You to here . Guest so im actually a researcher. I work on integrated optical sensing methods and technology. Host and before we get into what those are, whats your background . Guest so i have a background in solid state physics and semiconductor devices. Wheres your degree from . Guest berkeley. Ucberkeley. Host and ph. D. . Guest my undergrad was at iowa state university. Guest are you from iowa . Guest im from minnesota originally. [laughter] host well, youre here at bell labs now, and youre working again on guest im working on optical sensors. Host which is what . Guest really anything that detect light. The most familiar youd be familiar with is a camera which we all basically carry on us every day. Typical cameras only see twodimensional images. So what im working on are 3d Optical Imaging techniques very similar to radar but using light. Host arent those in practical use today already . Guest yes. There are some uses of these. One of them is in the medical field, its a technique called optical coherent [inaudible] and you can use it to do 3d imaging in the eye, for example. You can actually get a micronscale resolution of your entire eyeball and the retina just with light. Host and thats pretty common, isnt it, today . Guest yeah. These systems are becoming more and more common, especially in optometry. But the systems tend to be rather large. They take up a large table and cost tens, even hundreds of thousands of dollars. Host so whats your research doing . Guest so im working on taking that entire system and shrinking it down to the size of a chip, something that we could actually integrate maybe in all of our phones, possibly even on the clothes we wear. And allow us to not just monitoring ourselves, but also the environment around us in full three dimension. Host now, how far are we away from that technology . Something theres a lot of challenges we still have to overcome, but we right now have all of the Building Blocks that weve kind of developed with the Optical Communication Technology that bell labs has pioneered in the past. Were now repurposing this technology for this new tech anesthesiology. Host so in your work, is it completely healthrelated . Is that where youre headed . Guest thats definitely where im headed, is healthrelated. Really sensing the biochemistry of the human being and basically monitoring our health on a day to day basis. To me, its really one of Biggest Challenges we face as humanity, is how we feel, and its really something that we dont monitor. And so really theres a lot of opportunities and challenges to overcome to really make dontous Health Monitoring a reality. Host so in a sense, is it sensors and then the math that goes into the optics . Guest yeah, yeah. So theres really a whole spectrum of work that needs to be done. Both making the components that can be small and integrated, putting it in a system. Theyre actually copping the developing the algorithm to actually diagnose things with these tools. Im not a doctor. I dont have a medical background. So what i do is i really make tools, new tools that can help those professions actually do their job better. Host so, mr. Eggleston, you talked about the 3d eye image. Where else can or is this being used today . Guest yeah. So its really seeing a lot of uses. So the second most common is in skin care. So, for example, if i have an abnormal growth or something on my hand, i can actually do a full 3d scan of it and tell is it cancerous, is it benign, is it something i need to look at. And its completely noninvasive. You dont have to cut anything off, you dont have to take a blood test. Theres many other fields people are using it. Dentistry, you can actually look at the enamel of the teeth and see microfractures before they develop into cavities. Even inside the body theyre using this to do internal surgeries where theyre needing to see what theyre doing on the inside of of the body. So really the range of possibilities is kind of endless. Anything youd want to see or sense you can kind of use this technique. Host now, a lot of us are familiar with mris. How old is that technology, and is it related to what youre doing . Guest yeah. Mri is a pretty old technique. Functional mri was actually pioneered here at bell labs, and its a great technique. It lets you scan the entire body. But the resolution is on the order of millimeters to centimeters. Its very, very challenging to scale down, and thats why theyre very expensive and people dont have access to them. Now, with these optical techniques, they can be very small, very cheap, and they can allow us to do a lot of the things that were doing with mri but at much lower cost. Host so whats more important in your work, the scale or the technology . Guest so to me, i think they kind of go hand in hand. But at the end of the day, when you can scale something down to a small size, you can make it cheap, and really you enable it to use for a mass audience. So an mri is great, if we all had access to mri in our daily lives, it would significantly enhance our health. So, to me, really helping everyone is one of my main goals, so thats ooh why i think driving the size down with the technologies we have is so important. Host now, on your door here is a sign that i want to ask you about. Danger, invisible laser radiation. [laughter] guest yes. Host happening in here. What guest so a lot of the sensing that we do is in the infrared. So its longer wavelengths of light than what you can see. And we tend to use rather highpowered systems when were demoing them in the lab. So this is just a warning that depending on what experiments were running, you might have to wear safety goggles. Now, as you can see, my door is open. Thats how i usually work. Usually we dont use these lasers, because were really looking at something thats used to scan the eye, so its actually completely safe. Host where is a laser . [laughter] do you have one in here . Guest were surrounded by lasers. I can bring you over here and show you some examples. Host and these date back, in a sense, to einstein, dont they . Guest yeah. So, lasers date back to the late 50s, early 60s. Host based on his work . Guest yes. His ideas. They usually come in a form like this. Host so thats a laser in there . Guest this is a laser. This is a pretty nice laser thats used for medical imaging. And the actual laser is much, much, much smaller. The lasers can actually be, actually even, for example, if you have the air pods with your iphone, those have three lasers in them. Youactually make them very, very small which is part of the reason theyre so attractive. Now we can embed them in everyday devices, and they have an incredible amount of added functionality. So, yeah, this is a laser. We actually have several lasers on this table here. And theyre all different form factors, shapes, sizes and really its all about packaging. Lasers are tiny, but we put them in large boxes because then theyre easier to handle. Host so before we interrupted your day today, what were you working on . Guest so im working on a new method to actually implement these systems on chips. So a lot of it actually is about when you come to system integration, its writing codes and algorithms to actually to do what you want to do. When everythings tiny, you cant physically change it, so everything has to be automated. So i was actually writing code to do this. Host where did your interest in this topic, in these topics come from in. Guest so ive always been fascinated with light. Host with light . Guest with light. To me, like, my vision is my number one thing. And so when i went to school, i learned more about light. And light is amazing because everything it touches, it actually picks up the signature of that material. So every beam of light that bounces off you and hits me actually carries information about you. If i look at it in the right way, i can discern an incredible amount of information from that. So that kind of curiosity has led me down this path. Host Michael Eggleston with bell labs, thank you. Guest thanks. Host and the communicators continues its visit to Nokia Bell Labs in new jersey, and joining us now is a gentleman named sean kennedy. Mr. Kennedy, what is your position here at bell labs . Guest so im a department head, and i lead the math and networks team. Were interested in a lot of things, but Artificial Intelligence, augmented intelligence and, you know, a whole host of other Network Algorithms we try and develop here. Host before we get into those guest sure. Host what do you mean by map of networks . Guest well, networks are, from one point of view, just a structure you can look at. We study the properties of these structures, and then we, generally we build algorithms to solve problems. So whether its questions about how you get information from one part of the network to another part of the network, or, i mean, you have facebook which is just a large network, so lots of networktype problems come from studying these, the structures of these networks, and you can decide things like who you should be friends with next and these other questions that, you know, we experience on a daytoday basis. Host you mentioned Artificial Intelligence. How do you define that . Guest thats a good one, actually. Im not exactly sure how i would define it in general. I mean, i think people tend to think about replacing human intelligence and thinking machines. We have, like, selfdriving cars and things like that. I tend to think about it a little bit more like a mathematician, so right now Artificial Intelligence is just a giant optimization problem where we build ma cheaps or machines or mathematical models that do a good job of answering the questions. For example, we may have a lot of data that is cat images and a whole bunch of data thats not, and we build a model that is able to decide which images are cat images and which ones are not. Its a large optimization problem. Host is siri Artificial Intelligence . Guest yeah, i think so. Whether or not theyre thinking machines, i would say, no, but in the space of an optimization problem, they definitely fall into that realm. Host whats the difference between a thinking machine and a learning machine . Guest uh host are we there yet . Guest no, no, definitely not. We dont have intuition built into these machines. If i show you a cat picture, youll probably do 95, 96 getting it right. But the machines do ap excellent job as well of identifying cat pictures, right, but theyre not really sometimes theyre surprising us. You also probably drive a car really well, and machines are able to drive cars now. I think theyre not yet thinking machines. Theres no they dont really show a lot of intuition behind what theyre doing. Host dr. Kennedy, what kind of research or whats the, whats the back end of that machine recognizing the face of a cat . I mean, what kind of technology is that . Guest so its generally a big network. Its a convolutional Neural Network. Dont have a picture here, but they are Large Networks where you input the image just by the pixel values, you know . An image is nothing. You have these dots on the screen, and they show up with different intensities so you can turn this into some sort of digital signal. So you feed it into this network, and theres a whole bunch of nodes in the middle of this with edges that use this to compute different values as it pushes itself through the network. Im not sure thats the clearest answer here, but host well, lets go to the 1s and the 0s before we look at whats on the board. Why is everything in iss and 0s in 1s and 0s in your world . Guest well, because we digitize everything host right, but what do they mean . Guest they can mean anything. They can represent your heart rate in your watch thats reading your heart rate, or they can be the pixel images that we send to the screen or an image that we can produce on the screen. Its just the way that a Computer Stores information, is in 1s and 0s. So thats normally the way we deal with information. Host all right. We are in whats called the anomaly room here at bell labs. What are you going to show us . Guest so what were talking or what we have here is called augmented intelligence. So rather than trying to replace the human thinker, what were trying to do is were trying to build tools that help humans understand information. And so were not doing this in a way that is just for data scientists, we want to build a tool that is simple enough for children to look at. Indeed, i have examples of ingesting data from pokemon that kids love to play with. We took this to a robotics conference, and a whole bunch of kids were playing this and explaining the me how pokemon worked. But what we want to do is ingest data of all types across all spaces whether its iot sensors or image sensors or tax data, etc. , scraping the web for documents, etc. , or even, you know, your web site that has videos, etc. , we could scrape this information. What we want to do is build an interface that allows people to explore this information so they can see this information in a way that theyve never seen it before, allow them to compare things inside this data set so they start to understand why things are showing up on the screen, and then ultimately, we with want to use this information to create knowledge. What we mean by creating knowledge is we want to be able to make decisions about this information. You could imagine youre given some huge spread sheet or some web site, some document you want to ingest. Ultimately, whats useful for you is to take that data, so the 1s and 0s of the document, turn it into information, but ultimately knowledge that you can use to answer questions. Maybe your boss is asking you for some information from this document. So what we have here is what we call augmented intelligence and what we think of as assisted thinking tools. So theyre tools that allow us to think better, to answer questions better. Host well, lets hit be pokemon and walk us through what augmented intelligence can do. Guest gotta have a special touch. Host ah, i guess you got to be a mathematician. Okay, im going to stand back. You go ahead and walk us through guest this is just massive information from pokemon. Like, we ingested you have the different creatures, whatever you call them, the pokemon, i guess. And you have a whole bunch of extra features about these pokemon. The screen shows a global ranking of what the smart machines that underpin this machine thought were the most important things. So if theres something on here that you find interesting, another feature we feel if youre going to play with data and use data, you want to interact with data. So perhaps you find this guy interesting. So you simply touch him. Now, by touching him, what youve done is youve indicated to the machine that this is the piece of information that youre interested in. Its not showing here, but we also have the ability to tell the machine youre not interested in certain piece obviously information. So once with i did that, what you see is immediately the screen was reorganized around the pieces of information that were interesting. So you can see these are different, you know, properties of these different pokemon whether its their way that they breed, i guess, or their name, general category overall. Continue . Host yeah. Go right ahead. Guest this is a bit of a toy example and, you know, we did this mostly for kids to think about information, but, you know, or, you know, just to play with information or just to get contact. But, you know, when youre thinking about a problem, what you want to do is you want to take that information, and you want to quickly kind of streamline it to the pieces of information that youre interested in. I mean, i guess you could think about doing an interview. Youd quickly like to be able to scrape these documents and go through and extract the pieces of information. Something would catch your eye in a document, and youd collect that, and you would hope youre able to reorganize the screen to help you filter on these pieces of information. So, of course, when youre thinking about things that are, when youre thinking about a problem, lots of times we dont want to think about the things that are just immediately related to the subject, we want to look further out. And, you know, thats a what im able to do with this machine. Im able to zoom in and out on the data to look for things that are related. So maybe well go, ill see if i can come up with a slightly different example here. You have questions . Keep going . Host keep going. Guest lets see if this one works. I never get to play with this screen. Obviously, i dont have this in my office. I dont think this is the one that i want. Or maybe yeah, lets go with this one. So one of the curious things about Artificial Intelligence which i get maybe comes to crux of why these machines arent really, you know, theyre not intelligent in a sense is that if you were to train up a Neural Network or to recognize cats, to go back to our simple example from before. You did a good job of sticking images in. If you put in a cat image, 100 of the time it says cat, if you put in an image thats not a cat, it says not cat. So the question is whats it going to do. Did you learn anything about that cat. Or what about similar things like, whats going to happen, for example, if i stick in an image of a fire truck. Is it going to solve my problem . Have i learned something about the structure of cats thats going to help me recognize fire trucks . You know, its curious that it could kind of return either thing, right . It doesnt really contain any intuition that is helping you get closer to recognizing different types of images. Youve done a really good job of training on cats, for example. So is we get here in this space, we said, well, what we should be able to do is train a big Neural Network and then use this Neural Network to help us understand what things are similar and what is happening in the inside of a Neural Network. So, for example, is there something here that catches your attention . Ill give you a second chance, no third chance, to touch the screen. So tennis serve. Host will it teach me a better tennis serve . Guest well, i would hope so. A. [laughter] these are all images related to tennis serves. This is not based on keywords, more important this is on deep learning features that underpin each of these images. So you can see why these images are similar is not because of the tennis serve, theyre similar i mean, you can just tell as a human when you look at all these images in the middle, they seem very similar, right . Pe