First, we are going from a lot of infrastructure, a lot of capital, a lot of hardware to a world where software is eating everything. That's number one. We are very heavy software now focused than ever before in our lifetime even as a company. That's number one. It's all about software.

Number two is that the hardware we use to support the software, a lot of that is virtualized and cloud-based. A lot of that is built as commodity hardware and it's built in as minimal location as possible, because it's not just about the infrastructure, it's about maintaining the infrastructure, doing lifecycle management, upgrading the infrastructure, etc.

So we're moving to a world where maybe 10 years ago, 20 years ago, we used to be 90% hardware, 10% software. Now we are in a world where 90% is software, 10% is hardware. And even the hardware—we talk about white boxes and off-the-shelf hardware, a very different world than has ever been before. So that entire view of capital investments and infrastructure investment or even 5G, is completely different in 4G, and completely different in any other Gs before that.

All right. Well, it looks like we're running out of time here. If people want to keep up with AT&T Labs and you and all the things you all are doing, can you throw out a few things to follow or find or search for?

Yeah, absolutely. So number one is that we have a website www.research.att.com. You can search for AT&T Labs and AT&T Foundry, [and] you'll find a lot of information about us. You could look at a lot of the open-source. One of the things AT&T Labs feels very strongly about [is] we've started a movement a few years ago about openness and open innovation, and a lot of the software we build, we actually put it in the open source. We've put out over 10 million lines of code just in two years, so go to our innovation channel, also on LinkedIn, and you'll see a lot of information about AT&T Labs.

All right, I want to thank you so much for the time. It was a fascinating half hour.

Thank you so much, I appreciate the invitation, have a wonderful day.