Neural net visualisation

I'm able to produce neural nets that work but I want to know what the resulting network looks like. Ideally some sort of graph with edges for each neuron showing their weights. If this does not exist, could someone give me a description of the network object output that will let me find and understand the weights myself?

Answer 1

[Greg Heath]

The simplest solutions use integer weights, two hidden nodes with sign activation functions, purelin output activation functions and polar binary targets [-1 1 1 -1].

Since there are only 4 training equations for estimating (2+1)*2+(2+1)*1 = 9 weights, solutions are far from being unique without other constraints.

You can play around with the equations and determine your own version of "simplest solutions".

These can be easily transformed into solutions for continuous activation functions by replacing sign with tansig. The transformation of purelin to tansig or logsig is also straightforward.

Good Luck.

Greg

Answer 2

[Greg Heath]

You can print the weights. However, it is generally

impossible to understand the weights of a regression

MLP for a medium sized RW problem

On the other hand, the weights of a classification

RBF or EBF are more readily understood.

Moreover, once you get all the interclass distances you can use

multidimensional scaling (MDS) to get a 2-D visualization. The

form of MDS that I have used is called Sammon Mapping. You

will have to Google to find code.

Hope this helps.

Greg

P.S. I have never used a weight graph. However, you may be able to find code on the internet.

Answer 3

[John Dickson]

Thanks, I am looking into your suggestions. It may well be that it is not possible to understand medium-sized nets but it would really help me to understand what's going on if you might explain a simple net. The classic XOR problem can be solved by two and gates and an or: I.e. [A and ~B] or [B and ~A]. Using the code:

input = [0 0; 0 1; 1 0; 1 1]'; %'# each column is an input vector

ouputActual = [0 1 1 0];

net = newpr(input, ouputActual, 3); %# 1 hidden layer with 3 neurons

net.divideFcn = ''; %# use the entire input for training

net = init(net); %# init

[net,tr] = train(net, input, ouputActual); %# train

outputPredicted = sim(net, input); %# predict

The solution net has input weights:

net.iw{1} =

   -1.6875    2.0152
   -2.3719   -1.3529
    1.5600    2.8228

Layer weights:

net.lw{2} =

-2.6606    3.7609    3.7860

And biases:

net.b{1} =

    2.6314
    1.3409
    2.0613

net.b{2} = -1.2402

This is very confusing. Do you know the simplest net that will solve this problem? Is it possible to constrain the optimisation so that the simplest solution is found?