I’m just about to wrap up a weeks vacation before it’s back on the horse again – as I leave for the nucl.ai conference Sunday morning. If you’re going, maybe I’ll see you there? If not, you should go check out the live stream.
Basis
Anyway, so in my vacation I built a little project when I could find a bit of time here and there. It was inspired by work on a client project earlier in the year, where we looked at wiring up a neural network for analysis of a noisy sensor.
The problem at the time was that although I could find a few C# NN libraries here and there, they were either not really a good fit or had a not-so-open open-source license like GPL (good thing I don’t have comments enabled here, eh?).
Anyway, that felt a bit silly, so I figured that I’d at least get the ball rolling on a somewhat sane, MIT licensed, alternative. Either that or just have some fun with it until someone would point me at the obvious alternative solution, which I had overlooked in my earlier search.
Project
In its initial shape, the library just lets you construct a layered network, randomize its weigths, train it with a straight forward implementation of back-propagation with adjustable learning rate and momentum, and of-course run it.
There is no serialisation interface yet, I’ll probably want to do a simplification pass over it, and I have a few edge case tests I’d like to run. However all of that is noted on the project TODO and shouldn’t amount to all that much work – in case you fancy taking a stab at it.
I did set up a little Unity project for testing though (the library itself does not depend on anything in Unity). It has a few debug views, a test scenario of the XOR case, and a curve visualisation of training error. Look, I uploaded several pretty pictures:
Access
Feel like giving it a go? As mentioned, the license is MIT, so you can really do whatever with it. However should you want to contribute, pull requests are most welcome. I don’t know how much time I can spare on this next to contracts and Behave, but at least it was a fun thing to work-on-while-not-working.
Whatever the case, this is where you can find the repositories: