lliiffee

Is it typical to interface to Newton's method via a function that computes the inverse Hessian? I've never seen that. Typically, people claim it is numerically unstable to explicitly invert the Hessian, and that it would be better for the interface to take the Hessian itself, and then call a subroutine to do a linear solve.

Doesn't seem to work in firefox? (The search bar keeps getting activated.) I think I have the "automatically search on typing" option selected though.

Great questions! I can answer a few of them.

- Yes, you can get stick in a local maximum. However, see the k-means++ algorithm for a clever initialization scheme that gets you within a good constant factor of the global maximum.

- The best way to pick K is to fit the algorithm with a range of K and see which K seems to give the best application performance.

- It is "your job" in creating the feature space to ensure that feature-space nearness corresponds to application-space nearness. This is not easy at all, and essentially the only reliable test if you've done a good job is if the final algorithm performs as you'd like it to.

I think that the post is really observing that, once you understand it, ML algorithms aren't really that complex. However, once you understand it, most things aren't that complex, and that doesn't make fully understanding them easy in the first place.

I have to disagree. The potential downside for a flaw in the new bill could be rather catastrophic damage to the economy. The downside of a failure with the Mars rover is that the rover just doesn't work. Thus, it makes sense to be dramatically more conservative in designing a new currency.

I totally agree-- the problem with reviews isn't technological. Reviewers fundamentally need better incentives to do a good job, which I'm not sure how to do.

There are lots of things you might consider, that would be impossible with snail mail. For example, the reviewers might be able to send regular emails with questions to the authors, or the authors might be able to constantly update the paper in response to initial reviewer comments, there could be an anonymous "chat room" of some kind, etc. Not sure if any of these are good ideas, but I think it is a fair point that most journals simply use the internet as a "faster snail mail".

That said, I review for conferences that are taking advantage of the internet, with author response to reviewer comments, discussions among the reviewers, etc.

Actually, it takes real effort to make GCC not completely optimize-away this kind of stuff for C:

https://github.com/JuliaLang/julia/issues/1325

There is evidence of a bunch of mild benefits (and also mild risks). I'd currently say it looks like a small net benefit, but with a confidence interval that would include negative territory. In general, it seems reasonable to drink coffee if you like it. It certainly is an addictive drug, though.

http://en.wikipedia.org/wiki/Health_effects_of_coffee

If the current penalties remained in place for smoke-driving there would be no increased incentive to drive while stoned. In general, I think the problems you bring up are perfectly valid, but I see no fundamental reason society should deal with them differently than with alcohol.

Well, the confidence intervals are right there on the charts, and look pretty damn significant to me...

If you are worried about climate change, it seems like our problem in the future might not be running out of fossil fuels-- it will be having entirely too much, making renewables very expensive in comparison.

There is no reason to think that prices elsewhere will get worse. If they could get away with charging higher prices in non-metro areas now, they would do it already. It is likely, however, that non-metro areas will fail to see the price lowering that other areas do.

What, really? I've never had someone in front of me use the space beneath their seat. That's my space! (I'm also 6'2" and only bring a small carry-on because I need that space for legs.)

In my experience, the number of applications submitted is actually somewhat low compared to most recently graduated PhDs these days.

The central limit theorem only applies when random variables are combined linearly. I see no reason to suspect that this is the case with random inputs for testing. (Most functions are nonlinear!)

Wow-- apparently Mergesort can be implemented in-place!?

Interesting. Are you saying that the technology for thorium reactors wouldn't be so closely related to weapons technology?

> how many times do you want to do it?

Well, I've been to the St. Louis City Museum he talks about many times and I can't imagine getting tired of it. (Enjoy sticking your arms out and getting rolled upside down in an antique gigantic wine barrel? Or sliding down frighteningly steep slides intended for transporting shoes? Or crawling upside down in a thunderdome style metal rig on the roof of the building? Or watching a several hundred gallon barrel fill with water until it tips over? Or climbing in metal chutes outside a hundred feet in the air to arrive at a plane suspended in midair by cables? I do!) However, there are few (if any) other places like that.

noelwelsh is responding to donjigweed, not the article.

The issue is that it is nearly impossible to detect "the same". Imagine the sample size required to differentiate between .5 and .500000001. Now, technically speaking, sure-- with any effect (however small) in the infinite data limit a difference would be detected. But this isn't really what these tests are intended for.

This error builds on a simpler, even more common one (at least among students): Suppose you have just one treatment (A), and you find that A has no statistically significant effect. This does not show that A has no effect, or that A's effect is small. (It could just mean your dataset isn't large enough.) The error discussed in the article seems to build off this mistake.

Well, since Steam only stores the salted PW hashes, it doesn't seem like that would allow them to compromise your facebook account.

I don't see why not. Different classifiers have different bias/variance characteristics. If you want to increase variance and decrease bias, than boost your classifier. (This is why boosting is usually applied to simple classifiers.) But whether that will actually help depends on the characteristics of the problem and the classifier used.

I guess bagging is a different story. So far as I know bagging usually decreases variance with no bias penalty, so it is more a trade-off between variance and speed.

In a case like this, perhaps a perceptron?

I think it's more that machine learning builds off of earlier work in statistics. If I remember correctly, that class discusses lots of other things that are used in ML but were invented elsewhere (gradient descent, maximum likelihood, matrix methods, etc.)

Arguably this is all just semantics (nature knows no stats/ML divide), but as a ML person I know this drives statistics folks crazy.

I think we should note that logistic regression has been around forever, and should probably be considered property of "statistics", not "machine learning".

> I use Google Voice so losing my number makes no difference and I am not on the unlimited plan.

Is it possible to use google voice such that when you make outgoing calls, your google voice number appears on the receiver's caller id, rather than your phone's "real" number?

From what I can tell, what you get access to depends on where you bought the kindle more than where you are. Did you buy your kindle in Greece?

I was just thinking of using GVoice for sending/receiving SMS messages. The Kindle 3 has a webkit browser that is supposedly better than previous versions:

http://www.youtube.com/watch?v=OzIlYEp_kDo

Does anyone have much experience with the browser? I've been thinking about buying a kindle 3 3G, with one of the major motivations being the free worldwide web access. I know it is limited and slow, but it would at least allow for email access / chatting / google voice.