For example, I consider an NN with n number of input units. What should the construction of NN be changed if the number of input units increases to 10*n? I think the number of hidden units and hidden layer should be increased. And, we should have a higher number of samples in the dataset (e.g., using data augmentation).
I also think dimensionality reduction could be used (like word embedding).
What do you think about this problem? Thank you.