McAfee SiteAdvisor Blog

Michael Kearns is a Professor of Computer Science at the University of Pennsylvania and was previously the head of the AI group at AT&T/ Bell Labs. He is also an Advisor to SiteAdvisor and has been an active contributor to the development of some of our core algorithms. One thing we particularly like about Michael is his broad set of interests, ranging from computer security, economics, game theory, and machine learning to fiction and the visual arts. Here we provide a lighthearted interview we recently conducted with him.

Hi Michael. Can you start by telling us a little about yourself and your interests?

I'm a computer science prof at the University of Pennsylvania, generally interested in machine learning and artificial intelligence. In the past few years I've become particularly fascinated with the interactions between economics, game theory and "traditional" computer science problems. I've also always had a left-handed interest in security --- my long-ago PhD dissertation showed a natural connection between "hard" learning problems and the RSA public-key cryptosystem. I guess someone had to do it.

How did you get interested in computer science?

Like many, I suppose it was initially through the power of programming --- the realization that with very little effort, I could get a machine to accomplish things I could never do myself. Once I learned the mathematical foundations of computer science, it seemed incredibly deep and modern at the same time, a hard balance to accomplish. Eventually I became lazy and came to enjoy proving that something could (or could not) be programmed rather than actually doing it --- the polite prefix for this is "meta".

Can you explain your view of what Machine Learning is?

Well, I'll tell you what I like best about it, which is the fact that it spans the range between very practical, powerful and widely used algorithms to fundamental insights about inference, learning and other topics in philosophy. It's not alone in computer science in this regard. But being able to trace a research idea from something as abstract as Occam's Razor --- which asserts a basic connection between learning of any kind and a very general notion of what you might call data compression --- to algorithms as useful as boosting or support vector machines, is very satisfying.

You teach a really interesting undergrad course at Penn on "network analysis" where you talk about the Tipping Point, Small Worlds, and related things. Can you tell us a little about this?

The class is called "Networked Life" and I think one of its most novel features is the fact that it is open to all majors and all levels, despite the fact that we cover fairly technical material. In the beginning we focus on those aspects of network science that are dominated by the metaphor of viral spread --- word-of-mouth marketing, fads, etc. --- and examine the structural properties of networks that aid or hinder such spread. Later in the course we move on to what I call the "dynamics of rationality" on networks --- distributed processes involving economic or strategic behavior. The class is great fun to teach, and we hold lots of experiments with the students as the subjects. I give away a lot of money over the course of the term, which the students seem to enjoy.

A friend of mine said that a good test of the importance of any new technology is whether you could imagine Oprah talking about it on her TV show. The "Oprah test" strikes me as a nice mental exercise for determining whether a technology is just for "digerati" or could ever go mainstream. Should Oprah care about Machine Learning?

Probably not in its pure form, as opposed to being instantiated in an important application. A colleague of mine once said that machine learning is the second best way of solving any problem, by which he meant that it is inevitable that as people understand specific applications, they can gradually replace the "learning" with hard-coded domain knowledge. Its power is in its generality, which is unfortunately not a subtlety I see Oprah jumping on any time soon. Now David Lee Roth on the other hand...

In your opinion, what are the most interesting current research problems related to computer security?

I'm interested in modeling the behavioral aspects of security, which is something I don't see the "mainstream" security community taking seriously yet. They tend to focus more on the technical aspects of security, both the exploits and their defenses. Many security problems have a strong behavioral or even strategic flavor --- my willingness to invest time or money in security measures may depend strongly on the security practices of those I am closely "connected" to, in a variety of senses.

Can you tell us a little bit about why you think economics is relevant to computer science?

When I was an undergrad studying CS at Berkeley in the 80s, networking research tended to focus on highly structured topologies like butterflies, grids, etc. The unspoken premise was that centralized design of centrally managed system was the norm. Today nothing could be further from the truth. At the very least, economics provides the right language to describe the most important properties of complex networks like the Internet: decentralization (in every respect), heterogeneity, incentives, competition... There's a long way to go before it has practical impact but one could argue that, whatever you want to call it, CS has no choice but to adopt "economic" thought going forward.

If you were stuck on a desert island with only 1 algorithm, which one would it be?

Anything by Christos Papadimitriou: algorithm, reduction, novel... For the unfamiliar, he's the man who gave Bill Gates an Erdos number.

Oh yeah -- what's your Erdos number?

Mine is three: Kearns to Mansour to Alon to Erdos.

Gotta work on that, Michael! Ok, now for the big cliché AI question. Do you think computers will ever pass the Turing Test? If so, when?

No, because I don't think humans could pass the Turing Test as it tends to be phrased and administered. Imagine if actual human interaction consisted of us always suspecting each other of being "just machines", and agressively driving every conversation towards the other party's areas of greatest incompetence so as to unmask them, like at the end of a Scooby-Doo epsiode... it's entirely unnatural, and ignores all kinds of social conventions people routinely rely upon. I think the Turing Test is thought-provoking at a cocktail party level but not a productive yardstick for an important research agenda.

But we do need something to replace it --- the currently established metrics of AI tend to be overly specialized.

What books are you reading right now?

A 300-year history of Philadelphia. Don't ask. But did you know that Ben Franklin invented the internet?

Ha ha. Yeah, that guy really seemed to invent everything! Speaking of books, if you recommended one book to our readers, what would it be?

Infinite Jest, by David Foster Wallace. A gripping tale of tennis, addiction, and Matlab.

Been meaning to read that one, but it's more than 1000 pages! Michael, thanks so much for your time and talk to you soon.

-- Chris Dixon