Building Optimal Neural Architectures using Interpretable Knowledge