jwp

It also resonates especially when you're reading from hard experience. A year ago I would have quickly read & enjoyed this, then moved on. Now too many items give me pause as I recall some mistake I've made. It is very good.

Can you tell apart founders who sponge up advice vs those who are doomed to learn on their own? I suspect the answer is conversational resourcefulness (http://www.paulgraham.com/word.html) but wonder if you have more to say.

This is exactly the same handshake deal the founders of Fairchild Semiconductor did with each other.

What he said! I'd like to add that a key to squeezing more out of NathanRice's post is the phrase "conjugate prior." Another totally natural thing would be to use a Gaussian prior & likelihood, then update the posterior as ratings arrive. This would take advantage of the ordinality of ratings as NR suggests. Bishop's Machine Learning book goes into this sort of stuff in more depth.

Hey, AOLserver is open source. At the time it was the only web server with a built-in glue language (Tcl) and pooled database connections. It got me excited about web programming. Plus, the sheer amount of code already written was fantastic.

I never understood why people were put off by his favoritism to those tools. ACS was all about page flow and user experience, and in that department, got it right more often than not.

Kevin Garnett plays exactly the same way that you describe Jack Lambert. He goes only one speed: full blast. Whenever I need motiviation, I find myself asking myself, wwkgd?

Creating the list of URLs and prioritizing them is the hardest thing about building a crawler! That is, a good, web-scale one. A replacement for wget might be sort of fun, but the real way to make a fast crawler is to be choosy about which pages get updated frequently, which are likely to contain good content (by computing a pagerank-like stat on the fly), etc.

It is far from my area of expertise, but the Wikipedia page about this looks very useful. It cites a bunch of wicked smart people. http://en.wikipedia.org/wiki/Web_crawler

If you just want to suck down a bunch of pages, then there's nothing wrong with wget.

Sure, there's plenty of fluff in the article, but it isn't because the author (http://www.math.wisc.edu/~ellenber/) is clueless about science. The article is probably at the right level of technical depth. Seems ambitious to even touch SVD in Wired.

Suffix array.

I remember seeing this video in a very sobering HBO documentary: http://www.hbo.com/docs/programs/hackingdemocracy/

Nespresso is great. At $0.50 per pod, it's pricey, but for such painless cleanup and consistently good shots, worth the price. Hands down the best teas I've had come from Upton Tea (uptontea.com).

My intuition says it's not possible to get 1 without destroying the ability to get 2, or vice versa. But I'm no expert. The solution that leaps to my mind is to use a locality-sensitive hash function. It would make search stochastic, which means reversing the encoding would be easier. Seems like any hash-based solution will involve a tradeoff between 1 and 2.

What did you have in mind?

That's not true. You can submit your results and then refuse to release your algorithm, disqualifying you from the competition.

"Upon qualifying, as described above, the Participant is required to submit within one (1) week for judging a description of their algorithm along with all source code. The Participant warrants that the source code is either fully or substantially developed and functions or will function as represented by the description. Failure to deliver both the description and source code within one (1) week will disqualify that entry and additional qualifying entries will be considered."

What's your score on the Netflix Prize quiz?

We prefer to call it a family.

The non-friend vs anti-friend distinction didn't even occur to me, but it's clearly an important part of the experiment. I like all the discussion of data. Fun problem.

I did not realize squaring an adj matrix tells you what it does. Thanks for edjumacating me.

Did you go past f2hops? Seems like 3 would be reasonable and predictive. Since 1/2 of your tree is so small, and there were 12k nodes in the tree, that suggests to me a pretty easy task. Do you agree? It would be interesting to see if PCA or LDA pick the same features as the decision trees did. Just a click away in Weka, after all.

(An aside, and neat hack: Buddy of mine just walked in and saw the document on my screen. He saw the decision tree and said, "I remember those. In grad school I printed out decision trees as C if/else statements. Part of running my decision tree was a call to gcc.")

Yeah, yeah, so the engineering part wasn't that great. But were the results good? Which algorithms did you use? What were your features? Seems like an interesting experiment, especially with all the gratuitous "friending" people do. Reminds me of a related paper: http://www.hpl.hp.com/research/idl/papers/facebook/facebook.pdf

Good cites. I second them. Mackay and Jaynes are excellent but can get out of hand quickly. Duda, Hart and Stork is another good hardcore option. When I need the book "for dummies," which is pretty often, the Weka book and Mitchell's Machine Learning are helpful.

A very important question... Celestial Seasonings Fast Lane tea.

The MSFT guy is right. Full open speech recognition is very hard and a long way away. For some simpler tasks, it's fine.

On a related note, I was surprised to discover that "say" works over remote logins. Try "ssh some-mac say hello world". Good for Halloween.

When you file yourself (pro se), the PTO is gentler and kinder than they are if a lawyer files for you. If you have the time and energy, file now to get the early filing date and call a lawyer later.

This begs the question, how do you define a bot? From my understanding of the talk, they replay one person's session and don't use composites. But, point taken.

It depends on what your definition of "bot" is. Are you talking about when the ESP game records a person's session, then plays those moves back to other players (and other recorded sessions)? If enough about the recorded play is logged, why should the person playing interactively care?

I always thought that's one reason his games are so clever. You get better data out this way, too. Seems perfectly legit to me.

From what I can tell a lot of this CF stuff starts with a machine learning algorithm and data about likes/dislikes, making filtering a classification or search task. Is there something that differentiates it from the usual machine learning challenges? Like dealing with users interactively, perhaps?

Hmm, perhaps we should talk. Email me at [email protected] if you're interested.

The first link you cited is spot on. The authors are from Univ of Cambridge, and work on HTK <http://htk.eng.cam.ac.uk/>.

That paper is 10 years old. As I'm sure you can imagine, there have been improvements in the field since then. To be completely honest, I don't stay on top of search applied to speech, but the keyword you want is "Spoken Document Retrieval" (SDR). Ciprian Chelba and TJ Hazen do cool stuff in this area; they are giving a tutorial at ICASSP this year SDR.

An aside. Both of these approaches use the fact that when you process speech, you essentially form a graph of words (or phonemes). Paths through the graph represent possible transcriptions. So, since graph is a denser, richer thing to search than the transcript, and we've got graph algorithms sitting around, there are neat tricks you can do to build a search engine index for speech...

I've recently been reading some interesting work that uses locality-sensitive hashing to search audio. The Google speech people are presenting a lot of it at ICASSP this year. See this post for more, and chase the links in their papers for even more: <http://googleresearch.blogspot.com/2007/02/hear-here-sample-of-audio-processing.html>

I work on speech recognition. What Blinkx is doing isn't novel. (Sorry, Blinkx.) Google has top speech researchers working on search for speech and video. Same for MSFT. Remember the Kai-Fu Lee thing? The guy who built CMU Sphinx, an open-source HMM-based speech recognizer, in the late 80s? He and many other solid speech people are at Google to work on searching audio/video.

And there are other companies in this space, but they tend to center around US gov customers. Virage is one. It's owned by the Autonomy group, where, according to the article, the founder of this company used to work.

There's also Podzinger, a subsidary of BBN, which is another company that gets a lot of gov business. Podzinger runs BBN's speech recognition system on podcasts and videos, and pipes the output to a search engine: <http://podzinger.com/>.

I could go on... And if people are interested, I'd be happy to post links to some relevant papers and tools.

To my mind 2 interesting things are going on here. 1) The company appears to be thriving by applying 20 year-old stuff from the lab to a new problem, in apparently no special way. (And that's not a bad thing!) 2) They got an article in the NYT business section to talk about Hidden Markov Models. Although maybe that's not so surprising, since hedge funds have recently started speaking out about using machine learning.

IANAL either (although I do date one), and I too have heard of these things. I think they are called E-2 visas. They require you to raise or invest around $100k in a US business. The lawyer who told me about the E-2 was quick to warn that the applications get scrutinized hard. It sounded political, vaguely sketchy, expensive, and complicated.

I'm glad you linked to the talk. It's excellent.

I'm curious, why do you think that visual games are more likely to be successful? His question answering games seemed useful and engaging, too.

There must be great startup ideas lurking behind this idea of games and Human Computation. (Luis alludes to as much, with his tongue-in-cheek suggestion that the ESP game double as a dating service.)