Converged neural network states

Hi -

I am wondering why I don’t arrive at the same trained network (net1f and net3f) even though I believe I have started from the same initial network state.

clear all, pack [x,t] = simplefit_dataset;

%% 1st trial net1i = feedforwardnet( 1); net1i= configure( net1i, x, t) ; IW1i= net1i.IW ; LW1i= net1i.LW ; b1i= net1i.b ; net1f = trainscg( net1i, x, t); IW1f= net1f.IW ; LW1f= net1f.LW ; b1f= net1f.b ;

%% 3rd trial with controlled initialization net3i = feedforwardnet( 1); net3i= configure( net3i, x, t) ; net3i.IW= IW1i ; net3i.LW= LW1i ; net3i.b= b1i ; net3f = trainscg( net3i, x, t); IW3f= net3f.IW ; LW3f= net3f.LW ; b3f= net3f.b ;

I appreciate your help.

Thanks. Siva

Answer 1

[Greg Heath]

You have to explicitly reset the RNG state to the same initial value. To illustrate this. Check the RNG state before each training.

Hope this helps.

Greg.

Answer 2

[Siva]

Thanks Greg!

Siva