hidden layer values of MLP

Can we get data from the hidden layer of feedforward network? I am using these commands: net = feedforwardnet; [net, tr] = train(net, test, target);

Answer 1

[Paras Gupta]

Hi Asma,

I understand that you want to get the output data from the hidden layers of a feedforward network.

One way to achieve the same is forward propagating through the feedforward network and performing the necessary computations by accessing the transfer function and weights/biases for the different layers in the network. You can refer the following code in MATLAB:

% Define and train the network
[x,t] = simplefit_dataset;
net = feedforwardnet([10, 8, 10]);
net = train(net, x, t);

% Initialize the input
input = x;

% Initialize a cell array to store data from hidden layers
activations = cell(1, length(net.layers));

% Loop through each layer to compute activations
for i = 1:length(net.layers)
    if i == 1
        % For the first layer, use input weights and biases
        weights = net.IW{1,1};
    else
        % For subsequent layers, use layer weights and biases
        weights = net.LW{i, i-1};
    end
    biases = net.b{i};
    
    % Compute the activation using feval
    transferFcn = str2func(net.layers{i}.transferFcn);

    % set input for next layer as output of previous layer
    input = feval(transferFcn, weights * input + biases);
    
    % Store the activation in the defined cell array
    activations{i} = input;
end

The following documentation links can helpful to get more information on the class properties used in the code above:

• network - https://www.mathworks.com/help/deeplearning/ref/network.html
• network object properties - https://www.mathworks.com/help/deeplearning/ug/neural-network-object-properties.html
• network subobject properties - https://www.mathworks.com/help/deeplearning/ug/neural-network-subobject-properties.html