regarding Neural network toolbox

Dear Sir I want to know that,whether we will implement different transfer function for different neurons on the same layer in matlab Neural Network toolbox

Answer 1

[Greg Heath]

If you do this you will have to write your own code. An equivalent, but easier alternative might be to use skip layer connections with each layer containing a different type of transfer function.

However, why in the world would you want to do this? A standard MATLAB function like FITNET or PATTERNNET with a single hidden layer with tansig transfer functions is a universal approximator.

Hope this helps.

Greg

[Kshitij Tripathi]

ok Thanks Greg Sir,but Just help ,give me the hint,how to write my own code in Matlab to do this.

[Walter Roberson]

You can see the features you should support by looking at http://www.mathworks.com/help/nnet/ref/purelin.html

Your replacement transfer function should have the same functionality.

Answer 2

[Walter Roberson]

You can if you want to but it would be unusual.

You can supply your own transfer function. That function can do its calculation any way you want, including being position dependent.

I would remind you that if this is not the input layer then you probably do not really know which neurons are participating in which calculation. Neural Network weights are calculated as "whatever works to give the desired result" not by partitioning the network into portions that calculate specific features.