Differentiable Compositional Kernel Learning for Gaussian Processes