From Black-Box to White-Box: Control-Theoretic Neural Network Interpretability