Early Exit Strategies for Approximate k-NN Search in Dense Retrieval

TL;DR

This paper introduces an unsupervised early exit strategy for approximate k-NN search in dense retrieval, significantly improving efficiency with minimal accuracy loss by adaptively deciding when to stop searching.

Contribution

It proposes a novel unsupervised patience-based early exit method and a cascade approach for efficient dense retrieval, outperforming existing strategies in speed while maintaining effectiveness.

Findings

Up to 5x speedup in A-kNN search efficiency.

Negligible loss in retrieval effectiveness.

Reproducible results with publicly available code.

Abstract

Learned dense representations are a popular family of techniques for encoding queries and documents using high-dimensional embeddings, which enable retrieval by performing approximate k nearest-neighbors search (A-kNN). A popular technique for making A-kNN search efficient is based on a two-level index, where the embeddings of documents are clustered offline and, at query processing, a fixed number N of clusters closest to the query is visited exhaustively to compute the result set. In this paper, we build upon state-of-the-art for early exit A-kNN and propose an unsupervised method based on the notion of patience, which can reach competitive effectiveness with large efficiency gains. Moreover, we discuss a cascade approach where we first identify queries that find their nearest neighbor within the closest t << N clusters, and then we decide how many more to visit based on our patience approach or other state-of-the-art strategies. Reproducible experiments employing state-of-the-art dense retrieval models and publicly available resources show that our techniques improve the A-kNN efficiency with up to 5x speedups while achieving negligible effectiveness losses. All the code used is available at https://github.com/francescobusolin/faiss_pEE