Fast and Slow Gradient Approximation for Binary Neural Network Optimization