asymmetric distance computation

Definition

Asymmetric distance computation (ADC) is a technique for efficient similarity search, particularly with product quantization, that compares a query vector against pre-quantized database vectors. It optimizes the comparison process by treating query and database vectors differently, often avoiding full dequantization of stored data to reduce memory bandwidth.

At a glance

Executive summary

Asymmetric distance computation is a clever way to quickly find similar items in large datasets without fully unpacking all the stored information. It's particularly useful for making big AI models, like those for language, run much faster and more efficiently on smaller devices by compressing their memory usage.

TL;DR

A technique that helps computers quickly find similar data by cleverly comparing new information against a pre-organized collection, making big AI models run faster and use less memory.

Key points

Efficiently approximates distances between a query and pre-quantized database vectors.
Solves the problem of high memory bandwidth and slow similarity search, enabling large models on edge devices.
Used by vector database developers, large-scale search engines, and LLM deployment engineers.
Unlike full dequantization, it avoids unpacking stored data, significantly saving bandwidth and computation.
A critical research trend for efficient deployment of large AI models, especially LLMs, on resource-constrained hardware.

Use cases

Efficient similarity search in large-scale vector databases for recommendation systems or semantic search.

Compressing KV caches in large language models for deployment on edge devices.

Real-time image retrieval from massive datasets with reduced latency.

Accelerating approximate nearest neighbor search in various machine learning pipelines.

Definition

At a glance

Executive summary

TL;DR

Key points

Use cases

Also known as

Related papers

Related topics