fix for "hidden_layer_sizes" not being assigned when "@hidden_layers" is set

[illtellyoulater]

fixes #15

[irfansharif]

thank you for taking the time to investigate this 👍

[illtellyoulater]

Glad I could help, this is a really nice project :).
If this could be of interest to you, it seems that the original JS project got back to life.
I am not sure how though. The author has not been responding for long, but then - looking at harthur/brain#72 - it seems the community started to maintain it again.

[irfansharif]

ah. hadn't seen that.
it was an interesting original project but I'm not a very big fan of javascript/node, the lack of a strong standard library puts me off to it. going to closely follow the development of the tensorflow client in ruby, lots of potential there!

[illtellyoulater]

I fully agree! 👍 I just thought I would tell you in case you were interested in implementing new features that JS Brain might implement.

On a side note, I am seeing some weird results after training a Cerebrum NN.
Specifically some of the outputs have a value bigger than 1.
My use case is that I have 54 inputs, 2 hidden layers of 8 neurons each, and 54 outputs.
I trained the NN by providing an array of multiple instances of 54 inputs and 54 outputs.

I know this is not the perfect place for discussing it, but just wanted to know if you had seen this kind of behavior yourself before (outputs > 1), so that I may better decide what to do about it.

Thank you!

[irfansharif]

interesting, if possible email me the datasets and I can possibly investigate further, though word of note I never intended Cerebrum for serious use more along the lines of working pseudo-code as an instructional tool.

[robertleeplummerjr]

Just to note about #16 (comment), we have started maintaining it, and if there is anything you see that needs investigated, please open an issue. We have started to work on recurrent neural nets (BrainJS/brain.js#16), which I'm really excited about. Very cool to see this lib!