I have a multilayer Neural Network with 1 layer of hidden - z units, input layer - x units and output layer - y units. I have a question, Can I divide the NN into 2 new ones as:
- NN1: Input layer - x units/hidden layer z1 units/output layer y1 units.
- NN2: Input layer - x units/hidden layer z2 units/output layer y2 units.
with: z1,z2 \< z and y1+y2=y
Can I divide it? Please, help me.
Thank You very much.
Not if you want to keep the exact same behavior. Each output node could potentially depend on all the hidden nodes, so while you could split the output layer into two networks (duplicating the input and hidden layers), you can't split the hidden layer without (potentially) changing the behavior.
That said, if each output node depends only weakly on some of the hidden layer nodes, it might be possible to split it while changing the behavior only slightly.
My output units are large and they are independent groups. Can I split it?
What do you mean by "large" and "independent groups"?
While I have never used neural networks, I am very interested in AI and have read a bit about them. Could you possibly plug what you want to divide into another NN as the input? Just a possibility that I thought might work, sorry if it is of no use, like I said I have no real knowledge on using them just theorectical knowledge, I guess as what you could call it.
I think I had a mistake. Thank you for help.