langreiter.com plain, simple: comments-2008-07-13

LANGREITER.COM plain, simple

START

INDEX

Sonntag, 13. Juli 2008

A Conversation with Leo Breiman (2001)

"The interests and accomplishments of Leo Breiman extend outside the areas of professional statistician and probabilist. He was a waiter in the Catskills, a dishwasher in the Merchant Marine, a trekker into the heart of rainforest Africa, an active father to many children from a small agrarian Mexican village, a member and President of the Santa Monica School Board, the architect of his stunning home and an accomplished sculptor."

If you're only tangentially interested in ensemble methods like Bagging, [create AdaBoost] &c., this is highly recommended reading.

"On most data sets that people have looked at, Adaboost did quite a good deal better than bagging did. This was a startling discovery because you could take a sow's ear and transform it into a silk purse. That is, you could take a classiﬁer like, say, everyday vanilla CART, which was good but not a great classiﬁer, and by using this Adaboost algorithm, which was almost trivial to program, just iterated calls to CART, turn it into a world-class classiﬁcation algorithm that, by almost any standards, had accuracy as good as anything else out there, and better than almost everything else out
there."

I guess the interview was conducted at about the same time as Leo Breiman was working on the first Random Forests publications. Random Forests were recently shown (via) to perform splendidly in high-dimensional supervised learning settings.

2008-07-13-randomForestry

There's a relatively interesting discussion on Thrift and Protocol Buffers on Stuart Sierra's blog; apparently, my recollection that the first was massively inspired by the latter wasn't all wrong: [create Mark Slee] interned at Google before working at Facebook.

The coolest of all is still Q IPC, tho.

2008-07-13-thruff

Git: The simple things are simple. The amazing things are amazing.^*

I spent (quite) a few hours last week working through assorted Git materials (mostly prompted by a mind-bending crash course administered by no one else than earl; but also by a looming sense of eventual inevitability). As far as I understand it right now, Git looks like a neurosurgeon's chainsaw when compared to the humble table knife that is Subversion.

What follows is the curriculum I'd stick to if I had to start afresh; since I'm most certainly still a Git newbie, take it with an appropriately sized grain of salt.

Linus Torvalds's talk at Google (optional). Gives the backstory and motivation; describes the style of development organization Git was made for. Might even entertain some.

Git tutorial, part 1 (required, via Hannes). Short & sweet.

A tour of git: the basics (required). Nicely paced walkthrough; probably the best single-document introduction. Not surprisingly, covers essentially the same material as part 1 of the Git tutorial, but from a slightly different angle. This helps to reinforce.

Git tutorial, part 2 (required). Judging from del.icio.us evidence, many skip this part. I wouldn't. Goes well with Tommi Virtanen's graphical illustration (kthx earl).

Ryan Tomayko's The Thing About Git (required). Excellent propaganda, enticing use cases: How to untangle the tangled working copy.

Everyday Git With 20 Commands Or So (optional). When you understand the commands without having to look at their description, you're done with Git 101.

More: earl's git-notes, del.icio.us/chl/git.

^* With apologies to Alan Kay.

2008-07-13-got

no comments

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