Training Multilayer ANN using Kohonen Self Organizing Maps

[zameer]

Hi,

Is there any way to train a multilayer neural network using Kohanen Self organizing maps in AForge framework? I would prefer if there would be a such class like BackPropagationLearning in AForge, to train a Multilayer ANN in Self Organiation mode. Insted, there are classes to train a single layer ANN in AForge like SOMLearning , ElasticNetworkLearning.

Is there any way to do this? Please guide me.

Many Thanks.

zameer

[andrew.kirillov]

Hello,

Have seen such functionality in any other library? Do you know any method of doing this at least in theory? I really doubt. Supervised and unsupervised learning are absolutely different approaches in learning. More of it is not even clear what you want to get as a result.

No, there is nothing like this in the AForge.NET Framework.

[zameer]

Thank you Mr. Andrew,

In hear what we usually do is, to Train a Multilayer ANN, first we train all the layers except the Out put layer in unsupervised mode using SOMLearning rule by repeatedly applying to each layer. And the last (output) layer which is the Grossberg layer is trained in the supervised mode.

The output layer weight change is done using the formula W(new) = W(old) + c ( D - W(old)) Y , where c : is a constent, D : desired output, Y : output of the Kohenen Layer.

In hear we use a combination of supervised and unsupervised training algorithms. I got these information from some of the ebooks available through the internet.

It is actually a Counter Propagation Network

References : Google Books online : Principles of artificial neural networks
By Daniel Graupe , pg 166, Section 8.6 The Combined Counter Propergation Network.

I may have misled you in my previous post, without any proper clarification. Sorry for the inconvenience caused. Hope now you may be able to help me.

Thank you in advance.

zameer.

[andrew.kirillov]

It is actually a Counter Propagation Network

Ohh, this clarifies a lot. I really thought you wanted to mix architecture and learning algorithm ...

No, Counter Propagation Network is not supported by the AForge.NET Framework. But may be you can split the problem into 2 networks (SOM + feed forward network with one layer) and then just pass output from one network to input of another.

W(new) = W(old) + c ( D - W(old)) Y

Does it really work? To me D - W(old) looks meaningless. I would think it should be W(new) = W(old) + c ( D -O ) Y, where O is real output. In this case (if D is desired output) D-O givers network's error, which looks to be more correct. See standard learning algorithm of Perceptron (it is also implemented in the framework).

[sam]

Interesting,

Hope the AForge team will be incorporating the support of Counter propagation networks in their future releases.

Best regards.