@ChrisDiehl Note that since SVMs are single value output, you'll need one SVM for each hidden layer node. Trained with SGD.