What is vector quantization in deep unsupervised architectures?

Vector quantization maps continuous embeddings into discrete codebooks for compression and learning.

Vector Quantization (VQ) replaces continuous latent representations with discrete embeddings from a learned codebook. Models like VQ-VAE use this to compress data and learn discrete latent spaces. This improves generative modeling and reduces memory usage. It is widely used in image, audio, and speech representation learning systems.