Feedforwardnet trainning: Inputs partially disconnect with hidden layer after I train my network even if I set correctly the field inputConnect

Hi, I face the problem when trainning my feedforwardnet nn_h with multiple inputs X_h as a cell format and single target Y_h. I do not know why inputs partially disconnect with hidden layer after I train my network, as shown by the resulted figure

my simple code is as follows.

clear;
inputnum=3;
batch_size=10;
X_h=cell(inputnum,1);
for i=1:inputnum
    X_h{i,1}=zeros(1,batch_size);
    for j=1:batch_size
        X_h{i,1}(:,j)=i;
    end
end
Y_h=cell(1,1);
Y_h{1,1}=zeros(1,batch_size);
for j=1:batch_size
    Y_h{1,1}(:,j)=4;
end
nn_h=feedforwardnet(10);
nn_h.numInputs = inputnum;
for k=1:inputnum
    nn_h.inputConnect(1,k)=1; 
    % even I set connections for each input with hidden layer, it seems partially removed after 'train' as above figure shows
end
nn_h=train(nn_h,X_h,Y_h);

Answer 1

[Nicky]

Hi, I guess I found the answer myself.

As we can see from the above code, the first input X_h{1,1} is actually a constant for all 10 data points, which seems the bias constant b is enough to enter into the trainning process. It makes sense!

On the other hand, if I make the elements across the whole batch of X_h{1,1} different, then we can find the back of connection between the first input X_h{1,1} and the hidden layer.

As for other inputs X_h{2,1} and X_h{3,1} which are also constants, it seems that MATLAB only removes the connection of the first constant input, due to my observation.