Catalysis of neural activation functions: Adaptive feed-forward training for big data applications