Hallo,
I am trying to understand the way the bpxnc works. I read documentation and right now I am trying to get closer the the PRTools code.
Could You please send me the picture (jpg) how networks structure looks like when I type bpxnc([],[10 10], nr_of_epoche)? I need just the draft picture in which I could find the information where are weights and how the input and output of such network looks like for the ‘soft’ version. 

[10 10] means number of units in each hidden layer

Please help me with this problem

If you apply bpxnc([],[10 10], nr_of_epoche) to a dataset with 5 features and 2 classes it creates
a network with 5 inputs, 2 hidden layers (each with 10 units) and 2 output units. There are the
following weights:

input to first hidden layer:  5 x 10 = 50 weights
first to second hidden layer: 10 x 10 = 100 weights
second hidden layer to output:10 x 2 = 20 weights

Every unit has a bias weight as well that is not connected to an input, but that is always fed by ‘1’.
See the documentation of Matlab’s neural network toolbox (TRAIN) which is called by prtools.

Bob Duin

Thank You for Your explanation, it helps me a lot.
I have one more question. Let’s imagine that I am using perlc as a fuser. As a input to the fuser I put class numbers which I got from simple classifier denoted as Y1, Y2, Y3. There are three final classes. How the connections looks in such situation?
As I know single perceptron can give as only two classes, but in my case a have 3 ... Please try to explain me this.

Thank You for Your help

Martin

I am confused by the terminolgy: “fuser” with “class numbers” as input. I don’t understand what is meant
by this. Moreover, perlc is a one-layer neural network, identical to the standard unit used in multi-layer networks.

Anyway, maybe the following is helpful to undertstand how PRTools works: any 2-class discriminant can be
used to solve multi-class problems by schemes like one-against-all (’single’ in mclassc) or pairwise
(’multi’ in mclassc). These schemes make use of the confidences generated by the classifiers and the
max rule as combiner. perlc uses the one-against-all scheme.

Bob Duin