Compression Contest and Marctar's Axe

Marcus Hutter has posted a 50,000 Euro prize contest for compressing (a subset of) Wikipedia. While I applaud Dr. Hutter for his generous contribution to stimulating AI research, and I do believe that cognition often can be viewed as compression (as championed by Gerard Wolff), I do have one reservation with this idea:

The idea that a smaller theory is a better theory can be traced at least to Ockham's Razor. When I applied The Cruncher to the frames of a (very low resolution, but continuitous) movie, I found that the best compression was when the 1st frame was described, then every frame was described in terms of the previous frame (much the way that mpegs work). However, I don't think this is how people do it. I don't record my day by remembering when I woke up then every instant in terms of the previous instant. If I did, I'd have to spend a good deal of time "unpacking" my day to tell you what I had for supper (and presumably less time to tell you what I had for breakfast).

Perhaps a more extreme example would be Euclidean Geometry. This can be compressed down to the 5 postulates (plus some inference rules), but I don't think anyone rederives a commonly used lemma every time they use it.

Therefore, I propose an alternate to Ockham's Razor, which I'll call Marctar's Axe: The best model is that which has the smallest average query time (in terms of steps of computation to answer it). By "query" I mean things like "How many times have you seen this pattern?" or "What is the usual outcome of this pattern?".

My suspicion is that compression will usually fall out of Marctar's Axe, but there will be some cases where Marctar's Axe will tell you to "cache" results and thereby trade in a lot of computation time for a little bit of memory.

Human Subject Research

I recently had some dealings with the UMBC Institutional Review Board. All university research involving human participants and animal subjects is placed under review by this board to ensure some level of security in privacy, safely, and general well formed research methodology. We can, essentially, blame behavioral scientists that couldn't control themselves or their research (e.g. The Stanford Prison Experiment). During this review, I came across a recent undertaking by an MIT Media Lab professor, Deb Roy. A New Scientist article, Watch language grow in the 'Baby Brother' house, details the addition of audio and video recording devices to Professor Roy's home placed to capture the linguistic growth and development of his son.

Back during my time in the UMass Linguistic department, Tom Roeper gave me several notebooks full of observations made when his children were younger to transcribe. I'm hoping that more researchers take advantage of new technologies and share the collected data. Now, who is going to be the first to do what Deb Roy is doing but in a bilingual environment?

Without Numbers

Pirahã, a language spoken in Brazil by peoples number in the hundreds, lacks quantification and subordinate clauses. Daniel Everett, a Manchester researcher, has studied Piraha extensively and even supplies a sample of Pirahã sung speech and a gloss of a "Killing the Panther" story.

"Pirahã is a member of the Mura language family. It has been traditionally referred to as Mura-Pirahã but this obscures the distinction between the Mura language family and the Mura and Pirahã languages. The Mura language family includes Pirahã as well as the extinct dialects of Bohurá, Yaháhi, Mura and Torá. It has been suggested that the Mura family is member of the Macro-Chibcha phylum but there is insufficient evidence to support this view. Pirahã is the only language spoken in the region of the Maici. It has no external affiliates and is in danger of becoming extinct as well. Numerous attempts have been made to learn and document this language since first contact with the Pirahã in the 1850's. Most efforts were inadequate or relatively unsuccessful as the language is extremely difficult and unusual. The Everetts have devoted the most time, collected the most data and been the most successful at learning the language and providing linguistic analyses for several interesting features of Pirahã." (link)

In 2004, Peter Gordon of Columbia University published an article in Science, Numerical Cognition Without Words: Evidence from Amazonia, and recently brought linguistic relativity back into the mainstream media with recent coverage in a Spiegel article about Pirahã. Daniel Casasanto counters with a letter to Science with two main points: Gordon's work lacked a control, baseline performance group and Gordon demonstration fails to support the causality claim. Another letter provides additional criticism.

Cabspotting for Inference

Here is one of the Exploratorium's coolest projects: Cabspotting (BoingBoing and Metafilter discussions). The visualization plots in real-time a cab (a circle glyph that is filled in only if the cab has a passenger with a comet trail where the cab has come from) in San Francisco and prior trips (phantom, wispy lines).

Now, let's consider the above screen capture that I made of their Flash client. There's a clump of cabs on Market Street and near The Embarcadero which is clearer on the Google map of San Francisco. The southernmost cab seems to be going from Market Street to pick up someone and the westernmost cab was doing the same and has picked up the passenger. From these data we can infer "cab intent," major areas of commerce, and traffic dynamics. For example, consider the amount of time and distance it takes the cab to pick up a passenger and with a bit of taxicab geometry we might discover short cuts.

You can email the fine folks from their FAQ page which provides more project extensions. I just emailed them for access to their API and Flash client. Perhaps I'll strap GPS to something and display that and more.

More Squiggly Lines

I realized that someone could use spectrograms to compose as well as recognize sounds. That is, with enough training, I could think of a sound (virtually any sound), then imagine what the picture of the sound must look like and draw it. It's like mastering an instrument: The Pickettheremin Musical Whiteboard.

With this in mind, I threw together a quick script so that I could point my webcam at a whiteboard, press a button on Lappidactus, then listen to the (...ahem) beautiful muzak [sic] produced.

Below is my first composition (both the sound and its spectrogram). It's called "Squiggly Lines I: John Cage Shooting a Laser Shotgun".