Neural Network

[censix]

[nickoppen]

Hello,

While we are waiting, I thought I'd use the time to update my old, Windows/MFC neural network library so that it runs on a platform that is a little more generic.

My initial intention was to update the interface classes to use the C++ standard library that runs on the Raspberry Pi given that the Pi is the most similar system I have to the parallella. I found it hard to stop there so I got the whole thing going and added a number of upgrades. It is all very C++-like and as such not at all suitable for the parallella. The next branch will be a more matrix oriented approach that is based on the design outlined in my blog pages ().

Anyway, I've loaded it all up to a git repository (). I use Eclipse and egit and given the amount of fumbling around I've done to get all of the stuff loaded there, I profess to be a complete git ignoramus. "All" actually means "almost all" given that I've not yet managed to upload the most interesting data set, the Shuttle Statlog. I'm still working on that.

If you have a Linux system and can nut your way around git please give it a try and post any feedback on the Feedback... Post it here wiki page in the repository.

Cheers,

nick

[aolofsson]

Nick,
This is really cool. Thanks for staying with us through all these inctedibly frustrating delays. Can't wait to see how epiphany performs on neural networks once hardware hits the street!
Andreas

[nickoppen]

[aolofsson]

nickoppen,
Now that you have the Parallella in your hands, what are your thoughts on running NN on it? Is it what you expected?
Look forward to your thoughts.
Andreas

[Cyberman.ff]

Thanks this is helpful for some of what I was trying to think through.

The training part has been my big "how is that going to happen" issue. It's a highly asymetric process.

[Gerald]

check this out:

[petewarden]

Thanks for posting a link to my work! I actually just registered here to get a sense of how hard it would be to port the framework to the Parallela board. The bulk of the time is spent on a series of massive matrix multiplies, so I was wondering if there was a version of the BLAS GEMM function that had been optimized for the hardware?

[aolofsson]

petewarden,

Nice work with the pi!

Unfortunately there hasn't been an official BLAS library completed yet.

Take a look at the following example showing how it could be scaled out.

https://github.com/adapteva/epiphany-ex ... /matmul-16

What size matrices are you working with?

Andreas

[Gerald]

@petewarden: oh you are here too ...how much speedup do you expect when running it on the Parallella? how much would be the speedup compared to a cuda/gpu version?