# Accelerating Large-Scale Inference with Anisotropic Vector Quantization

Source: [Accelerating Large-Scale Inference with Anisotropic Vector Quantization](<https://www.sota2.com/research/paper/68fa68de1e209131cea1ed83>)

Collection index: [llms.txt](<https://www.sota2.com/research/llms.txt>)

Data updated: 2025-10-25T22:47:52.778Z

Authors: Ruiqi Guo, Philip Sun, Erik Lindgren, Quan Geng, David Simcha, Felix Chern, Sanjiv Kumar

Published: 2019-08-27T18:27:17.000Z

## Abstract

Quantization based techniques are the current state-of-the-art for scaling maximum inner product search to massive databases. Traditional approaches to quantization aim to minimize the reconstruction error of the database points. Based on the observation that for a given query, the database points that have the largest inner products are more relevant, we develop a family of anisotropic quantization loss functions. Under natural statistical assumptions, we show that quantization with these loss functions leads to a new variant of vector quantization that more greatly penalizes the parallel component of a datapoint's residual relative to its orthogonal component. The proposed approach achieves state-of-the-art results on the public benchmarks available at \\url{ann-benchmarks.com}.

## Recorded benchmark results

A research result applies to its stated dataset, split, metric and evaluation configuration. Compare like-for-like settings, preserve the metric's unit and direction, and distinguish paper-reported results from independent reproduction. The best result recorded here is not a claim of exhaustive, current global coverage.

- [Vector Search on MSMARCO 8.83M text documents (test)](<https://www.sota2.com/research/sota/vector-search-on-msmarco-8-83m-text-documents-test.md>): Vector Search; MSMARCO 8.83M text documents (test); Traditional Recall: 97.6.

- [Approximate Nearest Neighbor Search on ANNS Empirical Evaluation Suite](<https://www.sota2.com/research/sota/approximate-nearest-neighbor-search-on-anns-empirical-evaluation-suite.md>): Approximate Nearest Neighbor Search; ANNS Empirical Evaluation Suite; QPS vs Recall Tier: -1.

- [Approximate Nearest Neighbor Search on FASHION-MNIST](<https://www.sota2.com/research/sota/approximate-nearest-neighbor-search-on-fashion-mnist.md>): Approximate Nearest Neighbor Search; Fashion MNIST; Throughput (QPS): 2082000.

- [kNN Throughput on Glove 100](<https://www.sota2.com/research/sota/knn-throughput-on-glove-100.md>): kNN; Glove 100; Throughput (QPS): 339000.

- [Approximate Nearest Neighbor Search on DEEP 10M](<https://www.sota2.com/research/sota/approximate-nearest-neighbor-search-on-deep-10m.md>): Approximate Nearest Neighbor Search; DEEP 10M; Throughput (QPS): 147000.

- [Approximate Nearest Neighbor Search on GIST](<https://www.sota2.com/research/sota/approximate-nearest-neighbor-search-on-gist.md>): Approximate Nearest Neighbor Search; GIST; Throughput (QPS): 38000.

## Original sources and implementation

[Original paper](<https://arxiv.org/abs/1908.10396>)

[PDF](<https://arxiv.org/pdf/1908.10396.pdf>)
