Incorporating Prior Knowledge into Neural Networks through an Implicit Composite Kernel