ESSLLI 2007

So ESSLLI this year was at home, in Dublin. Hardly the best destination - rainy, cold, expensive and totally devoid of tourist attractions. But maybe it makes people concentrate more on the scientific content ;-) I had more fun last year in Málaga, but it was still nice to go to ESSLLI once more, and Trinity is definitely a nice central location (although a bit overrun by tourists).

Most of the courses I attended were decent though none was spectacular. The best one was definitely the one by Joakim Nivre and Ryan MacDonald on data-driven dependency parsing: easy to follow, clearly structured, and taught by people who really know what they are talking about.

I hoped the machine-learning course would be interesting but the first day was quite disappointing and I didn't go during the rest of the week.

Other than that, one of the invited talks was on modeling language acquisition. It was a strange affair: the speaker (Ronan Reilly) showed how a supervised learning algorithm, namely neural network with error backpropagation can accurately learn a toy grammar derived from a corpus of child directed speech, using as training material sentences generated with this grammar. Not terribly surprising is it? The puzzling fact was that the speaker proposed this as the model of first language acquisition. I had to leave early so I didn't get to ask the obvious question: where do kids get their error backpropagation? I thought it was more or less generally acknowledged that there is little feedback on errors, and whatever there is, children mostly ignore. So a strongly supervised learning model doesn't say anything about first language acquisition. Or is it me who is really confused?

I don't know if I'll be going to another ESSLLI: next summer I won't be a student anymore (hopefully!) so it might have been my last one. But who knows; it's definitely quite an addictive event...