Being Human with Algorithms: Robert Tarjan
The Digital Transformation is currently changing all aspects of our lives fundamentally. In this series I discuss with people about their personal experiences regarding this Digital Transformation.
This episode: Robert Tarjan, Turing Award Winner (“Nobel Price” for Informatics): 1986 for fundamental achievements in the design and analysis of algorithms and data structures.
More info about him: https://en.wikipedia.org/wiki/Robert_Tarjan
The Interview took place at the occasion of the Heidelberg-Laureate Forum on Sep 28, 2018.
Marc-Oliver Pahl: “Being human with algorithms” together with Bob Tarjan.
Marc: Bob, could you please introduce yourself briefly.
Robert Tarjan: I grew up in Southern California. I got interested in mathematics from a very early age, went to Caltech, took math courses, but I also took all the computing courses I could find. I got a chance to do programming very early. Went to Caltech, got a PhD, became interested in algorithms and algorithm design. Ran into Donald Knuth and also John Hopcroft who was on sabbatical from Cornell. Most of my career I’ve worked on design and analysis of algorithms. It’s kind of the plumbing of computing, shall we say, which is, if you want to do something complicated you need the basics. You need data structures to be able to access information, you need data structures and algorithms to process graphs. This is the sort of thing that I do.
Marc: With that said, you’re directly in the core of our motto, which is “Being Human with Algorithms”. When we thought about this motto, as we also discussed during the symposium some days ago, we thought about what does that, “Being Human with Algorithms”? The idea behind the motto was that we have this digital transformation that is ongoing, people are more and more interacting either directly or indirectly with algorithms and in the more recent past with computer algorithms. This digital transformation, how does it get most apparent to you? That this society changes and the way you interact with how things change.
Bob: I have a couple of different answers to that. I mean in my personal life I love to read the newspaper in the paper version, New York Times, but I can’t always get it, so I get it online. But if I’m reading it on the phone, it’s a much more constrained and different experience. It changes my interaction with the news. I’m sure that I’m like many other people, I’m too addicted to this device reading email, reading news and so on. I try to avoid social networks. I’m only on Facebook, so I can see pictures of my children, since they still use it. But it’s affecting all our lives in ways we don’t understand. And I would also say, you know, I’m proud of the algorithms I’ve designed. But algorithm now has become a dirty word because we have neural nets, we have “intelligent systems” that are getting trained in making decisions for us in ways that we do not understand and cannot explain. It’s having consequences in the real world and people start to question this for a very good reason. Indeed, I went to a conference at NYU some number of years ago, where this issue came up kind of in the early days when people were understanding the difficulties, the unintended consequences of all this technology that we have. It’s very important for us to think about the consequences of these systems.
Marc: Your answer is very interesting in many ways. So, you would say one of the positive effects is definitely being able to have more connection to the younger generation, because they are using these tools like Facebook more. When you say you have the pictures of your children there. The other aspect is these algorithms. I was also asking Tony Hoare and Steve Cook for their definition of an algorithm. The answer was it is basically a processing rule. I will ask you something a little different, namely, why do you think is there this recent shift in the meaning of “algorithm”? Or maybe it’s not even a shift in the meaning of algorithm but that this term became more used by the general public in connection with artificial intelligence. And then it started to have some of these negative connotations. Any thoughts to what happened to the term?
Bob: Yes, let me back up and say in my professional life, back to your first question, in my professional life it’s incredible to have essentially all the world’s scientific papers at my disposal. If i want to read something, I can go find it on the web. I’m fortunate to be at a university, so I can get through various paywalls, which are a big problem. This ability to get access to all the world’s information, this is very significant and very positive. Now getting back to your question, the classical definition would be a step by step process, like a recipe. I want to bake a cake, here the ingredients, the inputs, here are the steps to produce a cake. A neural network is really just an algorithm. We don’t have to go into the details, but it’s a big network. You feed in the inputs, it does some simple computations, produces some outputs. It’s not explicitly designed by any individual. How does it get constructed? You start out with essentially a blank slate. The structure is there, and you feed in lots of examples. The internals get adjusted until such time as an output is produced. Is this an algorithm? It’s not in the sense that I think of it, but it is an algorithm. The word has become to mean any computation or any computer process however it’s conceived. It’s interesting how this has evolved.
Marc: We were just talking about artificial networks and that they are getting trained and then they do something. When you translate them back into a processing rule, because this is what people kind of tend to do when they want to understand what the what the neural network is doing, would you say that is an algorithm again?
Bob: If it were possible to translate it into something that a human could make sense of, yes. Now the big difficulty is humans cannot make sense out of these things. These things cannot explain themselves. This is the challenge. If you get an output of one of these systems and that has an impact on humans, like I’m getting a loan or I’m not getting a loan, you want to know why. Especially if the decision is negative. Now, these systems cannot produce an explanation. This is one of them the big issues. The other issue is that the underlying technology is very simple, in fact. They are huge networks, but what they’re doing is very simple. The reason they’re having this kind of impact is because of the data that’s coming in. So, people question algorithms, they ask whether they have some inherent bias or not. Algorithms start out with a blank slate, whatever bias there is, is coming in from the data. Maybe we’re just getting revealed the kinds of biases that exist in society. Or maybe the data is somehow selected in such a way that their selection process introduces bias. In either case it’s the whole system, the data in combination with the computing, that is producing the result. I think it is inherent in the use of these systems that they have some human oversight. Especially if they’re being used in places where the decisions matter greatly to people’s lives.
Marc: This a very interesting and important point: the algorithms being an amplifier of what the humans put in. I’m thinking about this Microsoft chat bot that started being racist because the training data that it got was racist, it’s just been amplified.
Bob: This example also raises the issue that if you’re in an adversarial setting where somebody with malicious intent is controlling the system or using the system, may be able to feed inputs into it in such a way that you can produce any behaviour you want, even a systems that’s already been trained. This is no good.
Marc: And like you said before, it will be difficult to identify, because you essentially don’t know what these weights on the different functions actually mean in the end. If you continue training the neural network and someone puts adversarial data in it shifts. As you cannot understand what it does, you have the problem coming.
Bob: Yes exactly.
Marc: Would you say the key challenge for these artificial neural networks is to make it understandable for humans? Or what would you say is the biggest algorithmic challenge there?
Bob: I think making the outputs understandable for humans, making the behaviour understandable for humans and having some human filter in the process where these systems are used. If they’re producing some questionable behaviour, at least you have some kind of recourse. I mean autonomous cars, right, we’re going to have very interesting challenges with self-driving cars. People have raised these issues before. There it’s kind of hard to put the human in the loop, because the accident is over before the human can react. So, you need to be very careful.
Marc: In that direction, what would you say is the is the current challenge how one would formulate the Turing test? I’m having in mind this comic on the Internet: “Nobody can distinguish if you’re a human or a lightbulb”, because it’s just a limited interaction that you can have in a chat, for instance. Or when you think about the more advanced technologies like AlphaGo. If you would give it to someone who doesn’t know that the technology exists, this someone would probably not be able to distinguish if it’s a human or if it’s a machine.
Bob: No one has come up with a better definition of intelligence than the Turing test. But one has to lengthen the time period or expand the possibility of questions in order to capture the right notion. Somehow computers do very well in well-defined fields. They do very well at understanding images. But if you see the system, if you have the system, you can tweak a few pixels and turn a dog into a cat, or a horse, or a stop sign or something. These systems are very brittle. In the context in which they were trained or designed, even though we don’t understand what they’re doing, they perform better than people. But change the rules and all of a sudden, their limitations become apparent. So, we need to understand their limits as well as their potential.
Marc: Would you say that the artificial intelligence algorithms are used too prematurely at the moment by some companies, because they are overhyped? At the same time, if anything that promises AI is in the product sells better, they just use it even though they do not understand what it does.
Bob: There is certainly hype. I am not an expert enough in the field to be able to answer that specifically. I just say that one should be careful about hype and be careful about testing and evaluating. You know artificial intelligence has been through these cycles of hype and crashing. I am sure we’re now in a hype phase, so there’s going to be some crashing or gradual slow down or something. I think as we become aware of the limitations of these systems
Marc: Coming back to your fields, to algorithms. From my perspective as no complete out- or insider, it makes the impression that this artificial intelligence seems to be a key ingredient or building block to push algorithms to the next level. Because now we can bring something in that we could not have before, which one might call intelligence or fuzziness or something like that. Do you consider it an important building block having this artificial intelligence making some leaps now?
Bob: Not for the kind of things I do. I would love to be able to use this to analyse algorithms, but even very simple algorithms impose incredible challenges and AI is certainly not there yet. It’s really good for certain very important problems, such as image recognition, speech translation, it’s going to be important in medical situations and in law and so on. But again, I think you need the human interaction. Use the expertise of the AI system to get a start or get an idea, but then evaluate and put people in the loop somehow.
Marc: What would you say is currently the biggest challenge you are either working on or would love to find a solution for?
Bob: I have done mostly sequential algorithms over my career, where there’s one processor doing one thing at a time. I always get asked in lectures “What about memory hierarchies or multiple processors?”. So, the last several years I’ve been looking at concurrent algorithms, parallel algorithms. There have been decades of work in this area, but now with big data, big graphs, suddenly some of these algorithms can be used in practice. It turns out the engineers building the systems haven’t necessarily used the old theoretical results. So, there is vast opportunity to bring the old ideas into the new systems and actually develop new ones that match the current hardware. Those are the kind of problems I am now looking at.
Marc: A risk, as people often say, that comes with this digital transformation, is that lots of data is collected and then profiles of people are created. There are attacks to privacy by having digital personas and so on. What do you think about that? Is it that today, we have indeed multiple personas in the physical world and in the digital world that are coherent?
Bob: It’s not coherent, but they can all be tied together, which is a big problem. I think the EU took a big step. I think privacy is an important problem. People do not realize how much privacy we’ve actually lost. Security and privacy are only becoming more important and more important.
Marc: In that direction, at least in Germany I see a big problem for the education of children. How to get media literacy or digital-age literacy. This also has to do with truth and not-truth of information, for instance.
Bob: I have no good ideas as to how to educate school children, but we have the same problem with the scientific literature. I tell my students to always read papers critically, to question everything in the paper. Just because a great person wrote it or just because it is published, doesn’t mean it’s right. It doesn’t mean there aren’t things missing. Somehow, we have to train kids how to think, how to read sceptically, how to interpret what they see. I think it’s a big challenge.
Marc: Also in that direction would you say we have an over-flooding of information?
Bob: Yes and no. There is this phenomenon of compartmentalization of information, which is to say these systems try to feed you information on the basis of what you already looked at. There’s all this information but you end up getting focused down on some small part of it. You end up in some world which is different from everybody else’s world. We need some way to introduce randomness, tie things together, so people get a broader view. There are interesting challenges there.
Marc: Are filter bubbles something new?
Bob: Probably not, we just think of them as new. There’s nothing new under the sun, actually.
Marc: That is a good point. Digital transformation has positive and negative aspects. What would you say is the most negative aspect?
Bob: The filter bubble is a big one, but I think the privacy issue is another big one. I don’t have a view as to which one is more important. They’re both serious.
Marc: With privacy, I always see that there’s a huge difference between Germany and the US, for instance. In Germany people are at least a little bit more sensitive, thinking that data is first of all their own. Even if they don’t believe it themselves, the state takes care that data doesn’t get to everybody, you have to double agree and whatever. While in the US, it is more like “I give everything away” and then the fortunate one might catch it and do something with it.
Bob: I hope we don’t have a really bad experience in the US. I am a little optimistic, because since the EU is being more restrictive, and the US companies want to operate in the EU, maybe the restrictive policies carry over into the US.
Marc: Coming to the positive aspects. What would you say is the most positive aspect of the digital transformation?
Bob: Access to all the world’s information anytime you want it, as long as you can distinguish fact from fiction.
Marc: I have to ask a second thing; do you perceive that the skills of people changed from knowing something towards knowing how to find something? Is this a bad thing or a good thing?
Bob: It’s important to know how to find things. There’s only a finite amount of stuff we can hold in our heads. If there’s some kind of working set, you can really expand your ability to do work or to think about problems or to function in the real world if you can leverage information. But you can’t keep it all in your head at once. Learning how to learn is more important than knowing any particular thing.
Marc: I could continue interviewing you for hours, but the time is almost up. Thank you very much for the interview. Last question: “Being Human with Algorithms”, our motto, when you hear it, what does it mean to you?
Bob: Trying to understand how people interact with the digital world. How the real world interacts with the digital world. And how the world of people interacts with the world of machines.
Marc: Perfect. Bob, thank you very much.
Bob: You’re welcome.