The Unsung Hero of Vector Database: Metric Learning
summary
Metric learning is presented as the fix behind better vector search: using Siamese and triplet architectures with contrastive and triplet loss to stop representation collapse and handle negation. A demo shows CMcC-style training drops similarity between “I love pizza” and “I do not love pizza” from ~94% to ~66%. The open source pipeline EmbeddedThing is Rust‑based and claimed 40x faster for embedding creation, with practical negative mining strategies and a demo that improves negation handling in retrieval.
AI-written summary · about
chapters (7)
- 0:00Introduction to speaker
- 1:26Vectors and databases
- 3:02Classification vs similarity
- 5:02Self-supervised learning
- 7:09Uncertainty and metrics
- 12:31Contrastive learning
- 15:02Q&A and future outlook
related talks
38mApplying Metric Learning: A Logo Recognition Case Study
29mHallucination by Design: How Embedding Models Misunderstand Language
01h 23mVector Databases Demystified: Discover how they work
50mA fun and absurd introduction to Vector Databases
41mSemantic Multimodal Multilingual Similarity Engine
48mA fun and absurd introduction to Vector Databases
26mThe Rise of Vector Databases - Lessons from the Milvus Community
19mVectorize Your Open Source Search Engine