ECE Seminar Series – Dec 8 (Mon) @ 2:00pm: “Vector Search: High Throughput and Robust Query Processing and Modern Benchmarks,” Teemu Roos, Prof, CS, U of Helsinki

Location: Engineering Science Building (ESB), Room 1001
Come at 1:30p for Cookies, Coffee and Conversation
LECTURE at the ECE SEMINAR SERIES

Abstract

The rapidly evolving field of vector databases offers solutions for many real-world AI pipelines, including retrieval augmented generation (RAG) and multimodal semantic search. These pipelines require scalable algorithms balancing high throughput and accuracy. Many scenarios also require robustness with respect to mismatches between the data and query distributions. We present a machine learning-based approach to vector search that leads to light-weight index structures with high throughput-accuracy tradeoffs and good out-of-distribution properties. We also discuss some aspects of benchmarking vector search techniques, focusing on modern embedding datasets arising in generative AI applications, and present the VIBE (Vector Index Benchmark for Embeddings) benchmark suite.

Joint work with Ville Hyvönen, Elias Jääsaari, Matteo Ceccarello, and Martin Aumüller.

Bio

Teemu Roos is a Professor of Computer Science at the University of Helsinki and the Leader of the AI Education program at the Finnish Center for AI (FCAI). Roos is also the lead instructor of the Elements of AI online course that has over 1.9M users and has been rated as the world’s best computer science MOOC. Previously, he has held visiting positions at UC Berkeley, MIT, and Cambridge University. His
research interests include statistical machine learning and its applications in astrophysics, neuroscience and epidemiology.

Hosted by: Distinguished Lecture at the ECE Seminar Series in conjunction with REAL AI

Submitted by: Haewon Jeong <haewon@ece.ucsb.edu>