Random projection trees for vector quantization

TL;DR

This paper introduces a computationally efficient tree-structured vector quantization method whose error depends solely on the data's intrinsic dimension, not the ambient space dimension, improving quantization performance.

Contribution

It presents a novel random projection tree approach that reduces quantization error dependence on the apparent dimension of the data space.

Findings

Quantization error depends only on intrinsic data dimension

Method is computationally efficient

Outperforms previous approaches in error dependence

Abstract

A simple and computationally efficient scheme for tree-structured vector quantization is presented. Unlike previous methods, its quantization error depends only on the intrinsic dimension of the data distribution, rather than the apparent dimension of the space in which the data happen to lie.