Neural Network Question

[Nacre]

I have two neural network topologies. The first has two input nodes that are fully connected to 14 hidden layer nodes. The hidden layer is then split. 7 of those nodes go to 2 output nodes. The other 7 go to another 2 output nodes.

 

The second topology is very similar, but instead of the split hidden layer receiving input from the 2 input nodes, each section of the hidden layer receives input from separate input nodes (2 each). Effectively, then, the first network is split into 2 networks.

 

The thing is, what if these two networks in the second topology receive exactly the same inputs to their input nodes? Does it then produce the exact same results as the first topology? I am asking because the second topology is easier to implement (just two instances of a simple network, with identical input).

 

My understanding of the back-propogation algorithm is that it would not matter which topology I used, as long as the inputs were identical, but I wanted to verify this (as the network is not working in the way it is supposed to, and I am not sure if it is the network itself or some other factor).

[keelanz]

I have two neural network topologies. The first has two input nodes that are fully connected to 14 hidden layer nodes. The hidden layer is then split. 7 of those nodes go to 2 output nodes. The other 7 go to another 2 output nodes.

 

The second topology is very similar, but instead of the split hidden layer receiving input from the 2 input nodes, each section of the hidden layer receives input from separate input nodes (2 each). Effectively, then, the first network is split into 2 networks.

 

The thing is, what if these two networks in the second topology receive exactly the same inputs to their input nodes? Does it then produce the exact same results as the first topology? I am asking because the second topology is easier to implement (just two instances of a simple network, with identical input).

 

My understanding of the back-propogation algorithm is that it would not matter which topology I used, as long as the inputs were identical, but I wanted to verify this (as the network is not working in the way it is supposed to, and I am not sure if it is the network itself or some other factor).

 

hmm, MCDST, A+, N+ and even uni have (as of yet) never mentioned neural topology, just the old ones like star, net, mesh & token ring. do you have any visuals of your topology? might give a better understanding of whats going on, i would presume a neural network would create itself? i.e. find relative nodes to itself and only talk to them i cant see any algorithm that could define it though, otherwise AI would have existed a good while ago.