MATLAB: How to compute sensitivity analysis in neural network model

neural networksensitivity

How to compute sensitivity analysis in neural network model? I would like to find the level of importance of each input.
Input value is 12×1505 double. Target value is 1×1505 double.
Here is my code:
x = Input';
t = Target';
trainFcn = 'trainlm'; 
hiddenLayerSize = 3;
net = feedforwardnet(hiddenLayerSize,trainFcn);
net.input.processFcns = {'removeconstantrows','mapminmax'};
net.output.processFcns = {'removeconstantrows','mapminmax'};
    net.divideFcn = 'divideind';
    net.divideParam.trainInd = 1:903;
    net.divideParam.valInd = 904:1204;
    net.divideParam.testInd = 1205:1505.
net.performFcn = 'mse'; 
[net,tr] = train(net,x,t);
y = net(x);
e = gsubtract(t,y);
performance = mse(net,t,y)
trainTargets = t .* tr.trainMask{1};
valTargets = t  .* tr.valMask{1};
testTargets = t  .* tr.testMask{1};
trainPerformance = perform(net,trainTargets,y)
valPerformance = perform(net,valTargets,y)
testPerformance = perform(net,testTargets,y)
view(net)
Thank you very much
Janthorn

Best Answer

1. Use MAPSTD or ZSCORE to standardize the data BEFORE training.
2. Design 10 or more successful nets with the smallest number of hidden nodes as possible.
3. For each input: Add Gaussian noise to only that input. Tabulate and plot MSE vs noise standard deviation
There are also many, many, approaches for ranking inputs via backward and forward searches. In backward search, replacing an input with zeros (it's mean value) is equivalent to removing it.
Hope this helps.
Thank you for formally accepting my answer
Greg
Related Question