Paul Dix Explains Nothing

MatchupCamp in NYC next Wednesday

I just want to help spread the word that For Your Imagination is hosting a nextNY organized MatchupCamp next Wednesday here in NYC. From the site: MatchupCamp "is all about startup networking, creating a place for ideas and talent to meet."

As of right now there are 66 people signed up to attend with many of them listing a desire to find designers or developers. So if you're in that category you'll probably get a lot of attention.

Here are the time and location details:

Wednesday, November 28 @ 7PM

For Your Imagination
22 West 27th Street
6th Floor
New York, NY 10001

I plan on attending to meet as many people as possible and also keep an eye out for a good designer and possibly a co-founder.

Technorati Tags: nextNY, MatchupCamp, nyc, startups

Using Support Vector Machines for Predicting Structured Output

Today I attended a guest lecture by Thorsten Joachims on the topic of using SVMs for predicting multivariate outputs. Basically, predicting more complex output than a simple yes or no binary classification. The examples he gave were predicting things like parse trees for natural language processing, sequences, sets, and multi-class classification.

He also showed methods to train these in linear time with the number of examples in the training set. He showed results of a run on a set with 250,000 features and 36,000 training examples. It looks like really interesting work. You can have a look at Thorston's papers and research or you can jump to SVM struct, which is the realization of this research in code.

You can read about the results of A Support Vector Method for Multivariate Performance Measures and Training Linear SVMs in Linear Time in his papers.

It was cool to get to meet Thorston after using SVM light for an academic project last spring. I also worked with one of his Ph.D. students, Thomas Finley, last summer at Google.

Hopefully I'll have time to hack with this over winter break. Tom wrote the Python library for SVM struct and I think it deserves a Ruby library as well.

Technorati Tags: svm, machinelearning, classification

Aho Talks About the Quintessential Questions of CS

Today I attended a talk by Al Aho (of AWK and Dragon Book fame) about the quintessential questions (problems, challenges) of computer science.

In short, here are The Questions:

1. How do we determine the difficulty of a problem? P, NP, Exponential, etc. Does P = NP?
2. How do we model the behavior of complex systems that we would like to simulate? His main example was simulating a quantum computer so that we can develop compilers, programming languages, and other tools for them.
3. How do we build a scalable trustworthy information infrastructure? Protection from malware, privacy, security, etc.
4. Is there a scientific basis for making reliable software?
5. Can we construct computer systems that have human-like attributes such as emotion or intelligence?

Those were the five questions that Al had compiled and presented to Bill Gates in a 2005 round table with the CS faculty of Columbia. He also asked Bill what one question he would add to that list. Bill's addition was this:

How do we extend Moore's Law?

Dr. Aho showed some graphs of the progression of the number of transistors on a chip. He then switched to a graph that showed power consumption and heat output which roughly matched the steep positive curve of transistor counts. He then highlighted that if we continue on this path of expansion, chips will reach a temperature as hot as the surface of the sun in not too much time. So it definitely seems like a hard and important question.

I'll have some of my own musings on questions four and five in the next week or so, but I just wanted to get these up as food for thought for now.

RubyConf Wrapup

I'm a little late posting something about RubyConf, but here goes. It was a great event. Big thanks to the guys at RubyCentral for putting together an excellent gathering of talks and Ruby hackers. I ended up in the hallway track as much as in the actual talks. My quick favorites are:

nyc.rb's own Francis Hwang on Conversations vs. Laws. This was a more theoretical talk with very little code. Francis' presentation style made it entertaining and thought provoking. I thought it was interesting that he spoke about Ruby's lack of specs as part of what makes it dynamic. Mainly because the alternative implementation efforts like Rubinius, JRuby, and IronRuby are racing to definite a set of specs to nail the language down.

Evan Phoenix' talk on Rubinius. Again, one of the things that made this a favorite of mine was Evan's loose and easy presentation style. Rubinius looks very interesting and after talking a little bit to Evan, Brian, and Wilson, I'm really excited to see how it develops. I'm specially interested in how they're building parallelism into everything.

Nathan Sobo's talk on TreeTop was very good. He ran through a live coding session to show off how to write a grammar for building a combo lexer, parser, and semantic analyzer (basically everything you need to write your own fully featured DSL or compiler). It's an interesting project. The only drawback is that I have no where to apply it at the moment.

I'm also really excited to start using AP4r (here's a decent article on AP4r) on my project. I had given it a brief look over the summer, but decided it wasn't what I needed. Well after attending the talk at RubyConf and wrestling with parallel code, I've decided that I'll give it a shot. It isn't a perfect fit, but will most likely be the quickest path to getting my task done.

There were also a couple of things from the hallway track that were very interesting. I talked to Bob Amon about parsing feeds and the fate of FeedTools. He then pointed me to rFeedParser, which looks perfect for my needs. Speed was one of my primary problems with FeedTools and it looks like rFeedParser is much quicker.

I also heard some rumbling about Starling, a yet to be released messaging (like AP4r) library from the guys at Twitter. I've been waiting for this thing to get released since I first heard about it over the summer. I guess the original plan was to release it at the conference, but actual work prevented them from finishing things up. It doesn't look like it's been released yet, but I'll definitely be keeping my eye out.