That's what chemicals can do - academic

Sun Niagara 2 Open Source!

nospam@example.com (Greg Laun) — Wed, 12 Dec 2007 18:48:14 -0500

Have you ever thought to yourself, "I wish Sun would opensource its Niagra 2 CPU architecture?" I know I have. Really, I have.

Well they've done it. Why is this cool? Well, it means more people can chip in on chip design. The Niagara 2 can be used for supercomputing. If people are able to produce low-cost spinoffs, then CPU prices might come down for supercomputers or for computers overall.

This latter part is good because lower CPU costs are needed for projects like One Laptop Per Child.

Here's the slashdot link

Eigenfaces

nospam@example.com (Greg Laun) — Tue, 11 Sep 2007 13:42:41 -0400

Today in the class I'm TFing we covered some of the research on preferences for symmetric faces, which is apparently a preference that people have since birth. Basically, the idea is that people seem to prefer faces that are highly symmetrical to those that have noticeable asymmetries. This preference tends to be measured by asking which faces are the most attractive.

Things could have gone the other way. Many highly attractive celebrities have freakish asymmetries. Angelina Jolie has huge lips, for example. In fact, for all the times I've seen the symmetrical faces demonstration, I never find the completely symmetrical faces to be all that attractive.

Typically, symmetry is achieved in the lab by morphing several faces. If you average enough faces, the asymmetries of each becomes less prominent and you more smoothed. This has led to speculation that symmetrical faces are attractive precisely because they are average in some sense. Perhaps evolutionarily they represent the genetic central tendency and that somehow represents health. Or something.

But a more interesting (and I think more credible) theory is that symmetrical faces are somehow easier to represent. Our Professor Frank Keil mentioned a version of this account today, basically that average faces are prototypical and they allow you to represent other faces by the difference between that face and the target face. I was surprised at this because I had never heard a psychological researcher articulate this position. I am not strongly versed in this field, but I have heard it discussed several times and attended several talks on the subject. Nearly always, the preference for symmetrical faces is tied to an evolutionary story rather than a representational one.

I like the idea that average faces are easy to represent, and for a while I've been a fan of the idea of eigenfaces. Or maybe I just like anything with "eigen" as a prefix. Or "eigenfix" as I call it.

The idea of an eigenface is that you take the average of a set of faces and the eigenvalue of that average matrix (or some transformation of it) is used as the face prototype. Then each additional face is represented as a vector of what must be added or subtracted to the eigenface to get the target face.

Eigenfaces are used in computer vision to do face recognition. There's an interesting demo at the media lab that shows a little how this is done.

I am not sure how much research has gone into the possibility that humans represent faces as eigenfaces, but I think it's an interesting topic.

First, it could possibly explain the preference for average faces by the fact that these have less representational overhead. This representational overhead does not have anything to do with attractiveness, which makes the attractiveness of such faces (as separate from simply a preference) something more of a mystery.

Second, it explains mathematically and in a testable way how face representation might work. This makes for good science.

Third, it's possible that many sorts of concepts are represented with eigenconcepts. This is somewhat similar to a prototype theory of concepts, but with more realism about what representations could look like. For example, there are several easy matrix transformations that you could perform on eigenvectors to yield concept variations that couldn't be predicted simply be reasoning about concept stereotypes. I had this theory figured out much more when I was reading more about concepts, but I have confidence that it's not completely a dead end.

That said, while the general metaphor of a prototype face may be floating around in psychology, the actual calculation of eigenfaces requires too much math to be considered by most mainstream cognitive scientists. If we stop talking about matrixes of 3-d images and start talking about probability distributions over those 3-d matrices the math is even harder and less likely to be worked out. For those reasons, I don't imagine that computer science and psychology will share much information with each other as time goes by.

A friend of mine thinks that computer scientists and biologists are going to solve all of psychology's problems before we do. I hope that's not true. I would be very interested to see more collaboration among those fields.

What are Political Blogs For?

nospam@example.com (Greg Laun) — Thu, 06 Sep 2007 19:27:43 -0400

I had a (short) discussion with a friend of mine today about why I think blogs should be replaced by wikis in any place we want the blogs to have a political effect. Basically the rationale for this is that blogs produce information that's dependent on other blogs (since they tend to be read primarily by other bloggers). In addition, since these blogs usually exist in a community of like-minded people they tend to explore the conceptual space of any topic extremely inefficiently and predictably. The problem becomes worse the more bloggers communicate with each other, but then this is often the main point of blogging to begin with.

Wikis suffer from fewer of these faults and can be made more anonymous because blogging requires authenticating into the blog software. Wikis eliminate the need to feel that you are expressing a new opinion and hence reduces inefficiencies in exploration of the conceptual space. Moreover, wikis integrate information whereas blogs simply create a large series of poorly informed opinions. A series of poorly informed opinions can be aggregated into a useful set of information, but only on the condition that the opinions aren't dependent on each other, which they are in blogging. Integration in a wiki is more straightforward, especially if it's edited by people who disagree.

My friend responded by saying that she doesn't think the goal of blogs is to accurately represent anything but to offer an individual voice. It's possible that finding one's individual voice is a good unto itself. But if so, then it can be achieved equally well with a pen-and-paper journal. Such journals don't carry with them the cost of potential imprisonment or the technology necessary to blog. On the othetr hand, it's less likely that finding one's individual voice irrespective of accurately representing anything is useful as a means for encouraging democracy. Finding one's voice leading to democracy has been a long-standing belief of several American liberals, but has rarely produced anything of worth and has severely stalled many projects that could have helped people in need. This argument is made rather well in the book Nation of Rebels: Why Counterculter Became Consumer Culture. The basic argument is that "finding one's own voice" is taken to be the antithesis of the capitalist monolithic culture but in reality is just the same thing. The result is that people think that things like art therapy or breakdancing are effective cures to the world's ills and oppose actually useful things.

Blogging could have the same sort of voice-finding use that art therapy does, but I'm not sure that has any impact at all on democracy or freedom. Usually when we talk about blogs being good for democracy we have in mind a distributed network of journalists. But journalism is about accurately portraying something, not being an individual voice.

Best Statistics Book

nospam@example.com (Greg Laun) — Sat, 07 Apr 2007 11:36:49 -0400

I recently read Gelman and Hill's book Data Analysis Using Regression and Multilevel/Hierarchical Models. It very nicely explains classical regression (including a good amount of GLM) and then very nicely explains Bayesian inference and multilevel models. The exposition is both intuitive and mathematicall rigorous without being too demanding.

This is by far the best statistics book I've read for the social sciences and I've read a few by this point. I highly recommend it to anybody in psychology and plan to use it as a text whenever I teach statistics or research methods.