Hi,
Before implementation of MLP there was a reference at root server
to other variant of Neural Nets - NueNet.(e.home.cern.ch/e/ernen/www/NN/)
Is there an easy way to migrate from NueNet to MLP. I mean I have
*.dat files with neurons weights from NueNet.
Can I translate them easily to weights of MLP?
Do they use the same activation function ?(Iguess - Sigmoid, but I am not sure).
I would prefer not to train my Nural Net again, but rather take old dat
files and create new Neural Net, based on MLP.
Is that possible?
This will be difficult, I guess, but this is not excluded.
In both case (it looks like, I don’t know NeuNet) this are sigmoid functions (neurons) related by linear relations (synapses).
The format of the weight file for TMLP is not designed to be read by humans (well, there are comments), but it is like this: #input normalization
x1 y1
x2 y2
…
xn yn #output normalization
x1 y1
x2 y2
…
xn’ yn’ #neurons weights
w1
…
wm #synapses weights
w1
…
wm’
for n input neurons, n’ output neurons, m neurons, m’ synapses
Compared to other networks, the terminology comes from the following. If you have a sigmoid like this:
sigmoid(a1x1+a2x2+a3x3+c),
a1, a2 and a3 are synapse weights and c is the neuron weight.
It’s a bit difficult to build by hand such a file, because of the machine-oriented ordering. What you can do is to load a weight file, and then export the NN (e.g. as a C++ function). It will be more easy to understand the exported function and to compare with what you did in the weight file.
I let you judge what you will gain w.r.t. retraining the NN.