Back to my Paris index.

Date: 20060712

weather: sunny, hot-ish

wake: usually around 9:30

It's been so long since the last update that I have no idea where to begin this one. To be honest, the logs you just read between the first of July and this one were written with the benefit of considerable hindsight.

I can remember writing a lot of email, a lot of thesis-writing, and a lot of being stressed about the gf!

The email. First there was an email from Dan (Yoshua's masters student) soliciting thoughts and opinions about a generic optimization software that he thought he'd write. "I have some ideas" popped into mind. By the time everyone was through, five or six people were involved, thousands of lines of email had been written, two enormous wiki pages had been created, and many new ideas for ways of doing PLearn had surfaced. Dan was not a major contributor to the ensuing discussion, I suspect that he toiled away on the concrete and relevant problems he knew about, and let the rest of us duke it out in the domain of hypothetical consequences. I spent many hours on this.

The thesis. I made great progress in the 'background' section, filling in many of the sections that had long been marked by placeholders. Then I got to the experiment section, in which I describe in full detail, procedure I will use to test my hypothesis. One of my proposed experiments is to construct a table, in which several algorithms will be compared on the task of 'fitting songs' with probability densities. I maintain I had good reasons for doing this, really I did. However, when it came to describe, in full detail, the procedure I would use to evaluate the fit of a song, I stumbled. There was a faint, yet distinct figurative stench of circular reasonning. I wondered, speculated, rolled my eyes, and screwed up my face, but the circular logic got clearer and clearer. There was a major problem with my thesis.

But some quick backtracking circumvented the logical flaw. A long time ago the memoir plan had been to report on genre prediction on a big dataset instead of a small one (as I had already done). The reason I didn't do that, was that I thought it would be a pointless thesis. Evidence of the pointlessness of such a task was that there was no large database of songs labelled with genre. Further contributing to that sentiment of uselessness was the knowledge that musical-genre is a fickle taxonomy. Who would care if a program was able to correctly recognize an mp3 as containing rock? Hypothesis: no one, with high probability. But now things might be different if a sort of bargain can be struck with my new friends at last.fm. There are these well-defined (by consensus) things called 'tags' such that predicting them would contribute in a very real sense to an automatic, dynamic, record publishing agency. Perhaps naively, I think the RIAA should be abolished, and my thesis could help to do it!

So, I think: megalomaniacal rant goes in intro... editing required throughout remaining sections.

Stress about the gf! Aack! We are living together here in gaie paris, and thinking of moving in together in Montreal. Will this work out? Am I a grown-up? So many questions, hard questions. Wish us luck folks, if you like.