"a male snow leopard rarely exceeds 100-120 pounds — hardly more than a big pet dog."  Are you thinking what I'm thinking?

A bargain at E4000.

Conversation with Kristina suggests an alternative hypothesis: going to a Big Name school raises your income in every profession; students have a target income level, and would rather do something socially redeeming/fun if it doesn't cost them too much; therefore students who go to Big Name schools don't earn more, on average, but might have a broader range of jobs.  Testable... Anyway, there are obviously big problems of self-selection and self-regulation involved here.

They're not actually doing anything _flagrantly stupid_, except that they persist in talking about discrete generations when all their distributions massively overlap. Model specification FAIL.

I seriously hope that the report itself is at a higher level than the 1953-vintage headline and unfocused noodling of the piece. (Since Tom Mitchell is involved, I expect so.) N.B., as a socialist, i am all in favor of the technological elimination of toil and drudgery, a.k.a. automation getting rid of crappy jobs...

This is really not good. Strogatz knows much better --- why is he doing this?

Even if this is a joke, it's great. I should definitely consider assigning this the next time I have to teach financial time series...

"Self-taught learning" = use unlabeled data to learn some sort of abstract, higher-level representation/set of features; then use those features in straightforward classifier learning on the labeled data. Nice idea, cute examples; completely separable from the specific technique for learning sparse basis vectors.

The end of Swiss banking as we knew it.

Stephen Quake explains the facts of scientific life. As someone who is horribly unsuccessful with grants himself, I really sympathize. But on the other hand I'm often on panels reviewing grant applications, typically ones which are really looking for something which sounds both new and not-crazy. (I think "this is too ambitious" is best translated, in blunter terms, as "we don't see the evidence that _you_ can do it".) There is a type I/type II trade-off which is inescapable, but also an institutional incentive to favor type II errors, since the loss of projects you don't fund and never happen is much less visible than the waste of projects you do fund and go nowhere.

(And please don't talk to me about the private sector as an alternative. It never funds basic research, except when it has an enlightened and secure monopoly, insulated from shareholder pressure, like Bell Labs in the golden age; research goes to the wall as soon as any of that is threatened.)

Interesting, but all the usual problems with trying to squeeze causal blood from time-series stones seem to be present in abundance. Also: why oh why can't we have economists who are at least able to use generalized additive models?

Adopt a sloth!

Convex optimization, the book.

The jokes write themselves.

File under: people who do not grasp the meaning of the term "controlled experiment". See also: "ecological fallacy", "massive confounding".

Update: I gave in: http://bactra.org/weblog/592.html

Targeted at Michigan but probably good for much of the northeast.

Zoopsychopharmacology.

I don't know whether to snort or to sigh wistfully.

If only there was some institution which could - approximately, to be sure - aggregate all kinds of information about trade-offs and alternatives into simple, one-dimensional signals!

I must say that only now starting to look at supply seems - unusual.

*radiates intense guilt over his eating habits*

Learn less, live longer.

Superficial profile of the great Daphne Koller.

Needless to say, my prejudices are abundantly confirmed. (But I should really track down the original paper.)

Michael Pollan piece that makes me want to scream at him even when I agree with him on specific points. Gaaaaaarrrrrgh.

Unusually blatant pressure from big donors, on the lines of "Nice political party you have there --- be a shame if anything happened to it".

39 pp. tutorial on ways of bounding the error of your learning algorithm. Looks handy.

Query for the reader: can you find anything in this story which explains _why_ the go'vt settled? This would seem like an important fact...

Fuck, fuck, fuck.

Helpfully color-coded by season!

I'm sorry, but this has to be pure confounding.

"We consider the problem of estimating the parameters of a Gaussian or binary distribution in such a way that the resulting undirected graphical model is sparse."

Another neighborhood house just a little out of our price-range...

Note: those who think intelligence is innate learn less.

IPAM workshop on the mathematics of search and knowledge discovery, with links to slides and/or audio for some talks

More on lead as a source of poor self-control and hence crime

1996 Times write up of study. Interesting if true, but I always wonder about the stats. Done in Pittsburgh - see if data are available.

Coca Cafe is indeed very nice.

Nice photo set.