Hidden Layer Activations in NN Toolbox

I'm looking for a non-manual way to compute the layer activations of an arbitrary neural network created with the Neural Network Toolbox.

Consider following example for detailed problem description:

[x,t] = simplefit_dataset;       % toy data
net   = fitnet(5);               % initialize network
net.inputs{1}.processFcns = {};  % don't pre-process data
net.outputs{2}.processFcns = {}; % don't post-process data
net   = train(net,x,t);          % train network

The output of the network can be obtained using the sim function or manually:

testX = 0.5;                  % test value
testYnntbx  = sim(net,testX)  % automatic computation of network output
testYmanual = net.LW{2} ...   % manual computation of network output 
              *(tansig(net.IW{1}*testX+net.b{1})) ...
              +net.b{2}

The activations of the neurons in the hidden layer are:

testAmanual = tansig(net.IW{1}*testX+net.b{1})

I'm looking for a way to get the layer activations without manually specifying the equations, analogous to the sim function.

Answer 1

[Greg Heath]

You can create a 2nd net with no hidden layer. Next, make the output layer of the second net to be the same as the hidden layer of the first net.

net2 = fitnet([]);

will create a net with no hidden layer.

If the target matrix is not 5 dimensional, create a 5-fimensional target so that you can configure the correct topology. If t is 1-dimensional use

net2 = configure(net2,x, repmat(t,5,1));

Now you can replace the random initial weights of net2 with the hidden weights of net1.

I have not tried this. Therefore, there may be some details for you to work out.

Another approach might be to reproduce net1, net3 = net1, then remove the outer layer of net3. I can't see how to do this, but it may be possible.

Hope this helps.

Thank you for formally accepting my answer

Greg

[Ahmed]

Replicating the reduced network is a viable workaround. However, every change in the original network potentially requires manual re-definition of the reduced network and is therefore a serious source of errors.