Different results by backpropagation algorithm using different MatLab versions (2008 and 2010)

Hi,

I usually use the Neural Network toolbox of the Matlab version of 2008a. However in this moment I’m working with a different computer with 2010b version and my net doesn’t give me exactly the same results. I’ve notest that some functions are obsolete and I’m afraid that there are other reasons for this difference that I’m not aware (or other functions called inside other functions…). For example, 2008 version still uses “calcgrad” that disappear in 2010 version, which uses “staticderiv”…

The functions more relevant are:

net = newff(P1,T1,H,{},'trainscg');

net.layers{:}.initFcn='initwb';

net.inputWeights{:,:}.initFcn='rands';

net.layerWeights{:,:}.initFcn='rands';

net.biases{:}.initFcn='rands';

net.performFcn='mse';

[net,tr,Y1,E1] = train(net,P1,T1);

I updated newff as recommended.

Somebody has any idea why are my results different?

Thank you.

Answer 1

[Mark Hudson Beale]

First, I assume you are setting the random seed before running this code to try and reproduce exact results? Otherwise, every run will be different.

If that doesn't work, then exact results may be difficult to replicate. Bug fixes, algorithm enhancements, or even the order of calculations that change between versions, can make a difference in the final trained network.

You could try checking the weights after setting the random seed to consistent value and creating the network in both version, before training, to see if the change is happening in initialization or training.

[Ana]

Thank you for your answer.

First, yes, I'm setting the random seed before running :)

I'll do what you suggest to realize where is the change.

Thank you again.

Ana

[Harish Balaga]

what is random seed?

how to set it