Neural Search in Action

Type: Tutorial / Conference Session
Event: CVPR 2023 Tutorial
Category: Curated Resource Lists
Tags: tutorials, neural-search, vector-search

Overview

“Neural Search in Action” is a CVPR 2023 tutorial focused on practical neural search systems. It covers how to design and implement search over deep embeddings for large-scale multimodal collections, with emphasis on:

• Million-scale and billion-scale similarity search
• Vector search engines and approximate nearest neighbor (ANN) methods
• Using multimodal encoders (e.g., CLIP) to turn tasks into embedding-and-search problems
• How neural search supports applications like feeding context into large language models (LLMs)
• Designing query languages for real-world neural search applications

Time and Venue

• Event: CVPR 2023 (Computer Vision and Pattern Recognition)
• Format: In-person tutorial session (single-afternoon block)

Schedule & Sessions

1. Opening

   • Time: 13:30–13:40
   • Presenter: Yusuke Matsui
   • Materials: Slides

2. Theory and Applications of Graph-based Search

   • Time: 13:40–14:30
   • Presenter: Yusuke Matsui
   • Focus: Theoretical foundations and practical uses of graph-based approximate nearest neighbor search for neural search systems
   • Materials: Slides

3. A Survey on Billion-Scale Approximate Nearest Neighbors

   • Time: 14:30–15:20
   • Presenter: Martin Aumüller
   • Focus: Methods, algorithms, and system design for billion-scale ANN search, including performance and scalability considerations
   • Materials: Slides (PDF)

4. Break

   • Time: 15:20–15:30

5. Query Language for Neural Search in Practical Applications

   • Time: 15:30–16:20
   • Presenter: Han Xiao
   • Focus: Designing and using query languages tailored to neural search over multimodal data; representing, transforming, and searching embeddings in real-world systems
   • Materials: Slides (PDF)

Features

• Focus on deep embedding-based search across large multimodal datasets
• Discussion of foundation models, prompt engineering, and multimodal encoders (e.g., CLIP) as enablers of neural search
• Practical treatment of million-scale and billion-scale similarity search problems
• Coverage of graph-based search algorithms and their applications in vector search engines
• Survey of billion-scale approximate nearest neighbor (ANN) techniques
• Exploration of how vector search engines support real-world pipelines, including providing context for LLMs
• Introduction to query languages for neural search, focusing on representing, composing, and executing complex search queries over embeddings
• Slide decks available for all main sessions (opening and three technical talks)

Organizers / Instructors

• Yusuke Matsui – The University of Tokyo
• Martin Aumüller – IT University of Copenhagen
• Han Xiao – Jina AI

Media

• Main image/teaser: Neural Search in Action teaser graphic

Pricing

• Not applicable (research tutorial; no pricing information provided).

Source

• Official page: https://matsui528.github.io/cvpr2023_tutorial_neural_search/