Parallella Community

Hi,

I am currently using a neural network program called emergent from the University of Colorado, running it on Linux. the application uses a back propagation (Bp) model in my case to find correlations between energy consumption and solar generation for buildings and local environmental conditions and time-series data. The current version of emergent is great! it runs on Windows, OSX, and with Linux x86 and x86_64 on Ubuntu and Debian and uses I believe up to 4 cores conncurrently, and possibly can run on other architectures too. Fully open-source. Anyway I do like emergent a lot - but am interested in using Parallella hardware for this fairly compute intensive component of our energy management software platform which is called SolarNetwork, which is also 100% open-source (https://github.com/SolarNetwork), and uses the Pi and other low-cost boards for data collection.

My questions are:

-What would it take for emergent to support Parallela/Epiphany hardware?
-Are there any Parallella developers interested in looking at that project?
-Are there any existing ANN packages/applications that are 100% open-source and run on Parallella now?

Thanks, John

Hi John,

Please see the for previous discussion concerning neural networks.

Regards,

Andrew

Hi John,

I'm writing one at the moment which I'm open sourcing. Early days as yet, I've just got a simple feed forward kernel going (I'm writing in OpenCL).

I'm really keen to get hold of a big data set to do some training and test out performance and scalability.

I've written some blog posts on my ideas about getting the best out of the Parallella which you can read here: and my neural network repo is here: .

I'll have a look at your repo and see what tweaks you are using. What I have written to date has no fancy stuff at all so I've got plenty of scope to add them in.

This is my "after the kids are in bed" project to things are a bit slow, but I'd be happy to work with you on it.

nick

I noticed that you have a raspberry pi logo on your development page. I've got a previous version of my simulator () that runs on the pi.

It will not be very efficient given that it is single threaded and it is limited to one hidden layer but it will give you an idea of what I am heading for at the moment.