Data normalization for training ANN model and getting negative output values

I am using ANN for training a model, do I need to normalize my output "target data" also, or just normalizing the input variables are sufficient? also, all my target values are positive but after training the network I get some negative values which it does not make any sence in my case, be noted I have used logsigmoid function as my activation function.

Regards

Answer 1

[Rushil]

Hi

When working with ANNs in MATLAB, especially using the “logsig” (log-sigmoid) activation function in the output layer, it is important to normalize both input and output (target) data. The “logsig” function outputs values strictly in the range (0, 1), so the target data should be scaled to this interval before training to ensure the network can learn the correct mapping.

Below is an example workflow illustrating this process:

% assuming you have inputData and targetData
[inputNorm, inputSettings] = mapminmax(inputData, 0, 1);
[targetNorm, targetSettings] = mapminmax(targetData, 0, 1);

net = feedforwardnet(hiddenLayerSize); % assuming a general hidden layer network
net.layers{end}.transferFcn = 'logsig';
[net, trainInfo] = train(net, inputNorm, targetNorm);

% save the network if needed
testInputNorm = mapminmax('apply', testInputData, inputSettings);
outputNorm = net(testInputNorm);

output = mapminmax('reverse', outputNorm, targetSettings); % reverse the normalization 

In this workflow, both the inputs and targets are scaled to the [0, 1] range using “mapminmax” before training. After training, predictions are made using normalized test data, and the outputs are then transformed back to the original scale using the same normalization parameters. This approach ensures the network's outputs are meaningful and within the expected range. To know more about “mapminmax”, refer to to the following MathWorks Documentation link below:

https://mathworks.com/help/deeplearning/ref/mapminmax.html

Hope it helps