ASH vector quantization achieves state-of-the-art ANN speed and recall

Intel Labs researchers Mariano Tepper and Theodore Willke have introduced ASH (Asymmetric Scalar Hashing), a new vector quantization method that promises to redefine the speed-accuracy tradeoff in approximate nearest neighbor (ANN) search. Published on arXiv, ASH breaks from the conventional wisdom that additive quantizers like product quantization (PQ) are optimal. Instead, it revisits scalar quantization—compressing each vector dimension independently—but with a critical twist: a learned dimensionality reduction step that projects database vectors onto a lower-dimensional orthonormal subspace. By reducing dimensions while increasing the bitrate per dimension, ASH preserves more information per stored component. The key insight is asymmetric: queries remain in their original high-dimensional form, while only database vectors are compressed. This avoids information loss on the query side and enables fast distance computations via SIMD (single instruction, multiple data) operations.

Extensive experiments across multiple datasets show ASH outperforming both additive quantizers (e.g., PQ, OPQ) and data-agnostic scalar quantizers (e.g., SCA, LSQ) at every compression ratio. The method achieves higher recall at the same memory footprint and up to 2x faster search speeds due to SIMD-friendly integer arithmetic. Importantly, ASH has short learning and encoding times, making it practical for real-world deployment where index rebuilding is frequent. The paper positions ASH as the new gold standard for high-fidelity vector quantization in large-scale retrieval systems, particularly relevant for vector databases and semantic search applications that serve millions of vectors.