textminer

The pedant in me squirms at seeing a quadratic-time solution to something so linear. Use a heap or a scan, sir!

Best Kafka-short-as-an-HN-comment I've ever read. Won't be sleeping tonight.

I've wondered this, too!

It helps that elementary functions are in some sense dense in the set of all other functions. Results like Stone-Weierstrass show that on a compact interval, any function can be arbitrarily-well approximated by a polynomial.

Other things that might interest you-- how one might know if a function even has an antiderivative that can be expressed as an elementary function: http://en.wikipedia.org/wiki/Liouville%27s_theorem_(differen...

You just made me realize something. I attributed my own college grades slipping a bit my senior year to burnout, but it was also Fall '08, and I had just got an iPhone 3G, my first smartphone...

I agree with this. I find the convention most stifling-- my instinct in the afternoon is to go work somewhere else, but quickly guilt* starts to pop up and tell me I'm doing something wrong, even if it's just to do work more ably.

(* - this guilt is a weird thing I can't really describe. It also lessens as I advance further in my career and am more sure of myself.)

I can remember times being a student, learning a lot, but losing focus of how amazing it was I had dedicated time solely to study. As someone working full-time now, who only gets to learn new fascinating math in his precious free time, I implore you to fully embrace the awesome opportunity in front of you. Enjoy the lecture, and milk whatever you can from this man's teaching.

I've noticed people taking note of and commenting on Kara Swisher's take-charge, arrogant, brusque nature far more often than with many male tech journalists. Unexamined bias, or an expectation that female correspondents be softer and nicer?

About computationally as complex, no? The Manhattan distance would just be the sum of the absolute difference in the two coordinates.

The Manhattan and Euclidean distance on finite spaces both generalize to "l_p norms" (1 and 2, respectively) on an n-dimensional Euclidean space, where ||x||_p = (|x_0|^p + |x_1|^p + ... + |x_n|^p)^(1/p). You can even naturally define it in the limit where p goes to infinity!

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

We walk blocks, not as the bird flies.

Not useful for Kevin here, but for those with fat, frozen Python dictionaries that're useful to some vital process (e.g., feature lookup), the Python wrapper for the MARISA trie is wonderful. Tend to save 80-90% of memory whenever I capture a dict with it, and it works with memory mapping.

Quite efficient even for something silly, like casting a long_int -> short_int mapping by first converting the keys to strings. Hurrah, prefix trees!

The rewards game for first employees is perplexing. You have an order of magnitude less incentive than the founders who birthed the thing. Arguably, that's taken care of: you're on a real salary while they're making peanuts. You're also capable of moving to a new job tomorrow, while this is the founders' life.

But when an acquihire with retention bonus is almost equivalent in net payout to you as a successful acquisition or IPO might be, it is difficult to be super-invested instead of merely daydreaming and working toward starting your own company. At best you can hope to be flush from a quick and healthy payout, which might pay off student loans or capitalize the first steps of your own risky venture.

Is the game of Silicon Valley just to structurally encourage everyone to eventually devote their time to fully creating their own pet technologies and products? Who works for anyone else except for fools and transients?

(I'm currently a fool or transient.)

Maddening how so many classic textbooks have new versions every few years that add a few nominal features ("Calculus in the Real World!" "Online Tutorials That're Worse Than What Someone Else Made on YouTube!") with the real goal of shuffling around the exercises so a student can't do his or her homework without having the correct edition. Completely squashes resale markets.

Like you, that also floors me. Moved here in mid-2011, fresh out of grad school. Seems to be cheap when there's not a dotcom bubble, yet then we wouldn't be here. Ain't that a kick in the head.

In Oakland, I'd some nice parts of Lake Merrit, Chinatown and Temescal. Rockridge is nice, if thoroughly yuppified. Richmond's Marina Bay and El Cerrito are both thoroughly nice and BART-adjacent, if a bit too suburban and removed.

Silicon Valley is stucco-corporate ugly. San Francisco is charming, yet now grossly-expensive. The two are linked by decrepit transport systems that are being supplanted only by Google/Yahoo/Facebook's own shuttles (services that do nothing for low-wage Valley workers or entrepreneurs trying to begin new endeavors).

East Bay is charming and inexpensive. There are residents and young people who discuss more than option-pool restructuring and PHP/Rails. As long as I can continue to work in San Francisco, and as long as rents stay as high as they are in that city, I'll choose to live here.

R's not going to teach you about pre-measures, Carathéodory's theorem, or Radon-Nikodym. Break out the whiteboards!

(Then bust out R again when you want to play with stochastic processes and Brownian motion.)

"Pedigree Matters" as an untruth: I wish this were the case. Too many entrepreneurs I know are only entrepreneurs because they have the support system of comfortably wealthy families and the professional network gained at elite schools and fellowships.

Would love to hear more about entrepreneurs who live on the edge and try to start a business while also needing to make rent and student loan payments.

Seemed unfortunately successful in its advertising. Have a brother in first grade who is a champion at using my iPhone and iPad to play games and look up dinosaur pictures and videos. Suggested to my stepmom he get an iPad mini or Nexus 7 for Christmas, but she had already bought him a MEEP (at his request). What a shame. Kids can (and should) learn real computing devices. Giving them locked-down, "child-friendly" devices is a disservice to both them and the society they'll one day be adults in.

( - Pretty sure that learning Logo, QBasic, DOS, and Telnet/Winsock during grade school in 90's is the only reason I'm an employed programmer now.)

Incredibly interested in your work here. For small-dimensional problems (or problems with features that can be engineered to be small-dimensional), ensemble methods through random forests and bagging and the like are incredibly useful.

But for high-dimensional text problems that're pure classification, I tend to rely simply on 1NN classifiers (against a single centroid of training data of a target category, of which there tend to be many). I've spent a lot of time with NMF, for its potential as an incredibly interesting data-exploration tool ("There's a pronoun cluster! There's a Spanish cluster! There's a 404 Error axis!") or low-dimension projection step. I've even spent a good amount of time on implementing the algorithm in a number of memory-efficient ways.

Could you expand a bit on how you used NMF for these problems in practice (similar to how a sparse autoencoder captures reduced-dimensional features en route to supervised learning), or how others used ensemble methods?

Love it. It's that way with range, too, right? Does one end up casting things to lists more often if so much is naturally an iterator?

Love defaultdict. It and dict/set/list compressions are a big part of what makes Python so fast to write in.

Great practice for 2.7 that's probably quashed in 3.0. For large dicts, no need to create a giant set en route when iterating over keys, values, or both. Use "for k in d.iterkeys()", "for v in d.itervals", "for k,v in d.iteritems."

While I'm at it-- if you're ever finding yourself using a huge amount of awfully rigid objects from a single class, use __slots__ to allocate needed variables! Python will otherwise define the object's namespace in a dict (called __dict__) which allocates a whole kilobyte per object. Bad news if you have several hundred thousand... Guessing this is why Guido loves namedtuples so much for basic attributed storage.

My first downvote! Thanks for the constructive feedback, stranger.

Thanks for the advice. Honestly, being that I work in startups, am a year and a half out of grad school, don't have to well-to-do parents, and support my partner, I think you've helped me realize it's definitely not so smart to be bufferless.

Honestly, part of this mania began when I looked at my statements for just one provider (Chase) and realized the minimum payments they were asking for (~$300) covered ONLY the monthly interest. That sort of freaked me out, and I got spooked at the prospect of allowing that bank to rent-seek on my back for the next few decades, while I lack any sort of consumer rights to even ask a court for bankruptcy protection.

I agree here. The solution for me on not quite getting concepts in calculus and engineering was to delve into greater abstraction. When you get linear algebra, you understand the inverse function theorem in analysis. When you understand Hilbert spaces, the Fourier series makes a lot more sense.

A lot of the simple analogies and explanations I can use to talk about determinants or Fourier transforms now I can only conjure because of the complex study taking years to soak through.

Would you change that suggestion if the fresh grad had near 100k in loan debt, some of which was private and high-interest? The benefit of 10k sitting in savings when it could pay off a savage Chase loan seem less great.

(To be explicit, that person is me, and I tend to throw all buffer money right at student loans before they grow like a hydra, faster than I can kill them.)

Learning about the basis of logistic regression in Ng's Stanford class was eye-opening. Also like that he then motivated generalized linear models, and why they're nice (e.g., parameters are linear with input data; the maximum-likelihood hypothesis is the expectation of the sufficient statistic) and he explains why we see the logistic function in so many damn places (it's the response function whenever y|x is distributed as an exponential family).

(Pardon me/blame autocorrect: Breiman.)

You will find no understanding in the grown trees-- by definition, they're each seeing a different angle of the data, and each node a different subset of available features.

You can, however, calculate importance scores for the features used. Brennan's original paper gives a good algorithm for doing this (in short: for each tree, permuting the data along some feature for an out of bag sample and seeing how much worse it does.)

Love his tutorials. Motivates PCA by walking trough a creepy eigenface decomposition, and solves a political prediction problem with a ID3 implementation of a decision tree. I actually used the latter initial code scrap for writing an entire damn random forest implementation!

Really looking forward to his future discussion of computational topology and persistent homologies.

But District 1's representative is chair of a committee you need to sweet talk in two months, and is maybe a tie-breaking vote for that other thing you want.