Deep Probabilistic Kernels for Sample-Efficient Learning