Linkage

Posted on by

Apparently I spend half my time reading Crooked Timber.

Žižek gets a lashing for his lazy contrarianism.

A great piece by Michael Bérubé on the Sokal hoax and its aftermath.

Scott Aaronson thinks people should vote to cut funding for quantum computing via YouCut. Why? Because “seeing my own favorite research topics attacked on the floor of the House” would be hilarious (and it would too!).

Marc Lelarge has a new paper up on diffusion and cascade effects in random networks. Fun reading for the break, assuming I can get time.

Some new ways of measuring impact factors.

The top retractions of 2010

Posted on by

This is a bit of a pessimistic list, but here are the top science/scientist retractions in 2010. This reminds me of a pretty interesting New Yorker article I just read on the difficulty in reproducing scientific results. The lingering feeling after reading that article is that we need better statistics than just blindly applying chi-square tests and blah blah blah.

What are your favorite LaTeX macros?

Posted on by

I have quite a few collaborative writing projects going on in parallel now. One side-benefit of collaborating is that you learn neat LaTeX macros that your co-authors have developed that end up saving lots of time (or making the TeX equations more readable). Some people invent a whole set of macros for each paper (so that the macros stand for semantic concepts like “input variable”), but I do that mainly for small stuff, like different epsilons. What I do have are macros for are

  • font types and weights : using \mathrm{}, \mathcal{}, etc. is for the birds
  • functions : KL-divergence, mutual information, conditional mutual information etc. I get in trouble sometimes because I use the I(X \wedge Y) instead of I(X; Y) for mutual information, but we can change that in the macro!
  • norms, inner products : these are just functions anyway
  • epsilons : this helps keep different epsilons clearer, actually, and makes debugging proofs a little simpler.

I somehow never get around to making macros like \cX for \mathcal{X}, but maybe that would make my life easier. The nice thing about macros for functions is you can put in auto-sizing delimiters in the macro, saving some \left and \right‘s. What are your favorite things to make into macros?

Linkage

Posted on by

The Voice of the T announcements (h/t Erin). Features a special guest appearance by not-me.

Videos from the 2010 Kailath Lecture and Colloquium have been posted (h/t Pulkit Grover).

The Supreme Court issued a 4-4 per curiam decision affirming (pdf) the Ninth Circuit Court’s decision in Costco vs. Omega. Basically Omega sells its watches cheaper abroad, and Costco was re-importing them from a reseller to give the discount to US customers. This was a copyright violation, says Omega, and the Ninth Circuit reverse the lower court decision in favor of Costco. Why does this matter? As I mentioned earlier, a decision for Costco would let resellers sell cheap textbooks in the US. When I was in Delhi I picked up a copy of Feller Vol. 1 (the Vol. 2 was a little beat up so I decided to get it another time) for Rs. 400 (under 10 bucks), a savings of over $120 from the price on Amazon. Admittedly, it’s softcover and the paper is not quite as nice, but the way publishers gouge people on technical books is astonishing.

Wiener on control versus learning

Posted on by

I’ve seen this quote excerpted in parts before, but not the whole thing:

I repeat, feedback is a method for controlling a system by reinserting into it the results of its past performance. If these results are merely used as numerical data for the criticism of the system and its regulation, we have the simple feedback of the control engineers. If, however, the information which proceeds backward from the performance is able to change the general method and pattern of performance, we have a process which may well be called learning.
– Norbert Wiener, The Human Use of Human Beings

It is a strange distinction Wiener is trying to make here. First, Wiener tries to make “numerical data” a simple special case, and equates control as the manipulation of numerical data. However, he doesn’t contrast numbers with something else (presumably non-numerical) which can “change the general method and pattern.” Taking it from the other direction, he implies that mere control engineering cannot accomplish “learning.” That is, from numerical data and “criticism of the system” we cannot change how the system works. By Wiener’s lights, pretty much all of the work in mathematical control and machine learning would be classified as control.

I am, of course, missing the context in which Wiener was writing. But I’m not sure what I’m missing. For example, at the time a “control engineer” may have been more of a regulator, so in the first case Wiener may be referring to putting a human in the loop. In the book he makes a distinction between data and algorithms (the “taping”) which has been fuzzed up by computer science. If this distinction leads to drawing a line between control and learning, then is there a distinction between control and learning?

more binomial MADness

Posted on by

I posted earlier about the mean absolute deviation (MAD) of a binomial variable S_n with parameters (n,p). Here’s a little follow-up with plots. This is a plot of \mathbb{E}|S_n - np| versus p for different values of n.

The first is for n = 10. Looks beautifully scalloped, no? As we’d expect, the MAD is symmetric about p = 1/2 and monotonically increasing for the first half of the unit interval. Unfortunately, it’s clearly not concave (although it is piecewise concave), which means I have to do a bit more algebra later on.

When $n = 100$ the scallops turn into a finely serrated dome.

By the time you get to $n = 1000$ the thing might as well be concave for all that your eye can tell. But you would be deceived. Like a shark’s skin, the tiny denticles can abrade your proof, damaging it beyond repair.

Why do I care about this? If you take n samples from a Bernoulli variable with parameter p, then the empirical distribution (unnormalized) is (n - S_n, S_n). So \frac{1}{n} \mathbb{E}|S_n - np| is the expected total variational distance between the empirical distribution and its mean. More generally, the expected total variational distance for finite-alphabet distributions is a sum of MAD terms.

Linkage

Posted on by

Some interesting stuff has passed my way while being in India (and one or two things from before). Might as well post them before I forget, no?

Slavoj Žižek may be a curmudgeonly Marxist, but the animation helps soften it, I think. I don’t think I fully agree with him, but there’s stuff in there to chew on.

The Purdue anonymization project won a big NSF award.

Tips for tasks related to graduating (h/t Bobak).

Some interesting news about the future of the textbook market. It’s doubly interesting since I am in Pune, a treasure-trove of cheaper editions of technical books.

Apparently I sometimes wear a lab coat.