Being Human with Algorithms: Marc-Oliver Pahl interviews Vinton G. Cerf

 

Vinton G. Cerf co-designed the TCP/IP protocols and the architecture of the Internet and is Chief Internet Evangelist for Google.   He is a member of the National Science Board and National Academy of Engineering and Foreign Member of the British Royal Society and Swedish Royal Academy of Engineering, and Fellow of ACM, IEEE, AAAS, and BCS. Cerf received the US Presidential Medal of Freedom, US National Medal of Technology, Queen Elizabeth Prize for Engineering, Prince of Asturias Award, Japan Prize, ACM Turing Award, Legion d’Honneur, the Franklin Medal and 29 honorary degrees.

I had the great honor and pleasure to interview him on the digital transformation. Below you can find the transcript and the video of the interview.

Marc-Oliver Pahl: Vint, it is my big pleasure talking to you two years after I interviewed you last time about how the Internet works. By the way those people who have not seen the interview have a look it was a very very interesting discussion. You can find the video of that interview here.

I heard there are some people out there not knowing you, so could you briefly describe yourself?

Vint Cerf: Sure, my name is Vint Cerf I’ve been at Google for 13 years now, almost since 2005, but many years ago my partner Bob Kahn and I designed the Internet and and led its development. So we’re two fathers of the Internet, but anyone who’s been tracking any of this knows that there are literally hundreds of thousands maybe even millions of people who have contributed to the Internet’s evolution and they continue to do so today.

Marc: Today we have this thing that is called digital transformation where your technology plays a central role because without communication the digital transformation would definitely not have come at this pace, and so:

When you look at the digital transformation today, which are the most prominent effects for you that you encounter there?

Vint: The first obvious one, of course, is that we have computer mediated communication something that used to mean email way way back in the 1970s. Today of course it means all kinds of modalities: what we’re doing right now, using Skype, we could be using Google video conferencing or we could use FaceTime or many other applications like that. Voice over IP came along in the mid-1990s to transform the telephone system which today is heavily based on the Internet and Internet protocols perhaps the most important thing though is that we’re seeing computers intervene in our communication with each other in fairly significant ways. Not only are they the medium through which we communicate with each other but we’re also beginning to communicate directly with computers using more typical human means. Natural language processing natural, language understanding or at least recognition; has become increasingly accurate and so it’s quite common now for people to speak to their computers in order to get them to do things. We see this with Alexa, with Google home and some of the other applications of that type. Along with that is the use of machine learning in order to improve the ability of a computer to respond to inputs that it gets; whether they are oral inputs, image inputs or textual inputs. So these computers, as you say, are playing a very key role in intermediating human interaction.

Marc: What I find especially interesting with your perspective is that the first thing you mentioned is, you have humans that want to communicate with humans and in the middle you have the computer. It is a very positive and nice view. When I think about digital transformation, also the transformation of the work environment comes to my attention, and there, depending on the job, it is not so much about communication between two humans but also about automating processes. Can you say something into that direction?

What does it mean to you?

Vint: Yes, certainly. I mean, we were thinking about e-commerce, for example, as far back as the early 1990s. I built a shopping mall for the MCI company way back in 1994, but it was a little ahead of its time because not everyone had a
smartphone, which didn’t come in till 2007. So, we’re seeing several things
happening, one of them is the use of computers to organize business, to facilitate electronic commerce, to make products and services more visible to people. As you can recognize companies like Google and Facebook and others are using computers and algorithms in order to figure out which products and which
services should be brought to the attention of the public. So we find the computer algorithms intervening in our consumption of products and services, either facilitating their acquisition and use, or even being part of the product or the service where the computer software is an integral part of the application that you’re using. So it’s an inescapable fact that software is going to surround us in every possible dimension, as the rest of the decade unrolls.

Marc: In that direction I want to lay the focus on the aspect that the computers are also very good at storing and evaluating data and so the cue words would be Big Data and artificial intelligence for data analytics.

Which risks and which chances would you see in these directions?

Vint: Big Data, of course, is a popular term these days and so is artificial intelligence, although I think I would prefer the term machine learning to narrow the scope the functionality. The two things that strike me are that the machine learning tools are extraordinarily capable of finding signal in a great massive pile of data, looking for correlations, for example, we have to be careful not to mistake a correlation for a cause, that’s one very important scientific observation. But there’s another thing and that is that these algorithms are in some cases brittle, they may analyze the data and come to the wrong conclusions. We see this with image processing, for example, where classification fails, because a small number of pixels different from the training images cause the system to misrecognize something. It can be fairly dramatic, I mean I have pictures of dogs and it recognizes dogs and then you change a few pixels and it thinks it’s a crocodile. This is a consequence of the way in which the machine learning actually works. So one of our big tasks over the course of the next ten years is to understand much more deeply how these algorithms are brittle, or why they are brittle and how we can make them more resilient and resistant to significant failure. It’s a little bit like chaos theory, where you have small changes in input that produce unpredictably large changes in output. One more, it’s a better theory for machine learning tools anyway, so as to avoid some of those consequences.

Marc: Going into this machine learning direction, and also with other algorithms, do you think that humans alienate? Do they not understand anymore what happens in the algorithms? And do you see this as a problem?
Maybe, first for computer scientists like us in general; or even closer experts who both still don’t understand what exactly comes out of machine learning algorithms; and then, for society at large, so people that do not do anything with computers at all.

Vint: The interesting fact, I think, is that with computers were able to create systems that we don’t understand, whose performance is so complex that we cannot predict it very well. When you think of the billions of devices that are on the Internet, all interacting with each other, every time one computer sends a message to another computer it’s an experiment because those two computers might never have interacted before. So, we really have quite a bit of work to do, in order to be able to understand the way in which these algorithms function, how they might fail. In particular, we might end up training them in such a way as to introduce bias that we don’t even recognize. Many people are very worried about software to which we grant autonomy, in the sense of making decisions, taking actions. A dramatic case of that is self-driving cars, for example, but imagine algorithms that decide whether or not you can get a loan or what the interest rate is or whether or not you are or you aren’t insured. These algorithms need to be carefully analyzed and we need to be on guard that they have not accidentally introduced a bias based on the data that has been used to train them.

Marc: Thinking about this Uber accident, where this woman was tragically overrun. What I read in the German press was that the sensors were detecting everything properly, but the problem was that the the software was not configured correctly, so it was not taking into account all the things that it had. So, to me, this is tragic because there’s a great technology and then icon cases like that one, as tragic as it is, where something really bad happens, push peoples trust away.

Do you have ideas about what we should do so that people that are not into machine learning can gain trust in this technology and explore it and give it the chance that it should have, because it could bring great advance.

Vint: This is actually a very very important point, this whole loss of trust in general. We are losing trust in not just software but in our institutions as well. This has been eroding over time, the press has become less and less trusted. In the U.S. the Congress is very badly trusted or not trusted at all. But let me come back to your primary point, which is when these incidents happen, or they’re very visible. My understanding of the Uber incident is that the automatic braking system, which is in some sense independent of the artificial intelligence, was disabled and so the system would have stopped, we think, not because of the AI algorithm but because the system had hardware in it that was designed to not run into things. And for reasons I don’t understand, that was disabled. So, as you say, every time that happens people associate bad things with these kinds of systems and then assume that they are all not to be trusted.
It is going to be hard to build up trust, in the case of the Waymo system, that is part of the alphabet company, I would urge you and your listeners to look at the safety and security protocols that Waymo is using. They’ve documented these extensively, and you’ll find that very stringent testing is done both in the physical world and in the virtual world. The aspect of the virtual world which I find the most interesting is that it’s possible for Waymo to present to the software, emulated inputs that would normally have come from a sensor. They can create situations in the virtual world to emulate extreme cases, for example, a ball rolling out in the middle of the street followed by a young child is something that they can test without actually putting a child on the street with the ball and running into it. So, I’m very impressed to be quite frank, despite the fact that I work for a sister company at Google, with the degree of care that Waymo has introduced into its testing regime. That’s the sort of thing we need to document and share with the general public, as you say, so that they can build up confidence in these systems.

Marc: Indeed Google is probably the company that is doing the most extensive tests. Therefore they are also for the public audience not doing such visibly fast progress over the last nine years, because they do this testing. Then Uber comes and Uber says, I mean I just hypothesize, “we don’t do much testing, we just roll it out and then let’s see what happens”. This is the typical. problem if you are the first in the market, you can get the market share, but you can also lose the trust for the entire field. Google is doing the much more responsible thing and to me also the only way to go. Because it’s actually humans and machines and it’s large and heavy machines.

Would there be a code of conduct needed, saying company X: “it’s not allowed to do this unless we have certain regulatory levels met.”?

Vint: Yes, as you know, there are already definitions for certain levels of automobile operation, level 5 being full autonomy. I believe that it makes great sense to have a regulatory regime, which requires certain levels of testing and demonstration of capability before you put the cars on the road. We do the same with human beings we don’t let them drive cars until they’ve demonstrated that they know what the rules of the road are, that they demonstrated they can control the car, they can park it, they can drive in live traffic. The amount of testing that we do to grant a human the license is de minimis compared to the amount of testing that we do for self-driving cars. And so, one hopes, of course, that the self-driving cars will prove to be safer in their performance than human beings are, but we do need to show that and statistically it’s going to be important to show that the cars are at least as safe as human drivers and, hopefully, much more safe.

Marc: What would you think, in 10 years will we still have manual driving?

Vint: Yes, I think they will be very hard to eliminate. However, I can imagine that, one might insist that there will be no human drivers on certain roads which are exclusively for use of robot drivers. The reason for that is that we might have communication between the cars, if we can standardize that. So the cars can coordinate with each other about change of lanes and making turns and so on or even rapid braking. So that it’s more than just the sensors that are saying “oh my god I’m approaching the car” doesn’t mean a higher rated speed. You can be signaling at the speed of light the fact that you’re doing something and it might affect others, the car it’s behind you or in front of you. So that information exchange might in fact increase the safety of the entire system.

Marc: Yes, I think so too and, from what I read, it’s the biggest threat for the autonomous driving, or the biggest problem, that you have human drivers. Because if you wouldn’t have them, the automatic cars could just collaborate. Even if they’re not perfect, by just exchanging information and being neutral, this will be a much easier game on the road. Having these specific lanes or even roads for the autonomous cars to coexist but not interfer with human drivers, or even just for lorries at the beginning or something like that, is probably the best way.

Vint: Yes, interestingly you might not even need lanes, in some sense. These cars are coordinating properly, although I think as long as we have human drivers we need to have those visual clues about where you’re supposed to be.

Marc: Yes, coming back to what you’re currently doing: what would you say is currently the most important thing you’re doing to shape this digital transformation?

Vint: First of all I’m still spending a great deal of my time trying to get more Internet infrastructure in place. So, I’ve been encouraging governments and the private sector to work together to make that happen.
I’ve also been very concerned that just getting Internet infrastructure in place, like Wi-Fi for example or LTE, is is not sufficient, that’s simply the beginning. We’d like the system to bring useful services to the users, which means local information in the local languages, it means things that make people’s lives measurably better, whether it’s improved healthcare or the ability to drive when otherwise you couldn’t, variety of things that are demonstratably better than they would be otherwise. This positive feedback is important, so I spend a lot of time on that.
I’m spending also a great deal of time right now on the Internet of Things and the reason for this, of course, is that these things are full of software and the software has bugs, so we have to make sure we can fix the bugs. We have to be sure that we can identify them. We need methods for fixing them, to avoid for example the accidental downloading of malware, which is not the best way to fix a bug. The IOT avalanche, I think, is very important. There are privacy, safety and security questions. Or questions of autonomy, in the sense that you don’t want a house full of these devices that doesn’t work when the Internet connection goes away. I’m very much focused on that.
The third thing that has my attention is digital preservation, a big concern that digital content may actually evaporate over time, either because you can’t find a reader to read the digital medium or you don’t know anymore what the bits mean because the software that knew what the bits man doesn’t run anymore on operating systems of the day. That occupies a lot of my time.
Then just to add one more factoid I’m also working, as I’m sure you know, on the interplanetary extension of the Internet. This work is proceeding taking a long time to get to where we are; about 20 years. But we now have an operational two planet system plus the international space station. And as new missions to Mars are launched during the 2020s and to the other outer planet, I’m persuaded that we will grow this backbone mission by mission so that Mandarin robotic space exploration will get good quality Network support.

Marc: Wow, so it’s not only digital transformation on Earth but in the entire solar system or entire space.

Vint: Today the solar system tomorrow the galaxy.

Marc: Nice. Then, we already come to the end of the interview, maybe the last question is: for you personally what does it mean being human in such a world with lots of algorithms that you daily and even hourly interplay with automatically, voluntarily or involuntarily?

Vint: Two things: first, many of them are quite helpful. Translation, for example, of multiple languages helps me a lot because my language skills are sort of limited. So, Google translations for webpages, for example, are extremely helpful.
They are occasionally amusing, you probably, know the story about my looking for a weather report in Heidelberg getting a German weather page having it translated into English, so quickly I didn’t even know that it was a translation, and then discovering that the weather report was: zero probability of rain, zero probability of fog and zero probability of ice-cream. Because ice was mistranslated as ice cream instead of hail. That was a very funny thing.
So, I see a lot of the software tools and artificial intelligence tools as tools literally, to help augment my own abilities and I hope that that’s how most of them get used. That make us, in some sense smarter, more capable of doing things that we could not do at scale ourselves. For example, indexing the world wide web, is not something any human being could do and yet computers do it very well. So I’m frankly looking forward to the convenience and the augmentation of my own abilities.
At the same time I am very very conscious of the fact that these tools are not perfect, they can make mistakes, they can be misleading, they can take you to conclusions, for example, that might turn out to be wrong. So we always have to ask ourselves: are we sure that this algorithm is producing
good quality results? How can I validate the results using other information? Just like you would do in a math problem, where you make an estimate of the answer to see whether or not the order of magnitude of the answer matches your
expectations. We need to do something like that in the case of these software based tools.

Marc: Very good, so it’s a very optimistic point of view – as long as the people implemented right.

Vint: That’s right, there is a downside to all this: even setting aside the possibility that the software based tools don’t function perfectly, there are people who deliberately abuse the infrastructure and produce misinformation or, you know, generate fake news, for example, or, you know, inject deliberately harmful information, whether it’s malware or simply socially harmful content like bullying for example or stalking. This, of course, is a problem with human beings you know they are sometimes badly motivated and technology can’t necessarily stop that. We might try to detect some of it, but we also have to be on guard when we do see these kinds of things happening, that we have some mechanism for responding. At a recent conference at Ditchley house in the UK, one of the conclusions that was reached is that these online systems need to be designed with what we call traceability and this doesn’t necessarily mean that everyone has to identify themselves before they’re allowed to use the network, but what you want are, under certain circumstances and with the appropriate authority, that you can pierce the veil of anonymity in order to track down a party who is harming someone else through this online environment.

Marc: And then again, if it’s implemented properly then this non-repudiation is good because you can track the people down, if it’s in the wrong hands then it might again be a risk like always. What is always important to me is that the digital technologies are also an amplifier. They can amplify the good things and the bad things.

Vint: It’s all true, yes. We have to deal with scaling in both dimensions. Human beings are unpredictable and in some cases not trustworthy. We have to build, into these systems, mechanisms to cope with that.

Marc: Very nice. Vint, thank you very much for the great interview! We are looking forward to your talk in September and hope that many of the viewers of the interview will come to our event on the 20th and 21st of September 2018 in Heidelberg. I’m sure you will have many more interesting points there! I always enjoy talking to you! It iss always a source of big inspiration. It was very nice talking to you. Thank you very much.

Vint: Thank you so much, Marc-Oliver your questions are always stimulating.

---