I am trying to use the trainSoftmaxLayer for multi-classification application. I have few queried related with that. 1. Can I report Cross-Validation accuracy using the classifier. I have done 10-fold cross validation and I am getting good results. But what is troubling me is that at every round of cross-validation the initial weights and biases of the networks are initialized. I do not know whether I am right or wrong.
2. How to get optimal value of the regularization parameter. The default value of the regularization parameter is 0 and the function does not provide facility to change the regularization parameter.
3. How the initial weights are initialized. I did not found any information related with this in the function help of the trainSoftmaxLayer.
Please help me out.