Can I use the neural network toolbox to train and test customized feed-forward networks?

Hi,

I am using the neural network toolbox for classification tasks. In the toolbox, we can customize the # of input nodes (input feature dimension), the # of output nodes (# of targeting classes) and the # of hidden layers. The default network is a feed-forward network that each of the node in a hidden layer is connected to all the nodes in the hidden layer above it and below it. I would like to define my own network with customized connections. It is still a feed-forward network, however, I'd like to remove some of the connections. For example, a network like this, where the # number of nodes in each hidden layer may vary and the adjacent hidden layers may not be fully connected with each other:

Can I use the neural network toolbox or functions in the neural network toolbox to train and test a model like this? I have my own data set that I believe would work well with my customized model.

Thank you very much,

Qing

[Greg Heath]

It can be done. However, it will probably take longer to figure it out than you would like.

Good Luck

Greg

[Qing]

Hi Greg,

Thanks for replying to my question. Why do you say it can be done? Do you know if there is any reference I can refer to regarding this? Does customizing the network involve defining the "net" object? Is there any specific function I can look into?

Thanks a lot!

Qing

[Qing]

Hi Greg,

Thank you very much! I followed your suggestion and it works. :)

Cheers,

Qing

Answer 1

[Greg Heath]

http://www.mathworks.com/help/nnet/ug/create-and-train-custom-neural-network-architectures.html

Good Luck

Greg