[FreeSynapse]

I don't know if this has been discussed before or not, but under which programming language(s) will HTM-based systems be programmable and/or hackable? I don't know if limiting it to C++ alone would be a very good idea.

[FreeSynapse]

One important point to note is that generally speaking, the superset of AI programmers has enough work dealing with figuring out AI, statistics, etc. that they usually may not be experts in C++. Due to this, other languages such as MATLAB, Python, etc. gain popularity with them. I do understand however that the HTM core probably does need to be coded in C++.

[heckj]

Hasn't Numenta stated that they're making headway in coding their HTM's in python to enable both rapid development and a core that can be written in a speedy language (C++) where needed?

[FreeSynapse]

[heckj]

I'd guess python primarily, but the most recent available code that I've seen was all in MATLAB... pending development I guess.

[FreeSynapse]

Why is an official response so hard to get? I was under the impression that this forum was monitored by Numenta.

[Numenta-Phil]

In order to do all of the early research that we've done with our HTM theories, we used MatLab, as it has excellent mathematical functionality, and a powerful scripting language, allowing us to more easily experiment with our learning algorithms and such.

As we are moving towards more of a platform, MatLab isn't being used as much, as we want our technology to be available to everyone, and the price of MatLab eliminates a large percentage of our audience (there are other technical reasons, of course, but this was one of the largest stumbling blocks).

As we develop this platform, we want our core engineers to use the same technologies that our customers will use, and thus all of our algorithms are being ported to Python, as stated in our latest newsletter:
http://www.numenta.com/newsletter/index.php?module=newsView&id=1

Python enables us to easily make quick changes to our learning algorithms and experiment, without having to do a build process in between. However, once these algorithms are nailed down and polished, we absolutely expect to put the core algorithms in C++ which should be significantly faster.

Once we release this product, developers will be able to use our existing algorithms, or implement their own in C++ or Python, just like we do.

I hope this answers the question.
_________________
Phillip B. Shoemaker
Director, Developer Services
Numenta, Inc.

[FreeSynapse]

Phil,

It may be worthwhile to try using Psyco, which is a Python extension module that can massively speed up the execution of Python code. I don't pretend to know Psyco, much less understand how it works, but it may worth a try, both with full binding, and alternatively with selective functions binding. The link to Psyco is:

http://psyco.sourceforge.net/

Note that as I understand it runs only on Intel based machines. Nonetheless, it may still be worth experimenting with, at least to see how far Python can be pushed here.

[heckj]

I just returned from the OReilly Open Source convention where I saw Pysco in action (very impressive!), but the piece that really made my jaw drop was the use of pyrex. Details on pyrex can be found at http://www.cosc.canterbury.ac.nz/greg.ewing/python/Pyrex/ and a couple of pretty good overview articles at http://www-128.ibm.com/developerworks/linux/library/l-cppyrex.html and http://ldots.org/pyrex-guide/

You can basically write the pyrex python straight up and take no significant impact, but if you're willing to restrict the methods down to some specific signatures, then Pyrex does the work of generating x86 machine code and you can see some really amazing speed increases.

[Numenta-Phil]

Thanks for the headsup. I'll definitely get the information to the team and we'll have a look at it.

Thanks again! Keep the tips coming.
_________________
Phillip B. Shoemaker
Director, Developer Services
Numenta, Inc.

[FreeSynapse]

By the way, it would be nice to see native x64 compatibility also, if and as possible.

[lkelemen]

Hi,

Have you considered to implement a HTM in java or even better
to use an existing freeware java neural net framework like Joone?
See at http://www.jooneworld.com/

br
Lajos

[andy.goryachev]

Hi,

I wish Numenta had chosen Java over C++ and Python. Not only it gives people better chance at playing with HTM API, but also:

- offers faster design cycle (coding + testing + debugging)*Number of iterations
- provides comparable performance with C++ when compiled (GCJ)
- allows for better development tools (Eclipse)

It is unfortunate. We all have to wait six times longer for this revolutionary technology to appear.

[FreeSynapse]

[andy.goryachev]