Architecture
Vector Quantization (VQ)
2017ActiveUpdated: 22 September 2026Published
Key
innovation
Obtaining a discrete, learned representation of continuous data via a codebook.
Category
Architecture
Abstraction level
Building block
Operation level
Architecture blockLayer
Use cases
VQ-VAE / image and audio generationMultimodal tokenizationLatent-space language models
How it works
An encoder maps the input to a continuous vector that is replaced by the nearest codebook entry; a decoder reconstructs the input from the code.
Problem solved
Autoregressive models operate on discrete tokens, yet much data (image, audio) is continuous.