Quote: Original post by Phaz
According to my professor, the current best performing supervised machine learning algorithm/model on that dataset was a support vector machine.

I support that. SVMs where the best classifier in EVERY comparative study I've read.

I think the reason NN is the best solution is because it eliminates real-time calculations. It's suppose to just spit out an output depending on the data it is trained on.

Quote: Original post by Anonymous Poster
I think the reason NN is the best solution is because it eliminates real-time calculations. It's suppose to just spit out an output depending on the data it is trained on.

And do so magically without any kind of calculation? Come on. For NNs, the computation time will depend mainly on the size and topology of your network.

For a SVM, it will depend mainly on the size of the input vector (unless you use a very complex kernel), same for the method suggested by the OP.

There's a new type of pattern-recognition theory being devised. I've tried a demonstration program and I think it works pretty darn well. It can be downloaded here: http://www.phillylac.org/prediction/

The theory was formulated by Jeff Hawkins, creator of the Palm Pilot, and author of the book: "On Intelligence".

New 32x32 training-characters can be easily added to the program, and I would suggest deleting some of the other simple characters that come with it in order to bias it toward the more complex characters that you draw. Don't delete too many, however, or you'll run out of memory.

New patterns require atleast 4 or 5 prototypes to get started, followed by some testing and additional prototypes, though it usually doesn't require more than 10, even for complex shapes, like helicopters or people. :)

Numenta has just come out with their HTM toolkit: www.numenta.com

Looks pretty interesting, I must say.

Quote: Original post by Anonymous Poster
Numenta has just come out with their HTM toolkit: www.numenta.com

Looks pretty interesting, I must say.

Nice marketing, AP.
Until proven otherwise, that stuff is utter and complete bullshit. The only theory they released was full of hole and non-sense.

First of all, I'm not personally involved in the project.

Secondly, don't worry, conventional image-recognition techniques work just fine on simple patterns like text, so I wouldn't expect HTMs to be very competitive in that area. This is mainly for identifying more complex patterns that other software have not been able to before.

Btw, surprisingly enough, this works on almost everything BUT Windows. Looks like Virtualization will be required for the time being.