Supervised Vector Quantized Variational Autoencoder for Learning Interpretable Global Representations