Developing a ANN calibration for ash in meat meal today with Foss Calibrator I realize of the importance of the right selection of inputs in the network. If we take few inputs we can underfit the model and the results could be a bias in the predictions. So prepare a batch mode and don´t be afraid to select a wide range to the limit which in this case is 30. We will get a Heat Map showing the best option for the number of inputs as well for the number of hidden neurons.
First I select the default options and I got a bias validating with an independent set. After I make wider the batch for the input layer and the bias disappear with 30 inputs.
Validation with 22 inputs:
Heat mat recommendation with a wider batch:
Validation with 30 inputs: