苹果公司研究团队于2026年7月2日在ICML 2026会议上发表了题为《Amortizing Maximum Inner Product Search with Learned Support Functions》(通过学习支持函数摊销最大内积搜索)的论文 1。该研究提出了一种新的检索方法,旨在替代传统的基于聚类的索引结构,通过训练小型神经网络来学习查询分布 1。该方法主要应用于提升检索增强生成(RAG)和最大内积搜索(MIPS)系统的效率 1。
在技术实现上,该方法使用输入凸神经网络(ICNN)近似支持函数,并采用回归网络直接预测最佳匹配键 1。模型架构包含两个核心组件:负责路由功能的SupportNet和用于查询重写的KeyNet 1。实验表明,该方案在保持高召回率的同时,显著降低了推理延迟并提高了搜索效率 1。
以包含约260万个独特段落的Natural Questions数据集为基准进行测试时,该技术展现出明显的性能优势 1。具体数据显示,小型SupportNet达到95%路由召回的速度比质心路由快约1.9倍;而最大的KeyNet达到99%召回所需时间减少约8倍 1。
Researchers from Apple have introduced a novel retrieval approach designed to enhance the efficiency of Retrieval-Augmented Generation (RAG) and Maximum Inner Product Search (MIPS) systems 1. The study, titled "Amortizing Maximum Inner Product Search with Learned Support Functions," was presented at the International Conference on Machine Learning (ICML) in July 2026 1. By training small neural networks to learn query distributions, the proposed method replaces traditional clustering-based index structures, aiming to reduce inference latency while maintaining high recall rates 1.
The core technical innovation involves using Input Convex Neural Networks (ICNN) to approximate support functions, alongside a regression network that directly predicts optimal matching keys 1. The architecture consists of two distinct components: SupportNet, which handles routing tasks, and KeyNet, which manages query rewriting 1. Performance evaluations were conducted on the Natural Questions dataset, which comprises approximately 2.6 million unique paragraphs 1. Results indicated that a Small SupportNet achieved 95% routing recall roughly 1.9 times faster than standard centroid routing methods 1. Furthermore, the Largest KeyNet reduced the time required to achieve 99% recall by approximately eightfold compared to baseline approaches 1.
评论
还没有评论,欢迎留下第一条。