Swapping Neuron Groups in Backpropagation

[jjhan]

Awesome program, I'm trying to figure out how to make it do what I need.

I'm trying to run some deep learning backprop and I want to be able to have 3 different neuron groups for inputs where I swap alternatives of each group.  For example, let's say I want to train the network on equal sized neuron groups A,B,C,...,Z where I choose any three of those neuron groups for the input.

I've managed to get the connections outside the backprop, but that seems to prevent the backprop from training.

It seems like there should be some way to run a backprop where 3 neuron groups are connected to the first hidden layer.  The math doesn't change at all, only an easier organization.  

All help is greatly appreciated.

[jyoshimi]

Hi Jihan,

Thanks for the note.  Simbrain 3.03 does not support that. It has a pretty brittle, minimal backprop implementation.  You have to stay inside the confines of a single backprop subnetwork. What you describe should indeed be straightforward to implement, and we are working on being able to support things like that for Simbrain 3.1.   But 3.1 is going pretty slowly, since everyone involved has a million other commitments :)  But still it will eventually be possible!

- Jeff