yaroslavvb

Also as SWE, but for a specific position, as opposed to general pool new hire that gets assigned after starting

I've joined Google 4 months ago, and so far it is the most enjoyable job I've had, including grad school. I've worked at 3 start-ups before, and after working at Google, I think promise of future riches is the only reason to prefer a start-up. If you are proactive, you get full freedom to choose problems you work on, a ton of computational resources and many very smart and supportive people around you. I had an opportunity to join Google earlier as a "general pool" SWE, and I turned it down because I thought I might get stuck maintaining some boring piece of code. But if I knew what I know now, I would've taken it and would probably be in the same position as now -- SWE's have ability to reallocate themselves, and if you find a better team, your old manager isn't going to fight to keep you

Make these stats and I'll link them from the page :)

To add to that, the second floor area of the building where most of the data-entry people sit is open to public, meaning that your Googler friend could take you on a tour. Security/legal are always on look-out for people stealing trade secrets. I would guess "videotaping around google property", "refuses to give the tape", "not full-time Googler", triggered some red flags

I use Delicious to bookmark useful pages I come across, then import those bookmarks into Google Bookmarks. This makes my bookmarks hits show up using regular Google Search. This is better than Google bookmarklets because it's easy to synchronize your bookmarks across multiple google accounts

I think it's right because I got through using Mathematica's MIP solver http://bit.ly/fr7YXu

BTW, tree decomposition is a generalization of dynamic programming. In a general structured problem you may not have a nice table recursion like in edit-distance example, and tree decomposition is what you get when you abstract away from tables. When problem structure comes from the real world, like with probabilistic networks, computers are much better at finding a good recursion than humans. Here are some pictures to illustrate tree decomposition http://yaroslavvb.blogspot.com/2011/02/generalized-distribut...

Well, that looks easy...so why are people saying that writing memoization in Haskell is PhD-thesis level task?

Except here you could say "both are correct" or "neither are correct", whereas in the original question those also cause a paradox

That's because people are spoiled :P The service is free, meanwhile Facebook is a for-profit corporation so its responsibilities to the investors come first

J-1 visas with minimum residency requirements have something like "Two year rule does apply" written in the passport. People who come sponsored by program like Muskie have that limitation

Maybe...also the fact that I have 400 "friends" says something. You can't possibly know that many people well enough to trust them. And the fact that they accepted my friend requests means they didn't care about keeping their information secret in a first place. Maybe it's different for an average user (120 connections apparently), although that still seems too high to share sensitive information.

I agree. And personally, I find a big gap between what concerns "the media" and the people around me. Among my 400 facebook friends, I've gotten 2 or 3 status updates concerned with new privacy rules. Meanwhile on hacker news I see a dozen of articles in a day about it

Good point...it's not clear that any of those posts were really meant to be private

I sent it using gmail's web interface. My IP address shows up in "X-Originating-IP" part of header

Someone commented on that page that leaking user's IP address is a common spam-prevention practice. I just checked gmail and Yahoo Mail, and they both include my IP address in the header of outgoing messages

For instance, Yahoo Mail puts my IP address in a line like this

Received: from [xxx.xxx.xxx.xxx]

Note that the original paper doesn't say anything about crashing. Their point is that E.Coli regulatory network has a different topology than Linux call graph, in particular, a much lower overlap between modules and little reuse of low-level workhorses. They postulate that higher level eukaryotic networks would have more reuse and be more similar to Linux call-graph. http://www.pnas.org/content/early/2010/04/28/0914771107.full...

IMHO, this fits nicely under Eric Schmidt's philosophy that "If you have something that you don't want anyone to know, maybe you shouldn't be doing it in the first place." After all, it's the morally questionable act of profiting from someone's mistake that's the ultimate cause of his troubles.

The Russian version confuses two engines because of hyphenation (perhaps due to line-wrapping in the original article)

Sudden acceleration complaints by year/make http://www.npr.org/templates/story/story.php?storyId=1242358...

For model year 2009, Volkswagen had 11.5 complaints per 100k vehicles, whereas Toyota had 7.5 and BMW 5.8

A big problem is that there's often no objective way to pick the "right" statistical test. This gives experimenter the freedom to choose the statistical procedure that favors positive conclusion. Here's a classical example when the statistician has the freedom in choosing between binomial and negative binomial test

Two experimenters contrast treatments A and B. They both have A preferred to B in first 5 patients, and B preferred in the 6th. First experimenter planned to run 6 experiments and count the number of successes, so they get P-value 0.11 for the hypothesis that A is better. The second experimenter planned to run comparisons until B is preferred, up to 6, and got P-value of 0.03. (see Appendix A of http://www.annals.org/content/130/12/995.full.pdf+html)

A realistic example of this issue coming up http://www.jstor.org/pss/2336980

Another example is choosing between one-tailed and two-tailed t-test. When you ask for the probability of effect being as extreme as observed x under null hypothesis, should you ask for probability of effect>x or |effect|>|x|?

Eliezer Yudkowsky goes into some discussion on this http://lesswrong.com/lw/1gc/frequentist_statistics_are_frequ...

The most illustrative example of subjectivity of hypothesis testing is probably the issue of testing strings for randomness. There are many tests for testing whether a particular string of bits is generated by a Bernoulli process, with not one having a legitimate claim to being "the right one"

One way to remove the bias is three-way triple-blind testing, ie to measure effects of existing, new, and placebo treatments, have statistician analyze datasets while blinded to their true labels.

There's some confusion between "compressed sensing" and "L1-minimization" in the comments and the article.

L1 minimization here refers to trying to represent the image as a linear combination of basis functions a1 f1+a2 f2+ ... by minimizing objective function fit(a1,a2,...)+sum_i |a_i| where "fit" represents how closely the reconstruction fits the image (for instance, squared distance). The second term in the sum is the L1 norm of the parameter vector, and it has a special property of being non-differentiable at 0, and often causing individual components a_i to become exactly 0 during minimization, hence resulting in "sparse" solution vector.

Compressed sensing refers to the idea that when the image is sparse in the chosen basis {fi}, it's sufficient to take a small number of random measurements of the image. We can then compute "fit" approximately which gives us a different objective function that we minimize. The result will be close to what we'd get if we used the original (full) "fit" function

I found scanned pdf versions of half of the books in the list after about 5 minutes of searching

Slides from related presentation of his

http://www.niehs.nih.gov/news/events/pastmtg/2007/cehr/docs/...

Also, even when the papers are not posted on author's webpages, they are often happy to send you a copy over email.

Funny enough, I have the same issue with my old student ID -- it works to get access to those databases long after I graduated

I was in your boat several months ago when I had to learn image processing from scratch to implement a requested feature

"Digital Image Processing Using MATLAB" ended up the most useful, along with some basic Matlab tutorials

here's what I ended up with http://yaroslavvb.blogspot.com/2009/08/robust-ocr-in-video.h...

A curious recent variations is allowing the number of hidden units to be infinite. It was observed that increasing number of hidden units in a two level perceptron reduces the number of local minima encountered during training. Allowing the number of hidden units to be unbounded makes the optimization problem convex http://books.nips.cc/papers/files/nips18/NIPS2005_0583.pdf

He uses Bayesian methods in some cases, and non-Bayesian method in others. For instance on page 1412 he describes a problem for which Bayesian methods are "not appropriate." http://omega.albany.edu:8008/ETJ-PS/cc14g.ps

This is a bit of an anathema to purist Bayesians like Radford Neal who say that Jaynes Maximum Entropy method is not consistent with Bayesian methods and that it "doesn't make any sense"

http://groups.google.com/group/sci.stat.consult/msg/2cf57ceb...

It's good, with a heavy focus on frequentist statistics and maximum entropy approach

Pearl provides a formalism for expressing causality, however it's questionable (to me) whether that formalism corresponds to what we humans consider "causality". Take 3 variables as an example, A,B,C. We observe some instances of A,B,C triples, and try to fit a probability distribution P(A,B,C) to data. If A,B,C are binary, there are 7 parameters, we can fit them directly to training data to get perfect fit, however, when we run it on separate set (validation data), it'll probably not perform very well. To improve fit on new data, you need to reduce number of parameters, so you could make some simplifying assumptions and write P(A,B,C) as P(A)P(B|A)P(C|B). In this form, there are 5 parameters to fit, and the resulting model could perform better with validation data. Alternatively you could fit P(A)P(B)P(C|A,B), 6 parameters, or P(C)(B|C)P(A|B). Select the one that performs the best on your validation data, and is hence the "best" model.

Now comes the cuasality connection -- if the best model is P(A)P(B)P(C|A,B) you read it as A->C<-B. If best model is P(A)P(B|A)P(C|B), you read it as A->B->C. (ie, A causes B, B causes C)

I think causality connection is questionable because the model that corresponds to A->B->C will have the same fit as model for A<-B<-C or A<-B->C. In fact, you could take any causal network without loops and "unshielded colliders" (connections of the form A->B<-C), pick any node as a root, and re-order the arrows to face away from the root to get a model with a different semantic causal structure, but the same mathematical structure, meaning it'll give identical fit to data.

What would be really interesting is if someone deduced causality using Pearl's approach, then verified it using a direct experiment