langreiter.com plain, simple: comments-2007-09-23

LANGREITER.COM plain, simple

START

Sonntag, 23. September 2007

Alles Gute, Vitus!^*

Gross is er g'worden! Und immer noch kein Blog. (Weiter so!)

Der Papa baut derweil lässige Sachen wie den Google Link Enhancer.

^* Hoffentlich führt das zu keinem Backdating-Skandal ...

2007-09-23-gluwu

Kevin Kelly, triggered by Gadget Ads: "When I was part of the team making Wired magazine a decade ago, half the battle at launch was landing the right cool ads. We had to convince the advertisers to join (and pay) us. But what if we could just choose the cool ads we wanted, without having to ask permission? [...] This may be where we are headed."

2007-09-23-kkad

Hal Daumé: '[W]e should never be marking up data in ways that it's not "naturally" marked up. For instance, part-of-speech tagged data does not exist naturally. Parallel French-English data does. The crux of the argument is that if something is not a task that anyone performs naturally, then it's not a task worth computationalizing.'

2007-09-23-halnl

Three-quarters thru Josh Tenenbaum's ICML 2007 lecture (Bayesian Models of Human Inductive Learning). So far, really intriguing; inductive biases are necessary for learning, strong inductive biases greatly facilitate learning from very few examples (remember the old Chomskyan question?) — and, best of all, inductive biases themselves (e.g. the Shape Bias) can be learned (he cited a rather spectacular result from developmental psychology: 17-month-olds who were trained on labelling [artificial] categories of things [e.g. "wibs" and "zubs"] for 2 months [only 20 mins/week, no worries] commanded roughly twice the real-world object name vocabulary compared with their control group - while the number of other words used was comparable).

2007-09-23-tenCat

no comments

Please log in (you may want to register first) to post comments!