My lab has a new website. The URL is
http://www.cog.brown.edu/research/causalitylab/index.html
Most of the design was done by my grad students and undergrads, but as part of the redesign, I wanted all of my grad students to write a short paragraph about their research. One of my students wrote the following:
"I play games with preschoolers, and then try to make computer programs that are as smart as preschoolers. The children usually win."
I thought this was interesting. It reminded me of a conversation I had when I was a graduate student 12 years earlier. I was in my fifth year of grad school, almost finished, and I was talking with a new student about what she wanted to work on for her first year project. Her response was that she wanted to build computer programs of children's behavior (she actually wound up doing something quite different).
It also reminded me of my own undergraduate thesis, written in 1991, in which I described a (pretty flawed) computer program that was designed to predict human behavior on a learning task. It was based on a model of animal cognition that was built ten years previously, which in turn, were based on a set of models designed in the early '70's.
The goal of building computer programs that model human behavior extends even earlier than that. In my University's Psychology Department, one of our distinguished professors has been working on a model of timing (in non-human animals) since the ‘50's. I heard him describe his research once (last year): He's still working on the model.
So, here we are: 50+ years of modeling. One of the promises of the "Cognitive Revolution" was the "Computational Model of the Mind" - the idea that the brain might act like a computer, and thinking about this metaphor might offer us insight into how the brain produces behavior and thought. In graduate school, I learned this as "Mind is to Brain as Software is the Hardware", a mantra that undergraduate students in the Introductory Cognitive Science class I TAed were forced to recite.
In cognitive development, various kinds of models have bubbled up to explain children's behavior - symbolic AI, connectionism, dynamic systems, causal models, Bayesian inference (which is now all the rage) - and each has advantages and contributions. But, I wonder whether the promise of building these models for developmental processes is akin to the promise of jetpacks when I was a kid: a nice idea, but more wish-fulfillment than actual science. A "wouldn't it be cool if we had these" idea, as opposed to a realistic endeavor?
Yes, this is a straw man (and yes, I'll knock it down in a few paragraphs). But, I've been thinking recently it's important to think about why this argument is a straw man.
My favorite computational model comes from the animal cognition literature. It's the Rescorla-Wagner model, first published in 1972. Basically, it was designed to explain a phenomenon in conditioning called blocking, which it did quite nicely. The model itself is flawed, and there have been numerous (and I do mean numerous) variations and recreations of the ideas they present in that original paper.
Here's what I like about the model. First, it does make some really counterintuitive predictions, which turn out to be true. Rescorla and his students went on to demonstrate several phenomena that emerge just from sitting around calculating possible inputs to the model. But more importantly, it also suggested a set of paradigms that emerged from trying to break the model. Almost immediately after its publication, Wagner and his students authored a few papers that demonstrated the model did not account for all of animal conditioning, and many of those phenomena came again from thinking about the predictions made by the model.
It's not the model that I like (although it's quite elegant) - it's that almost 40 years after it was published, researchers (including myself) are still thinking about what it does and does not predict, and using those predictions to design new experiments that explain human (adult and child), as well as animal behavior.
This is certainly part of the promise of the Cognitive Revolutions - models should act as formal systems - allowing developmental psychologists (and psychologists in general) with yet another way of thinking about the relation between theory and data. A good model doesn't just explain behavior, it makes predictions about what other behaviors we should expect to observe.
I suspect that the previous sentence I wrote is widely agreed-upon within cognitive science, and certainly among computational modelers. But, what strikes me about the field right now is that so few people actually bother to write it down (or even say it out loud). Even worse, I'm often struck how few papers actually practice this. A challenge faced by cognitive scientists in general, and researchers interested in modeling children's behavior is to stick to this principle. Otherwise, aren't we just pretending to build jetpacks?
