I'd really like to see the follow-up on how to decide which data to actually use. Right now, it's pretty unsatisfactory and I'm left quite confused.
(Unless this was an elaborate plot to get me to read Judea Pearl, whose book I just picked up, in which case, gratz.)
The short of it, having read a few of Pearl's papers and taken a lecture with him, is that you build causal networks including every variable you think of and then use physical assumptions to eliminate some edges from the fully connected (assumption free) graph.
With this partially connected causal graph, Pearl identifies a number of structures which allow you to estimate correlations where all identified confounding variables are corrected for (which can be interpreted as causation under the assumptions of your graph).
Often times, it seems like these methods only serve to show you just how bad a situation "estimation causation" actually is, but it's possible to design experiments (or get lucky, or make strong assumptions) so as to turn them into useful tools.
The level of "trust" you have in a person should be inversely proportional to the sensationalism of the claim that he's making.
If a person tells you he was abducted by a UFO, you demand evidence.
If a person tells you that on the way to work he slipped and fell down, and you have no concrete reason to doubt the story in particular or the person in general, you take that at face value. It is a reasonable assumption that a perfect stranger in all likelihood will NOT be delusional or a compulsive liar.
DP
That makes sense if you're only evaluating complete strangers. In other words, your uncertainty about the population-inferred trustworthiness of a person is pretty high and so instead the mere (Occam Factor style) complexity of their statement is the overruling component of your decision.
In the stated case, this isn't a totally random stranger. I feel quite justified in having a less-than uninformative prior about trusting IRC ghosts. In this case, my rationally acquired prejudice overrules in inference about the truth of even somewhat ordinary tales.
It's supposed to be a statement, not a question - have a look a the two in English.
Egad, true. I jumped on seeing the juxtaposed question mark and then messed the English grammar.
Perhaps it's not about the ad hominem.
"Rationality is whatever wins."
If it's not a winning strategy, you're not doing it right. If it is a winning strategy, overall in as long of terms as you can plan, then it's rationality. It doesn't matter what the person thinks: whether they'd call themselves rationalists or not.
Picometer nitpick, for accuracy:
你有鼻子, as phrased, is not a question and thus ironically even more bewildering, not that someone who couldn't understand the utterance would be able to determine that. To phrase it as a question you need a different form; one of
你有鼻子吗？ 你有没有鼻子？ or 你有鼻子，对不对？
would work. The first is a simple question. The second leaves a bit more credence to the possibility you don't have a nose. The third probably is trying to imply that if you don't agree then you're foolish.
See antiprediction.
That's certainly sensible, and in But There's Still a Chance Eleiezer makes examples where this seems strong. In the above example, it depends a whole lot on how much belief you have in people (or, rather, lines of IRC chat).
I think then that your strength as a rationalist comes in balancing that uncertainty against some your prior trust in people. At which point, instead of predicting the negative, I'd seek more information.
Doesn't any model contain the possibility, however slight, of seeing the unexpected? Sure this didn't fit with your model perfectly — and as I read the story and placed myself in your supposed mental state while trying to understand the situation, I felt a great deal of similar surprise — but jumping to the conclusion that someone was just totally fabricating is something that deserves to be weighed against other explanations for this deviation from your model.
Your model states that pretty much under all circumstances an ambulance is going to pick up a patient. This is true to my knowledge as well, but what happens if the friend didn't report to you that once the ambulance he called it off and refused to be transported. Or perhaps at the same time his chest pains were being judged as not-so-severe the ambulance got another call in that a massive car pileup required their immediate presence.
Your strength as a rationalist must not be the rejection of things unlikely in your model but instead the act of providing appropriate levels of concern. Perhaps the best response is something along the lines of "Sounds like a pretty strange occurrence. Are you sure your friend told you everything?" Now we're starting to judge our level of confidence in the new information being valid.
Which is honestly a pretty difficult model to shake as well. So much of every bit of information you build your world with comes from other people that I think it pretty decent to trust with some amount of abandon.
I like to think Einstein's confidence came instead from his belief that Relativity suitably justified the KL divergence between experiments in 1905 and physics theory in 1905. He was not necessarily in full possession of whatever evidence was required to narrow the hypothesis space down to relativity (which is a bit of a misformulation, I feel, since this space still contains a number of other theories both equally and more powerful than Physics+Relativity) but instead possessed enough so that in his own mental metropolis jumping he stumbled across Relativity (possibly the next closest convenient point climbing from the prior of Physics to the posterior including new evidence for the time) and sat there.
His comment just reflected a belief that new experiments were unlikely to yet be including the same new information he already used. In some sense, their resolution was not yet strong enough to pinpoint something more precise than Relativity.
Not to knock Einstein, of course. Just because you have new evidence drawing you to a different posterior hypothesis doesn't mean that the update is going to be easy. That's perhaps where the philosophy of Bayes runs into the computational limitations of today.
= f037147d6e6c911a85753b9abdedda8d){kind=logo}
= 27d567bc2d48d9f40e9ccc20ea370b15)
= b715d02e85f7b3b4ae65ca3d3e450868)
Subscribe to RSS Feed
Right, so the challenge is to incorporate as much auxiliary information as possible without overfitting. That's what AdaBoost does - if you run it for T rounds, the complexity of the model you get is linear in T, not exponential as you would get from fitting the model to the finest partitions.
This is in general one of the advantages of Bayesian statistics in that you can split the line between aggregate and separated data with techniques that automatically include partial pooling and information sharing between various levels of the analysis. (See pretty much anything written by Andrew Gelman, but Bayesian Data Analysis is a great book to cover Gelman's whole perspective.)