Use PCA,KPCA on your dataset with variance 40% - binary datasets,60-70% - multiclass datasets. PCA reduces dimensionality and u obtain minimal numbers clusters KPC1 to KPCn that represent your minimal numbers of hidden layers. Fed those KPCn clusters to K-Mean clustering and u obtain corresponding numbers of neurons for each cluster. Then test trial your NN topology with performance and accuracy. U can use other clustering methods for estimating numbers of neurons for each cluster using modern Elbow method and then compare NN topology results in trial manor to see which performs how for given output against known values. To know which NN topology perform better use different values for regularization, best NN topology doesn't change or doesn't change much for different regularization values on accuracy. For known values against the predicted values u can trial check accuracy in excel which ones performs better. Look online for further info on how to do it. This is the by far best NN topology estimating method to start on. Further u can use Grid search or similar methods when pointed to right direction. Hope that helps all the other users who struggle with hyperparameter estimation.