Learning to Index in Large-Scale Datasets

In this paper, we present a novel ranking scheme that learns the nearest neighbor relation embedded in the index structure. Given a query point, a direct way to rank clusters of the index structure is based on their Euclidean distances to the query from near to far. However, the data quantization loss will inevitably impair the index accuracy. To address this problem, the proposed method ranks clusters based on the nearest neighbor probabilities of clusters rather than their Euclidean distances. We present two algorithms for offline training and online indexing that leverage the deep neural networks to learn the neighborhood relation. The proposed method can replace the distance-based ranking scheme and can be integrated with other nearest neighbor search methods to boost their retrieval accuracy. Experiments on one million and one billion datasets demonstrate a promising result of the proposed ranking scheme.