Gradual Surrogate Gradient Learning in Deep Spiking Neural Networks