The life of a computational linguist III. - Interview with Hugo Liu taste researcher

This interview is a little digression from our series, The life of a computational linguist. We talked with Hugo Liu who's the Chief Scientist at Hunch.com and a research affiliate of the MIT Media Laboratory and MIT Comparative Media Studies Program, where he has taught courses on artificial intelligence, and philosophy of aesthetics. He's got a blog at larifari.org and you can find him on twitter as @dochugo

Számítógépes nyelvészet: Please tell us something about yourself.


I'm a taste researcher at the MIT Media Lab, and the Chief Scientist of Hunch.com, an ambitious recommendation engine for the web. I'm interested in making technologies that help people live a more stylish and fulfilled life.

You have broad interests (from AI to Aesthetics), but in today's world, there is a trend towards being specialists. How can you manage to be informed on these various topics? Would you advise a student to become a specialist early, or just to go with the flow and explorer what she/he wants


Always try to pursue interests that you are passionate about. If you care about what you're doing, you are much more likely to succeed. However, it's equally important to cultivate a set of technical competencies, so you can deliver your ideas into reality. For me, having a technical skill set consisting of data modeling, machine learning, computational linguistics, and information design has enabled me to effectively research problems in the aesthetic realm, from emotions and gender, to food and identity.

I think that inevitably, some students will be ready to specialize early, while others will do better to explore for a long time.

But that does not matter as much, in my opinion, as making a habit of tinkering and creating things on a daily basis. It can mean building a robot, writing an essay, or completing a statistical analysis. People at MIT call this "hacking." It's the sine qua non of student culture at the Media Lab.

There is a shift in your work from AI and natural language processing to 'taste fabric'. Does this reflect any kind of change in your thinking? (I mean in your early work, you made Common Lisp and Python NLP tools, and published on using ConceptNet for reasoning about knowledge. I sensed a tone of criticism towards machine learning approaches in those papers. But then you end up with taste fabric which turns to 'soft semantical methods' and propose ML methods to investigate taste similarity.)


You're right about there being a shift in my work. I'm growing and evolving like everyone else. I started my career as an undergraduate researcher working for Marvin Minsky, one of the founders of Artificial Intelligence, and his brilliant protege who became a great colleague of mine, Push Singh. Marvin and Push belong to the "Strong AI" school, which means the pursuit of algorithms that understand and mimic human cognition at a high level of abstraction. This greatly shaped my early research approach, which emphasized cognitive models (such as Emotus Ponens and What Would They Think?) and inference (ConceptNet).

When my research took an aesthetic turn, I became fascinated with modeling creativity, food perception, happiness, identity, gender difference, and so on. I found that, for this class of problems, statistical methods offered better research opportunities than strong AI approaches such as inference. One reason is that aesthetics tends to be shaped by unconscious processes that we don't yet have great cognitive models for. Another reason was that, from 2004, the web saw explosive growth in blogs, social network profiles, and other rich web corpora that could drive statistical analyses. I don't mean to suggest that one approach is more valid than the other. I just think that a researcher should be driven by pragmatism (rather than dogma) and apply whichever methods are most promising in a given situation.

How did bring your knowledge of taste fabric into the real-world at Hunch?


Hunch applies state of the art machine learning algorithms to the problem of recommending anyone anything. It's quite ambitious in that regard. But as important as the algorithm, Hunch has an amazingly rich dataset about people's preferences, personality and behavior, gathered from tens of millions of people's answers to multiple-choice questions about themselves. Hunch is quite literally powered by a 'taste fabric' which we call the taste graph, and as of August 2010, it consists of over 10 billion connections.

You are an academic researcher and Chief Scientist at Hunch.com. Do you think that it is essential for profs to get some industry experience? Is there any conflict between the two personas of yours?


I don't know if industry experience is essential to being a successful academic. Many of my successful professorial friends aren't at all interested in industry. For me however, it's been incredibly fulfilling to see my dissertation research into computational taste modeling have a direct impact in the real world. The only conflict I feel is not having enough time to do everything I'd like to.