Product quantization
Product quantization: A technique that compresses vectors by splitting each embedding into subvectors and quantizing them separately. Each subvector is represented by the code of a centroid in a codebook. It reduces memory and bandwidth but introduces approximation errors that must be evaluated through recall.
In italiano: Product QuantizationLegge I · Embedding, feature e spazi vettorialiEach subvector is represented by the code of a centroid in a codebook. It reduces memory and bandwidth but introduces approximation errors that must be evaluated through recall.
Examples
- IVF-PQ
- Codebooks for subspaces
- Compression of document indexes