It isn't everything you need know in 30 minutes, but it's a concrete coverage of lots of topics in machine learning in under 150 pages. Here's why I'm recomending this paper:
* The algoritm is easy to understand.
* It can handle classification, regression, semi-supervised learning, manifold learning, and density estimation. The paper gives an introduction to each of these topics as well as a unified framework to implement each algorithm.
* It can handle categorical data and missing data [2]
* It gives as good results as other state of the art algorithms.
* The paper is well-written and easy to understand for someone without a deep background in machine learning.
[1] It's mostly a review paper. Using random forests for density estimation is new.
[2] This review paper doesn't cover categorical data or missing data.
Nice talk. The example of google translate is not a good one though. Say you translate from language A to language B with 99% accuracy, and vice versa, which would be pretty awesome, you'd still have a substantial quality decay after only a few back and forth translations (0.99^x where x is the number of translation steps).
That's not a very realistic model of communication. If A communicates to B with 99% accuracy in the text, it's likely that B will 100% understand what's going. He will reply back to A with a 100% accurate message that's 99% accurate after translation, and so forth.
I agree. But my impression was that she took the fact that google translate rapidly decays into gibberish as an indication that it's not doing a good enough job. I don't think you can argue that exactly because google translate does not have the interpretation capability.
Don't read too much into that example -- I chose it as a humorous metaphor, not a mathematical argument, and I messed up the delivery in the talk, anyway.
I'm really interested in doing more machine learning work (my current projects, as interesting as they are, dont really require it).
I've done a few weirdo projects with NLTK, tho, and its great fun. By stream hacking do you mean offloading learning sets (active or initial) and that heavy overhead into the "cloud", or am I misunderstanding the terminology?
In most data analysis work, we assume that the data resides in some database and that you have the luxury of iterating over that data as many times as you like to get to a final result.
The challenge with stream analysis is that you are dealing with a continuous stream of data where you can see each element of the stream only one time and must still be able to cluster/classify/analyze it. There are still few algorithms and tools designed explicitly for that purpose.
Not that I am doing Machine Translation (MT), but saying accuracy is a bit vague. The whole notion of what is lost in a translation using MT is to the best of my knowledge not fully captured with any well-established measure.
Fair warning, I haven't had time to have a look at the video (short break at work). I'll do it once I get home.
For more advanced audiences, here's a great resource I discovered recently: http://videolectures.net/mlss09uk_cambridge/ -- 60 hours of lectures from the giants of machine learning delivered at a summer school held at Cambridge in 2009.
It isn't everything you need know in 30 minutes, but it's a concrete coverage of lots of topics in machine learning in under 150 pages. Here's why I'm recomending this paper:
* The algoritm is easy to understand.
* It can handle classification, regression, semi-supervised learning, manifold learning, and density estimation. The paper gives an introduction to each of these topics as well as a unified framework to implement each algorithm.
* It can handle categorical data and missing data [2]
* It gives as good results as other state of the art algorithms.
* The paper is well-written and easy to understand for someone without a deep background in machine learning.
[1] It's mostly a review paper. Using random forests for density estimation is new.
[2] This review paper doesn't cover categorical data or missing data.